The clinical effectiveness and cost-effectiveness of technologies used to visualise the seizure focus in people with refractory epilepsy being considered for surgery: a systematic review and decision-analytical model

Epilepsy

Epilepsy is the most common serious neurological condition,1 with a prevalence of between 0.5% and 1% in developed countries, and a cumulative incidence of up to 3% by the age of 75 years. 2,3 A review of epidemiological data from European countries reported prevalence as ranging from 0.32% to 0.78%, depending on the country and population studied. 4 The only UK study included in that review4 reported a prevalence of 0.43% in children aged 6–14 years in England. 5 With approximately 456,000 people in the UK having a diagnosis of epilepsy,6–8 overall prevalence can be estimated at around 0.8%. Most patients achieve good seizure control, but up to 30% continue to have seizures despite ongoing treatment with one or more antiepileptic drugs (AEDs). 9

Classification of epilepsies

Epilepsies are a large heterogeneous group of disorders, including those that are genetically determined (idiopathic, tending to present in childhood and adolescence) as well as those that are symptomatic of a brain injury (e.g. stroke or head injury). The epilepsies are classified into syndromes according to seizure types, aetiology, age at onset, electroencephalographic changes and magnetic resonance imaging (MRI) findings. One of the major classifications is between localisation-related epilepsies (affecting a limited area of the brain, i.e. focal or partial), which result in focal seizures and can be simple (no associated impairment of consciousness) or complex (impairment of consciousness), and secondary generalised seizures (resulting in distortion of the electrical activity of the whole, or a large part of, the brain). These epilepsies are subclassified according to the site of the focus (frontal, temporal, parietal, occipital lobes) and then further divided according to aetiology:

• Symptomatic epilepsy Where there is evidence of a lesion or damage to the brain (e.g. mesial temporal lobe sclerosis, cortical dysplasia, tumour, vascular malformation, haemorrhage, infarct, infection and trauma).

• Cryptogenic epilepsy Where there are symptoms to suggest brain damage, but there is no evidence of a lesion or damage to the brain.

• Idiopathic epilepsy Assumed genetic aetiology.

Psychogenic seizures are non-epileptic seizures, although patients can be misdiagnosed initially as having epilepsy. Patients with such seizures can generally be identified using video-EEG monitoring, as they do not have the electrical discharges on EEG that are characteristic of epilepsy.

Epilepsy pathophysiology

An epileptic focus may be located anywhere in the cerebral cortex, and there are a number of potential aetiologies. Approximately 20% of patients who are evaluated for intractable epilepsy are thought to have non-epileptic psychogenic seizures. 10 From an epilepsy surgery perspective, epilepsy tends to be grouped into two categories: temporal and extratemporal. Temporal lobe epilepsy (TLE) is the most prevalent focal (partial) epilepsy, and is the epilepsy that is most frequently treated surgically. 1 The focus is most often located in the amygdalohippocampal region of the mesial (medial) temporal lobe,1 with mesial temporal lobe sclerosis (MTS) being the most common form of TLE for which surgery is undertaken, and is thought to be associated with hippocampal neuronal loss [hippocampal sclerosis (HS)] and gliosis,1 although the cause is still uncertain. Lateral temporal onset can occur, but this is less common. 1 The most common type of extra-TLE is frontal lobe epilepsy. Causes of epilepsy other than MTS include focal cortical dysplasia (congenital abnormality where neurons fail to develop normally), tuberous sclerosis (non-malignant growths caused by an inherited genetic mutation), and brain injury (caused by illness, for example cerebral infarction, alcohol or drug misuse, and trauma), while a large proportion remains cryptogenic.

Surgery for epilepsy

Of the 20–30% of people who continue to have seizures despite drug treatment, the majority have symptomatic or cryptogenic localisation-related epilepsy. A review conducted in 2001 reported response rates in patients with refractory localisation related epilepsy to adjunctive therapies (adjusted for placebo response rates) of between 12% and 29% for antiepileptic drugs (drugs evaluated were gabapentin, lamotrigine, levetiracetam, oxcarbazepine, tiagabine, topiramate, zonisamide) and 12% for adjunctive vagal nerve stimulation. 11 For patients who do not respond to adjunctive therapy, surgical resection of the epileptic focus may be considered, and can result in the patient becoming seizure free (SF). The main aim of surgery is to remove the seizure focus, leaving the patient free from seizures without causing other disability. Most epilepsy surgery programmes focus on temporal lobe surgery (temporal lobectomy); excision of the temporal lobe has a higher rate of success and a lower chance of causing harm than resection of extratemporal lobe foci. Temporal lobectomy has been estimated to result in the long-term cure of TLE (free from seizures and the discontinuation of AEDs) in 25–30% of patients undergoing the procedure, with a further 25–30% becoming SF with continued AED treatment. 1 Surgical procedures other than lobectomy are sometimes performed, for example hemispherectomy, corpus callosotomy, multiple subpial transaction and vagal nerve stimulation. The first randomised controlled trial (RCT) evaluating epilepsy surgery randomised 80 patients with drug-refractory TLE to either surgery or continued AED treatment. At 12 months, 58% of the surgical group and 8% of the AED group were free of complex focal seizures, giving a number needed to treat of 2 [95% confidence interval (CI) 1.3 to 3]. 12

Burden of the disease

According to Hospital Episode Statistics (HES), in 2009–10 epilepsy was responsible for 54,428 consultant episodes and 42,385 admissions (35,515 emergency, 4192 waiting list and 2678 day cases), which accounted for 153,035 bed-days. 7 In 2009–10, 344 operations were conducted where major excision of brain tissue was the main operation, and 3890 where excision of a lesion of brain tissue was the main operation; it is unclear how many of these procedures were conducted on patients with epilepsy. 7 The 1999 Department of Health-commissioned report from the Clinical Standards Advisory Group (CSAG) on services for patients with epilepsy estimated that between 5% and 10% of patients with refractory focal epilepsy might benefit from epilepsy surgery. 3 Using this estimate, the CSAG calculated that between 5000 and 10,000 UK patients might benefit from epilepsy surgery, with between 750 and 1500 cases being added to this cohort each year. A study using the UK National General Practice Study of Epilepsy estimated the number of patients with newly diagnosed epilepsy who may eventually require surgery, based on an estimate of 30,000 incident cases in the UK, to be around 1.5% (95% CI 0.5% to 2.5%) or 450 patients a year. 13

In a 1998 study,9 the direct cost of epilepsy was estimated at £1568 per patient per annum. When seizure frequency was taken into account, the total cost of care for patients having more than one seizure per month was eight times that of patients who were SF (£3508 vs £443). Drug costs accounted for 23% of the overall cost, 43% of which was spent on the then new AEDs vigabatrin (Sabril^®, Sanofi-Aventis) and lamotrigine (Lamicatl^®, GlaxoSmithKline UK), which were prescribed for only 6% of patients. Since this study, several drugs, used either as monotherapy and/or adjunctive therapy, have obtained licences in the UK, including levetiracetam (2000) (Keppra^®, USB Pharma Ltd), oxcarbazepine (2000) (Triteptal^®, Novartis Pharmaceuticals UK Ltd), pregabalin (2004) (Lyrica^®, Pfizer Ltd), zonisamide (2005) (Zobegran^®, Eisai Ltd), rufinamide (2007) (Inovelon^®, Eisai Ltd), lacosamide (2008) (Vimpat^®, UCB Pharma Ltd)14,15 and, most recently, retigabine (2011) (Trobalt^®, GlaxoSmithKline). A 2007 review estimated the cost of epilepsy in European countries at 2004 pricing levels. 4 The total cost (direct and indirect) ranged from 2000 euros (Estonia) to 11,500 euros (Switzerland); approximately 18% was direct health-care costs. 4 The total cost in the UK was estimated at approximately 8250 euros. 4

Single-photon emission computed tomography

Single-photon emission computed tomography uses a radiolabelled compound that binds preferentially to certain areas of the brain, depending on the compound’s properties. The most commonly used methods during epilepsy surgery work up are technetium-99-labelled compounds [hexamethylpropylenamine oxime (HMPAO)] and ethyl cysteinate dimer (ECD). Once the tracer has been given there is a maximum 6-hour window in which to do the scan. Scans can be conducted interictally (not during a seizure), but more reliable information about the site of seizure onset is provided by injecting the radiolabelled compound at the start of a seizure (ictal) or just after it (post-ictal). Scans show an area of increased uptake at the site of seizure activity. Patients require simultaneous video-EEG monitoring and the presence of a member of staff who can give the radiolabelled compound as soon as the seizure starts.

Subtraction ictal single-photon emission computed tomography coregistered with magnetic resonance imaging

Subtraction ictal SPECT coregistered with MRI (SISCOM) subtracts the results of ictal and interictal SPECT images, and overlays the result on an MRI image.

Positron emission tomography

Positron emission tomography uses radiolabelled tracers; ¹⁵O-labelled water is used to assess blood flow (interictal only), [¹⁸F]-fluorodeoxy-d-glucose (¹⁸F-FDG) to assess glucose metabolism (ictal and post-ictal) and flumazenil to visualise the distribution of ionotropic gamma-aminobutyric acid A (GABA_A) receptors in the brain. PET provides better spatial resolution than SPECT. FDG-PET is thought to provide more reliable results and provide better spatial resolution, and is thus more commonly used in selection for surgery.

Volumetric magnetic resonance imaging

Stereological techniques are used to estimate the volume of brain structures, most commonly the hippocampus, amygdala and temporal lobe. 25,26 Differences in volume, usually a reduction when compared with normative data, indicate focal pathology and, potentially, the site of the onset of seizures. Patients may have normal volumes, a unilateral abnormality or bilateral abnormalities that need to be taken into account when evaluating this technology. In addition, volumetric information is usually interpreted in conjunction with quantitative T2 data.

Magnetic resonance spectroscopy

Magnetic resonance spectroscopy is a magnetic resonance technology that is used interictally to measure the relative concentration of certain molecules in an attempt to find focal abnormalities that are consistent with the seizure focus. Proton spectroscopy can provide information about N-acetylaspartate (NAA), creatine lactate and choline-containing compounds, whereas phosphorus-31 spectroscopy provides information about phosphorus-containing compounds, such as adenosine triphosphate.

High-density scalp electroencephalography

High-density scalp electroencephalography uses surface electrodes as in routine EEG, but differs from routine EEG by using more electrode contacts (up to 256, compared with 21 usually used in standard EEG) and sophisticated data analysis strategies. 27

Magnetoencephalography

Magnetoencephalography measures the magnetic fields produced by electrical activity in the brain using sensitive devices such as superconducting quantum interference devices (SQUIDs). Advantages of MEG have been reported as high spatiotemporal resolution, insensitivity to conductivity differences (including skull defects and lesions), complementary sensitivity to EEG, high signal–noise ratio in superficial areas, focus localisation and functional mapping. 22 Disadvantages have been stated as metal implant artefact, cost, insensitivity to radial sources, less sensitivity to deep sources (gradiometers) and limited long-term monitoring feasibility, i.e. low likelihood of ictal recordings. 22

Magnetic source imaging

Magnetic source imaging (MSI) is MEG coregistered with MRI. The two sets of data are combined by measuring the location of a common set of points of reference; these are marked during MRI with lipid markers, and with electrified coils of wire that give off magnetic fields during MEG.

Diffusion tensor imaging

Diffusion tensor imaging is a magnetic resonance technique that quantitatively measures the magnitude and directionality of diffusion in a three-dimensional space.

Identification of studies

The screening of titles and abstracts was conducted by two independent reviewers. All potentially relevant studies were retrieved where available, and two independent reviewers applied the inclusion criteria to the full papers. Disagreements were resolved by discussion. Where consensus could not be reached at the title and abstract stage, the full paper was ordered. Where consensus could not be reached at the full-paper stage, a third reviewer was consulted. Abstracts were included if no associated full paper was identified, and there were sufficient outcome data to extract. Foreign-language papers were excluded unless there was an English- language translation available or if the study had been extracted for the previous review. 36

Inclusion and exclusion criteria

Following development of the protocol review questions (see Chapter 2), the protocol was amended to include a broader range of studies. The protocol amendment impacted only on the designs of studies eligible for the review (see Chapter 3, Study designs). Studies were sought in order to address three clinical questions relating to the use of non-invasive technologies in patients with refractory epilepsy being considered for surgery.

Data extraction strategy

Data extraction forms were developed using Microsoft Access 2007 (Microsoft Corporation, Redmond, WA, USA), which was piloted on a small number of studies. Information relating to study design, patient populations, index test(s), comparators and methodological quality were extracted by one reviewer and checked by a second; disagreements were resolved by discussion. Where multiple publications of the same study were identified, data were extracted and reported as a single study. No RCTs were identified that met the inclusion criteria.

Data from the diagnostic accuracy studies were extracted in order to construct 2 × 4 contingency tables (Table 1). Sensitivity and specificity were extracted where reported when a contingency table could not be constructed.

To gain the maximum clinical utility from the data, we classified the index tests based on whether they were concordant, non-localising, partially concordant or discordant (therefore defining the rows of the 2 × 4 table) with either the final consensus localisation of the comparator tests (where the reference standard was the decision to go to surgery or not) or the site of surgery (where the reference standard was outcome following surgery):

• Concordant The focus identified by the index test is the same as that identified by the comparator tests/final diagnosis, or the site of surgery (not concordance with the final decision whether to go to surgery, or outcome following surgery); this includes results where none of the tests was localising, as there was concordance that no focus could be identified.

• Non-localising The comparator tests localised a focus, or surgery was undertaken, but the index test was non-localising.

• Partially concordant The focus identified by the comparator tests/site of surgery overlapped, but was not identical to, the focus identified by the index test.

• Discordant The focus identified by the comparator tests/site of surgery was a different focus to that identified by the index test.

Where possible, attempts were made to determine whether the final decision as to the site of surgery for the partially concordant and discordant test categories was based on the results of the comparator tests or the index test, when the decision was to go to surgery was the reference standard. From the 2 × 4 contingency tables in which outcome following surgery was the reference standard, hit, miss and error rates were calculated; the definitions and method of calculation are described in detail in Methodological limitation of the diagnostic accuracy studies. This terminology was adopted from Wheless et al. ,42 who reported the results of their test evaluations in this fashion.

Given the nature of the data reported in the diagnostic accuracy studies, there was a level of subjectivity surrounding the extraction of the data into the 2 × 4 contingency tables. This resulted in some disagreements between reviewers that could not be resolved by discussion. In these cases, the final decision was made by the team member with the most experience in epilepsy and the conduct of systematic reviews of diagnostic data (JB). In order to allow reproducibility of the extraction of the diagnostic accuracy data, and for readers to assess the methods used to derive the 2 × 4 contingency tables, a detailed description of these methods is provided for each study alongside the data extraction tables (see Appendix 3).

From outcome prediction studies the analysis used, measure of association, the dependent variable and independent variables included in the model, and the p-values for the individual index test(s) of interest and the model were extracted.

Data from the decision studies, the number of cases in which the index test resulted in a change, (or no change) in the management strategy to be undertaken, the outcome following surgery when an excisional procedure was undertaken, and the predictive values for each of the tests conducted, were extracted.

Critical appraisal strategy

The basis of the quality assessment tool for all three included study designs was the Quality Assessment of Diagnostic Accuracy Studies (QUADAS). 43 Some criteria were used either during the study selection process or were considered not to be appropriate for this review, and were therefore omitted from the final quality assessment. These were:

• The appropriateness of the reference standard and partial verification bias The review was restricted to studies that reported the final decision to go to surgery or the outcome following surgery.

• Differential verification bias Given the nature of this area of medicine, some foci are easier to identify than others, and patients will invariably have different combination of tests during their presurgical work-up. This reflects clinical practice and there may be ethical issues associated with withholding a test that could be beneficial, or conducting a test that could have adverse events associated with it if the prior tests have confirmed a seizure focus. In order to reflect clinical practice, we required iEEG to be an option where the decision to go to surgery was used as the reference standard.

• Progression bias This criterion was not considered appropriate, as patients were known to have severe epilepsy, and it would be very unlikely that the seizure focus or severity of epilepsy would change between tests, even with considerable delays.

The QUADAS43 is a quality assessment tool specifically for diagnostic accuracy studies, and although these were the majority of the included studies, we included other study designs that required criteria relating more generally to observational studies. Two quality assessment tools suitable for this assessment are the Newcastle–Ottawa Scale for cohort studies44 and Downs and Black. 45 There are a number of areas of overlap between these scales and QUADAS,43 such as the assessment of spectrum bias, blinding of assessors and attrition bias. Criteria added to the quality assessment tool for the cohort studies were:

• Was there sufficient description of the groups and distribution of the prognostic factors?

• Was the test reliably evaluated in the outcome prediction studies?

• Were the groups in the outcome prediction studies comparable on all important confounding factors?

• Was there adequate adjustment for the effects of all important confounding factors in the outcome prediction studies?

One criterion from the observational scales was applied to all studies:

• Was follow-up at least 1 year on all patients?

Criteria included in these tools for assessing the quality of observational studies that were not assessed during the quality assessment were:

• The dose–response relationship between intervention and outcome This item was not considered relevant to these studies.

• Adverse events We extracted the rate of adverse events wherever reported, therefore the absence of these outcomes were captured elsewhere.

• Reporting of probability values These were only relevant to the outcome prediction studies, and a minimum requirement for inclusion was the reporting of a p-value for the index test of interest in the multivariate regression.

• Was there evidence of ‘data dredging’ This did not apply to the included studies.

• The use of a power calculation We restricted inclusion to studies based on a minimum number of required participants; the reporting of a power calculation is not common in diagnostic studies, and any further judgement as to what was an ‘adequate’ sample size on our part would have been subjective.

• Date of recruitment Although in Downs and Black45 this relates to whether the intervention and control groups were drawn from the some populations in time, this does have some relevance for our review due to the rapid advancement in the diagnostic technology in this area. However, we restricted our review to studies that used MRI with a magnet strength of ≥ 1.5 T, so excluding older studies using technology deemed to be out of date in the UK; clinical advisors to the project were unable to identify significant changes in the specification of the other technologies being evaluated that would now be considered out of date, therefore any other criteria excluding technologies on this basis would be subjective.

Study quality was assessed by one reviewer and checked by a second; disagreements were resolved by discussion, or referral to a third reviewer.

Methods of data synthesis

Given the substantial clinical heterogeneity across the included studies, a narrative synthesis was used for each type of study design separately. Differences between studies are discussed in the text, and study details were tabulated in Appendix 3.

Some calculations were conducted on data extracted from the diagnostic accuracy studies.

Where individual patient data (IPD) were reported, binary logistic regression was conducted where there were sufficient data.

Likelihood ratios were calculated for each test category of the 2 × 4 contingency table to indicate whether the odds of a decision to go to surgery were greater or outcome following surgery was different between concordant, non-localising, partially concordant and discordant tests. The calculations used are given in Box 1.

There are also limitations with the use of likelihood ratios in this indication. The likelihood ratios cannot be used to indicate the odds of a decision to go surgery or good outcome following surgery associated with a particular index test result, as the rows of the 2 × 4 tables are defined by concordance of an index test, either with the final diagnosis based on the comparator tests or on the site of surgery, depending on the reference standard used. They do not reflect the chance of a ‘correct’ decision, but rather reflect the change in the odds of the outcome. Therefore, the information that can be gained from these data is by a comparison of the likelihood ratio between concordant tests and the other index test categories. When considering the decision to go to surgery, if the results of an index test are given credence by the interpreter, the likelihood ratios will be higher in the concordant and partially concordant categories (as the confidence in the decision would be strengthened) and lower in the non-localising and discordant categories (as there would be less confidence in the location of any identified focus and therefore are less likely to go to surgery). Conversely, if there is no difference in likelihood ratio between the concordant and discordant categories, or the likelihood ratio is higher in the discordant category, the index test is of little or no use (or even counterproductive) when deciding whether surgery should be conducted and where. For the outcome following surgery, if the results of an index test alter the odds of a good outcome then the likelihood ratios will be higher in the concordant and partially concordant categories than those in the non-localising and discordant categories. Conversely, similar likelihood ratios across these categories would indicate no difference in the change in the odds of a good outcome based on the conduct of the index test. However, there are a number of factors that can have an impact on surgical outcome other than the accuracy of localisation. The likelihood ratios presented must, therefore, be interpreted with caution; the differences in likelihood ratios between categories need to be substantial and consistent across similar populations, if they are to be used as indicators of the benefit of conducting the index test.

Lack of an established reference standard

In some areas of medicine there is no agreement as to what constitutes an ideal reference (gold) standard. Where there is no widespread agreement, a test cannot be considered as a gold standard. Consensus is required in order to gain reliability; a diagnosis made using agreed-on criteria in one setting is more likely to be made the same way in a different setting than would be the case when no such consensus exists. 46 Using an imperfect reference standard will directly lead to bias in accuracy estimates. 47 The direction of the bias depends on whether errors by the index test and imperfect reference standard are correlated (a positive correlation will inflate the estimates of accuracy) and the magnitude will depend on the frequency of errors by the imperfect reference standard and the degree of correlation in errors between index test and reference standard. 47 There is currently no agreed, consistently measurable, independent, reference standard against which new technology to identify a seizure focus can be compared. 22 This was highlighted as a problem in the prior HTA report. 36 The most commonly used reference standards include continuous video-EEG, iEEG, consensus from a combination of tests, latent class models, and outcome following surgery; the limitations of each of these are discussed in turn.

Continuous video-EEG monitoring for a number of days to record ictal activity directly assesses the seizure onset but ictal EEG using surface electrodes may fail to find a seizure focus or may localise it inaccurately, particularly when the seizure activity arises from parts of the cerebral cortex some distance from the scalp. 36 Although video-EEG is a useful diagnostic modality, the proportion of patients in whom useful data were collected could not be considered sufficient for a gold standard test. 48

Invasive/intracranial EEG is considered the final decisive test of surgical decision-making in the large subpopulation of surgical candidates who do not have either localised ictal EEG or MRI findings, and therefore considered by some to be a suitable reference standard. 49 iEEG has very high sensitivity;22 however, there are significant limitations for its use as a reference standard. First, the placement of the electrodes is crucial; a single electrode has a field of view of only several millimetres, and the electrodes need to include the site of the focus to successfully provide an accurate location. 22,36,50 iEEG has been shown to have a relatively high percentage of non-localisation, negatively affecting diagnostic measure of the index tests being evaluated when used as a gold standard. 51 In addition, there is a potential for complications such as infection and bleeding,23 and the ability to repeat the test is limited because of local scarring after an initial evaluation,22 therefore there are ethical considerations in using such a test in all patients, which means that verification will be in selected cases only and can lead to biased estimates of sensitivity and specificity. 24

Combination of test results is often used as the reference standard, which is the currently the case when localising a seizure focus. 24,36 The composite reference standard used in studies needs to include a combination of tests which provide results that are meaningful when defining the target condition, and the likelihood for residual misclassification considered. 47 However, there is no consensus as to which tests should constitute this combination, and the combination/sequence of tests varies considerably across studies. Some concerns with the use of a composite of tests as the reference standard for the localisation of a seizure focus are:

1. The subjective nature of the interpretation of the test results, the subsequent observer variation, and the potential that the post-test probability could be a function of the accuracy of individual observers rather than the overall accuracy of the diagnostic pathway. 52,53

2. The destabilisation of a correct diagnosis with continued testing due to doubt being cast by additional information or the new information sufficiently changing the post-test likelihood as to suggest an alternative diagnosis. 52

The use of an expert panel to discuss test results and produce a consensus decision as to the presence or absence of the target condition based on all sources of information helps address these concerns with the use of a combination reference standard, and is an appropriate approach to address the issue of imperfect reference standard. 47 A number of factors have been identified as affecting the reliability of a consensus-based diagnosis by a multidisciplinary team – namely the number of experts and expertise mix, the way patient information is presented, and how to obtain a final classification54 – and these factors would need to be reported in order to determine the reliability of the reference standard.

An alternative to the consensus method is the use of a statistical model for combining multiple test results: latent class models. 47 Latent class models relate the observed patterns of test results to latent categories of patients with and without the target condition; using this analysis, specific test results associated with the patient categories can be identified, and sensitivity and specificity of the tests can be estimated. 47 A potential benefit of these models is their objectivity, as they examine the strength of statistical relationships among variables; the main disadvantage is that the target condition is not defined in a clinical way, so there can be lack of clarity about what the results stand for in practice. 47

Outcome following surgery could be considered the ultimate reference standard; however, surgical outcome can be affected by known and unknown variables that are unrelated to the accuracy of test localisation,22 and is therefore subject to a number of limitations as a reference standard. First, many individuals may not proceed to surgery, therefore surgical outcome does not provide information for the proportion of patients who are assessed and for whom the decision not to undertake surgery is made. Second, the index test may have been used in making the decision to proceed with surgery. These two factors would tend to increase the estimated diagnostic accuracy of the index test. 36 Third, the success, or otherwise, of surgery can be influenced by factors other than the results of the localisation decision, such as intra- or postoperative complications, incomplete resection of the seizure focus, excision of adjacent tissue to the focus identified that included the seizure focus, and post-surgical management regimen. Finally, surgical follow-up has to be sufficiently long in order to determine the persistence of the surgical outcome. 50

The major limitation that afflicts all reference standards that are currently available for diagnostic technologies for the work-up for epilepsy surgery is the inability to verify whether the index test was accurate, and whether the decision not to undertake surgery was appropriate. Neither iEEG nor SF surgical outcome is considered to reliably reflect epilepsy localisation, even in well-designed studies that minimise ascertainment bias and lack of independence,22 and there is no consensus as to the most appropriate combination reference standard.

In light of these limitations, and the need to determine the impact of test results on the decision-making process, this review concentrated on studies that reported the decision to undertake surgery or not based on the results of tests (where iEEG was an option in order to reflect clinical practice), and outcome following surgery using a range of thresholds for the definition of a good outcome; these are currently the best available, although imperfect, reference standards. The difficulties with extracting and interpreting the data from these studies are discussed in the following sections.

Difficulties with producing 2 × 2 tables of test performance and standard diagnostic outcome measures

The prior HTA report stated that the inability to dichotomise the data in order to construct 2 × 2 tables of test performance was a major limitation of the data from the diagnostic accuracy studies, as standard measures of diagnostic performance, such as sensitivity, specificity, and predictive values, could not be calculated. 36 In other areas of medicine for which an extended contingency table such as this is presented (risk stratification, cancer staging, threshold categories), the data can be dichotomised by selecting a single threshold (e.g. stages 1 and 2 vs stages 3 and 4). These approaches always risk reducing the applicability of the information derived, but with an advantage of increased statistical precision.

However, the previous review dichotomised the data when outcome following surgery was used as the reference standard in order to produce relative risks. 36 In order to collapse the data into a 2 × 2 table, the number of concordant and partially concordant tests were combined, as were the number of non-localising and incorrectly localising tests. Although combining the concordant and partially concordant tests may be valid (to produce an ‘at least partially correct’ category), combining the non-localising and discordant tests would mean combining two distinct populations for which the consequences of the test results would be very different. The consequences for patients non-localising and discordant tests would be significantly different; a non-localising test would likely result in a patient receiving medical management (MM), but an incorrect localisation could lead to surgery at an incorrect location, which could leave the patient with continued seizures, transient or long-term complications, alterations of function, and potentially alteration of personality. In addition, the combined non-localising and incorrectly localising categories will represent false- and true-negative tests; although the non-localising tests can be allocated as true- or false-negatives, discordant tests identified a focus and therefore are positive, and cannot be classified as negative tests.

Difficulties interpreting the data from the 2 × 4 contingency tables

Given the disadvantages of collapsing the data into a 2 × 2 table of test performance, we tried to determine (1) the proportion of scans that contributed to the decision whether or not to undertake surgery when a decision to go to surgery was the reference standard and (2) the proportion of patients in whom the index test identified a focus correctly, incorrectly or missed a resectable focus in those who underwent surgery. In order to do this, we would need to be able to classify the scans in each of the cells of the 2 × 4 contingency table (Table 2). However, this proved problematic. In order to determine whether or not the test was concordant for entry into the 2 × 4 contingency table, subjective decisions had to be made by the reviewer for some test results, which means that the numbers in each cell could vary with differences in the opinions of those interpreting the data. The full data extraction is reported in Appendix 3, along with a detailed description of how the 2 × 4 tables were derived for each study.

For studies reporting the decision to go to surgery or not, the following assessment of each cell of the 2 × 4 table (see Table 2) were considered.

• A and B Concordance between the comparator tests (generally a consensus diagnosis) and the index test means that the decision to go to surgery (or not) may have been the same whether the established tests or the index test were used alone or in combination. Therefore, it could be argued that there was no value in conducting the index test in these patients. However, it could also be argued that the index test increased confidence in the site of the seizure focus and subsequently the decision to go to surgery or that the index test could have been used instead of the original combination of tests, avoiding the use of iEEG. None of these alternatives can be proven for any individual study or case within a study.

• C and D Given a focus was identified by the comparator tests, the non-localising index test provided no additional information regarding the potential site of excision, and may cast doubt on the results of prior tests. Where the decision was not to go to surgery, we cannot tell from these data whether it was due to uncertainty of the location of the seizure focus as a result of the non-localising index test, discordance of individual comparator tests, or other reasons. Therefore, the change in the confidence, if any, and influence on the surgical decision cannot be reliably determined in these cells.

• E and F The partial concordance of the index test with the comparator tests means that there may have been some added value in the performance of the test, in terms of increased confidence in the focus where there is partial overlap. It may be that different tests contained in the combination of comparator tests gave slightly different foci, and the index test may show closer concordance with some tests than others. Alternatively, additional potential sites for the seizure focus indicated by the index test may decrease confidence in the potential benefit of surgery and may have led to a move away from an operative approach. The confidence the surgical team places on the results of each of these tests will determine whether or not the index test seems to be confirmatory of the focus. The data presented do not allow the true additional value of the index test to be determined by this route.

• G Despite the discordance between tests, surgery was performed. If the site of excision was based on the results of the comparator tests then the index test provided no added value, and indeed may have made the decision to undertake surgery more difficult. However, if the excision was at the site identified by the index test, there was not only value to undertaking the additional test, but there was believed to be greater value in the results of the index test than the other tests.

• H Although the index test was discordant with the comparator tests, and there was a decision not to go to surgery, this may be due to uncertainty surrounding the focus as a result of the index test, discordance between the individual comparator tests or maybe other reasons why surgery was not considered appropriate, therefore the value of the index test in making this decision is completely uncertain.

From the descriptions for each cell of the 2 × 4 table, it can be seen that it is not possible to determine whether or not an index test was of value in the decision-making process using data from studies of a diagnostic accuracy design. This was compounded by the difficulties encountered when categorising test results to the cells of the 2 × 4 table during data extraction; results of the range of comparator tests were sometimes reported separately, with the site of surgery being the only consensus decision. In studies in which the site of surgery was based solely on the comparator tests, classification was relatively straightforward; however, the index test contributed to the final localisation decision in some studies. Detailed descriptions at to how we extracted data from each study are reported in Appendix 3, indicating which patients were difficult to classify and how we made our final decision. Our method was just one of several that could have been used to classify these patients; differences in interpretation of the data by different reviewers could significantly alter the conclusions drawn.

Even where IPD was presented, the reviewer could not determine the contribution of a test result to the final decision, as different tests would carry different weight and some test results would be clearer than others. When data could be extracted into a 2 × 4 contingency table, any generalisation applied to all of the test results in any one cell of that table is inappropriate; some concordant, partially concordant or discordant tests could be crucial in the decision-making process, whereas others are not, and non-localising test results could cast sufficient doubt on the site of a seizure focus to alter the management plan for a patient.

It is apparent that without clear documentation from the clinical teams detailing the process of decision-making, it is not possible to reliably classify tests in order to determine whether or not they provide added value during the decision-making process from such data sets.

Only the clinical team interpreting the test results can give an indication on whether or not a particular test result was informative for the decision-making process, and this level of information needs to be recorded at the time the study is being conducted. Therefore, in order to determine the impact of an index test on the management strategy for patients being considered for epilepsy surgery, the consensus decision pre- and post-index test(s) needs to be recorded at the time the study was conducted; the diagnostic accuracy studies as reported are inappropriate to evaluate the results of diagnostic test in this context.

Similar problems were encountered when the outcome following surgery was used as the reference standard. As with the decision to go to surgery, data regarding outcome following surgery were extracted from the diagnostic accuracy studies into 2 × 4 contingency tables (Table 3).

For each cell of the 2 × 4 table, we tried to determine whether the test was a hit, miss or error, as described below:

• A The index test identified the focus at which surgery was conducted, and the outcome from surgery was good; this shows that the index test of interest was correct (Hit). It could be argued that the index test may have been sufficient to confirm the seizure focus, and an iEEG avoided. However, there is the possibility that the tissue resected was more extensive than that identified by the tests, and the seizure focus could be the contained within the additional resected tissue that was not identified by the tests.

• B The index test identified the seizure focus at which surgery was conducted but the outcome from surgery was poor, therefore these tests could be classified as Errors. However, a poor outcome following surgery could occur after surgery at the correct location due to intra- or postoperative complications or an incomplete resection of the focus, and therefore these tests are Unclassifiable.

• C The index test was non-localising but the comparator tests identified a focus; surgery was conducted, and surgical outcome was good. Therefore, the comparator tests identified the correct seizure focus, and the index test failed to identify a resectable focus; the index test was a Miss.

• D The index test was non-localising but the comparator tests identified a focus and surgery was conducted, but the outcome following surgery was poor. Therefore, the correct focus was identified by the comparator tests, but there were intra- or postoperative complications or incomplete resection of the focus; the focus identified by the comparator tests was not the correct focus, or only part of the focus; or the non-localising result of the index test was correct and the patients has no resectable focus. The correct interpretation and classification of these tests is therefore uncertain; these tests may be Misses or Correctly non-localising, therefore these tests are Unclassifiable.

• E Partial concordance with the site of surgery means that the index test identified a focus that was either more or less extensive than the area excised. Given the good outcome following surgery, these tests could be deemed ‘Partial hits’; however, it is unclear whether such a good outcome would have been achieved if the area for excision was informed solely by the focus identified by the index test, therefore these tests are considered Unclassifiable.

• F As with cell E, the index test identified a focus that was either more or less extensive than the area excised, these tests could be deemed ‘Partial errors’. However, not only is it unclear whether or not such a poor outcome would have been achieved if the area excised was informed solely by the index test focus, it could also be that the localisation was correct and intra- or postoperative complications or incomplete resection of the seizure focus resulted in the patients continuing to have seizures, therefore these tests are also Unclassifiable.

• G As the index test was discordant with the actual site of surgery, the area of excision is based on the results of the comparator tests. Therefore, the good outcome following surgery indicates that the comparator tests identified the correct focus, and the index test result was an Error; these are the only index test results that can definitely be classified as errors.

• H As with cell G, the site of surgery and the index test were discordant, and therefore the area of excision was based on the results of the comparator tests. Given the poor outcome following surgery, the comparator tests could either have identified an incorrect focus, or surgery could have been conducted at the correct location, but intra- or postoperative complications or incomplete resection of the seizure focus resulted in continued seizures. If it was assumed that the surgery was complete and complication free, it is still unknown whether or not the patient would have had a good outcome following surgery if the alternative site indicated by the index test had been excised. Therefore, these tests are Unclassifiable.

These descriptions highlight the complexity and limitations of the data retrieved from such studies, how interpretation of these data is difficult, and clinical utility limited. The only cells that provide unequivocal results are Cells C (Misses) and G (Errors) (the numbers of patients in Cell A who had a good outcome following inadvertent resection of the seizure focus by removing tissue adjacent to site identified by the tests were likely to be small, and we classified these as Hits in Table 6 in the results section (see Results, below). This section demonstrates that the accuracy of two or more tests when there is not complete concordance cannot be reliably evaluated using these data. For example, when patients had an index test that was discordant with the site of surgery and had a bad postoperative outcome (cells B, F and H), the poor outcome following surgery may not be because the surgery was conducted at the incorrect location but because of intra- or postoperative complications – such as gliosis, infection or bleeding – or there was incomplete resection of the seizure focus. Therefore, if a second operation is conducted at the alternative focus, the patient may undergo a second inappropriate and unnecessary surgical procedure. If the original surgery was conducted at the incorrect location, there is no way to know whether the alternate seizure focus identified by the index test is the true focus. If it was, and a second surgery was undertaken, there may be consequences of the first surgery, which may mean that reoperation would not give the good outcome it would have done if it was the original surgical procedure, therefore the index test may not appear to be accurate when in fact it was.

Quantity and overall quality of research available

The searches identified 3251 citations; 534 were retrieved for full paper screening, of which 161 were abstracts. Figure 1 shows the flow of studies through the review. Eighteen studies met the inclusion criteria (n = 1312; range 24–469). 42,49,55–70 Nine studies49,55,56,58,59,62,66,68,70 evaluated PET, three49,57,61 evaluated SPECT, one67 evaluated SISCOM, two55,56 evaluated vMRI, three55,56,65 evaluated MRS, three42,63,69 evaluated MEG, two49,64 evaluated MSI, and one60 evaluated MRI/MRS. No studies met the inclusion criteria that evaluated HD-EEG, DTI or EEG-fMRI. Of the 18 studies, 13 were single-gate (cohort) diagnostic accuracy studies,42,49,55–58,61,63–66,68,69 seven reported outcome prediction using either a multivariate regression analysis or Bayesian classifiers,49,56,60,62,67,68,70 and one was a decision study. 59 Brief study details are provided in Table 4; full study details are provided in Appendix 3.

FIGURE 1.

Flow of studies through the review.

Overall, the quality of the available evidence was poor. Figure 2 shows the proportion of studies that met each of the quality criteria; full results of the quality assessment are given in Appendix 2. Of the 18 studies, 1242,49,55–59,63–66,68 described the inclusion criteria clearly but only two61,65 recruited a spectrum of patients that was representative of clinical practice, 1642,49,55–63. 65^. 67–70 (89%) reported the execution of the index test, 1142,49,55,56,58,59,61,64,65,68,70 (58%) the comparator test(s)/surgical procedure, in sufficient detail to permit replication; seven studies55,57–59,62,67,68 (39%) interpreted the index test without knowledge of the comparator test(s)/site of surgery, five55,56,64,67,68 (26%) interpreted the comparator test(s)/decided the site of surgery without knowledge of the index test, and clinical data were available when test results were interpreted, as would be available in clinical practice in two studies59,66 (11%). Blinding of the interpreters of test results and other clinical data allows the independent assessment of the test results; however, it is worth noting that this does not reflect the work-up for surgery clinical practice, where decisions are made based on all the test results available. Only 10 studies42,55–59,61,66,69,70 reported uninterpretable results; given the nature of the index tests being evaluated, it is not uncommon to have uninterpretable results, and if these are omitted from data sets, accuracy could be overestimated. Twelve studies49,56,57,59,61,63–68,70 (68%) reported follow-up of at least 12 months in all patients. Of the seven outcome prediction studies, four60,62,68,70 (21%) reported dropouts to be similar across groups, one49 (5%) reported surgical outcome assessment being blinded to test result, it was unclear in any outcome prediction study whether the groups were comparable on all important confounding factors,71 and none adequately adjusted for its effects.

FIGURE 2.

Summary of the quality of the included studies. Q1: Was the spectrum of patients representative of the full range of patients who will receive the test in practice? Q2a: In the diagnostic accuracy and decision papers, were the selection criteria described clearly? Q2b: In the outcome prediction papers, were the prognostic factors described clearly and in full? Q3a: Was the execution of the index test described in sufficient detail to permit replication? Q3b: Was the execution of the reference test(s)/surgical procedure described in sufficient detail to permit replication? Q4a: Was the index test interpreted without knowledge of the comparator test(s)/site of surgery? Q4b: Was the comparator test(s) interpreted/site of surgery decided without knowledge of the index test? Q5: Were the same clinical data available when the test results were interpreted as would be available in clinical practice? Q6: Was outcome assessment blinded to exposure status? Q7: Was follow-up at least 12 months in all patients? Q8: Were uninterpretable/intermediate results reported? Q9: Were withdrawals from the study reported/explained? Q10: Were dropout rates and reasons for dropout similar across groups in the outcome prediction studies? Q11: Was the test reliably evaluated in the outcome prediction studies? Q12: Were the groups in the outcome prediction studies comparable on all important confounding factors? Q13: Was there adequate adjustment for the effects of all important confounding factors in the outcome prediction studies?

Assessment of clinical effectiveness

Key findings

The main finding of the review of clinical effectiveness is that the available evidence is inadequate, both in terms of methodological design and population sizes, to reliably inform clinical practice. Importantly, there is a distinct lack of studies evaluating the clinical utility of diagnostic tests, and therefore the impact of these tests on clinical decision-making and patient outcomes; data from the diagnostic accuracy studies are not sufficient to address these issues. One study met the inclusion criteria that reported the impact of one of the index tests of interest (FDG-PET) on the decision to go to surgery, and the outcome following surgery in those who underwent an excisional procedure. 59

Strengths and limitations of the review

The systematic review was based on an extensive search. Abstracts were included when there were sufficient data to be extracted; authors were not contacted. Foreign-language studies were included only if an English-language translation was available. Therefore, although publication and language bias are likely to be present, this was the pragmatic choice, given the nature of the evidence base. The inability to inform clinical practice was not due only to the lack of volume of evidence, but also to the type of evidence currently available; if the number of single-gate diagnostic accuracy studies were doubled, firm conclusions still could not have been drawn. In addition, we screened 161 studies that were retrieved as abstracts; none of these described studies that were relevant to the research question or that were different in design to those that were retrieved as full papers. Therefore, as the focus of the review was the limitations of the currently available studies rather than an assessment of the relative diagnostic accuracy, we feel that the lack of unpublished and foreign-language data does not compromise the conclusions of the review.

Studies were selected using inclusion criteria defined a priori. These criteria were extended in terms of study design to encompass all available evidence relevant to the research question. A further modification was made to the inclusion criteria concerning the requirement of iEEG in studies only reporting the decision to go to surgery. Originally, iEEG was a requirement; however, this was changed so that routine MRI was required (either as a component of the comparator tests or as an index test), and with iEEG being available as an option, as this reflects clinical practice more closely.

Limitations of the available evidence

There were a number of limitations of the available evidence, and overall, the quality of the available evidence was poor. Four studies were identified that evaluated the use of one of the index tests of interest on the decision-making process, of which three were decision studies59,72,73 and one was a RCT;74 unfortunately, only one of the decision studies passed the inclusion criteria for the review. 59 The specific limitations of the data from the diagnostic accuracy studies were discussed in detail in Chapter 3 (see Methodological limitations of the diagnostic accuracy studies). Further limitations were the generally small study sizes, patient selection bias, substantial clinical heterogeneity across the studies, terms of the combination of tests to determine the seizure focus, the technology used during the index tests and the populations recruited, even across those that evaluated the same index test, precluding the pooling of studies.

Comparison with previous systematic reviews

Some of the diagnostic technologies evaluated in this review have been evaluated in previous systematic reviews. 36,38,75 One was the prior HTA report36 which addressed a broad-ranging question evaluating any neuroimaging technology compared with any reference standard. The review included 75 diagnostic accuracy studies evaluating SPECT, PET, MRI, CT, SISCOM, MRS or NIRS, and nine outcome prediction studies evaluating SPECT, PET, MRI, SISCOM or MRS in univariate and/or multivariate regression analyses. This was a well-conducted, thorough review of the published and unpublished evidence available to 2003. Given the conclusions of that review, that MRI seemed to be the most consistent predictor of surgical outcome, probably due to MRI being effective in the localisation of lesions such as tumours and vascular malformations, and that the evidence available at the time could not reliably inform clinical practice, the present work did not attempt to update the previous review. Instead, this review concentrated on a focused but valuable clinical question that constituted a small section of the prior report, attempting to concentrate on the added value of non-invasive technologies over and above routine MRI and EEG, and restricting the inclusion of studies based on the reference standard used, and omitting routine MRI, CT and NIRS from the review. In addition, this review expanded in terms of the exploration of the data from diagnostic accuracy studies by maintaining the four different populations as categorised by test concordance rather than collapsing the data into a 2 × 2 table of test performance, by searching specifically for studies that reported the impact of the index tests of interest on the decision-making process, and including non-neuroimaging on-invasive technologies, HD-EEG and MEG, as index tests.

Two further systematic reviews are worth noting. 38,75 Uijl et al. 38 evaluated a range of technologies used to lateralise or localise the seizure focus in patients with TLE and used the consensus diagnosis or the decision to perform surgery as the reference standard. 38 Lau et al. 75 reviewed the use of MEG in a heterogeneous epilepsy population. Data in both of these reviews were dichotomised in order to produce 2 × 2 tables of test performance from which standard diagnostic outcomes were calculated. The method used to dichotomise the data was not reported in the review by Uijl et al. ;59 however, the review was conducted by the same authors as the decision study included in our review, and that study combined inconclusive and normal results, which included patients with bilateral temporal foci, non-localising tests and indeterminate/uninterpretable results. The review by Lau et al. 75 dichotomised data into concordant and discordant tests; it is unclear how non-localising tests were dealt with. As with the prior HTA report, both reviews concluded that there was insufficient evidence to inform clinical practice; Uijl et al. 59 recommended further studies evaluating the added value of consecutive tests on the decision-making process, and Lau et al. 75 recommended more controlled and consistent studies to determine whether or not MEG could replace iEEG.

Ours is the first systematic review of clinical evidence to attempt to determine the added value of a wide range of non-invasive technologies in a heterogeneous epilepsy population, without dichotomising the data and therefore avoiding combining distinct populations for whom the index test results could lead to different management strategies and long-term consequences for the patients.

Methods of systematic review of existing cost-effectiveness evidence

The search strategy for the project is reported in Chapter 3 (see Search strategy). In addition to the databases listed in that section, the NHS Economic Evaluations Database (NHS-EED) was searched for economic evaluations (via Wiley, The Cochrane Library website, issue 7 of 12 July 2010). A broad range of studies were considered in the assessment of cost-effectiveness including economic evaluations conducted alongside trials, modelling studies and analyses of administrative databases. Only full economic evaluations that compare two or more options and consider both costs and consequences (including cost-effectiveness, cost–utility and cost–benefit analyses) were included in the review of economic literature.

Two reviewers independently assessed all obtained titles and abstracts for inclusion. Any discrepancies were resolved by discussion. The quality of the cost-effectiveness studies was assessed according to a checklist based on that developed by Drummond et al. 76 This checklist reflects the criteria for economic evaluation detailed in the methodological guidance developed by the National Institute for Health and Clinical Excellence (NICE)77 and is reported in Appendix 5.

The systematic literature search identified only one study that met the criteria for inclusion in the cost-effectiveness review. 78 The following section provides an overview of this paper and assesses its relevance to the scope of this technology assessment.

Review of O’Brien et al.78

This study by O’Brien et al. 78 evaluates the cost-effectiveness of a range of presurgical strategies for the lateralisation of temporal lobe epilepsy in patients; a summary of the paper is provided in Table 12. The costs of the strategies are evaluated over the lifetime of the patient from the perspective of the Australian health-care sector. It is not clear if the benefits of the strategies are also considered over the patient’s lifetime. There is no discussion of whether costs or benefits are discounted. The author used a deterministic decision tree to model the different strategies; there is no discussion of any formal long-term modelling.

The four strategies evaluated are:

• Strategy 1 No presurgical evaluation. All patients receive MM alone.

• Strategy 2 Video-EEG and routine MRI alone. Patients with concordant (positive in both tests) results receive surgery; patients with concordant (negative in both tests) and discordant results receive MM.

• Strategy 3 Video-EEG/routine MRI plus ictal SPECT. As Strategy 2, but patients with discordant results from video-EEG and routine MRI receive an additional test (ictal SPECT). Patients with positive results to ictal SPECT receive surgery and patients with negative results receive MM.

• Strategy 4 Video-EEG/routine MRI plus FDG-PET. As Strategy 3 but patients with discordant results from video-EEG and routine MRI receive FDG-PET (rather than SPECT). Patients with positive results to FDG receive surgery and patients with negative results receive MM.

The main focus of the study is on the three imaging strategies (Strategies 2, 3 and 4) and specifically on the cost-effectiveness of FDG-PET (Strategy 4) compared with the use of video-EEG/routine MRI alone (Strategy 2). The data to inform the FDG-PET strategy (as well as the video-EEG/MRI strategy) are derived from an observational study undertaken by the authors, which is also used to estimate the concordance of video-EEG and routine MRI, whereas the data used for the SPECT comparison are derived from a separate published study. 79 The incremental cost-effectiveness of Strategies 3 and 4 are reported compared with Strategy 2.

The model used is based on a decision tree, which captures the pathways for each strategy depending on the results of the tests and the decisions concerning the use of surgery or MM. According to whether surgery or MM was provided, health outcomes differ: the proportion of patients having Engel class I/II outcomes (‘good outcome’) rather than Engel class III/IV outcomes (‘bad outcome’) depends on the treatment strategy. The Engel outcomes were then linked to estimates of lifetime cost savings through a reduction in longer-term direct costs due to improved Engel outcome.

The data collected by O’Brien et al. 78 was derived from a cohort of 176 patients who received FDG-PET scans at the Royal Melbourne Hospital as part of a presurgical evaluation for chronically refractory focal epilepsy from November 1996 to July 2001. In addition to F-FDG-PET all patients had initially undergone interictal EEG, prolonged video-EEG and MRI (16 of the 176 also received iEEG but the results and impact of undertaking this test are not reported). Although some patients in O’Brien et al. ’s sample78 had ictal and interictal SPECT (n = 15), given the small numbers the data that inform the decision tree are taken from an alternative study (Won et al. 79). However, this highlights that patients may have had a sequence of tests, which influences the decision to proceed to surgery and surgical outcomes; the implications of this sequencing are not well considered by the author or incorporated into the decision tree.

Clinical effectiveness of treatment strategies is based on patients achieving Engel class I/II outcomes. Because the cohort study also collected Engel outcome data, the results of the tests were linked directly to outcome. The sensitivity and specificity of each test for a class I/II outcome was calculated as well as the prevalence of class I/II outcomes and the proportion of indeterminate video-EEG and MRI scans. However, it is unclear from the paper how the author used these values to inform the inputs needed to fully specify the decision tree.

In terms of unit costs of visualisation technologies and surgery, the perspective of the purchaser is taken, and the author uses the Medicare schedule of costs (published by the Australian Government Department of Health and Aged Care). The lifetime cost savings after an Engel class I/II outcome (e.g. medication, hospital visits) are estimated for MM and surgery as $440,000 per patient. It is unclear how lifetime cost savings were estimated and applied in the model. It is also unclear whether or not surgical outcomes were assumed to impact on the life expectancy of patients.

The results of the decision tree analysis were provided for Strategies 2–4. It is not clear why Strategy 1 is excluded or not used to determine any of the ICERs; it can only be speculated that the author did not want to consider the broader question of the cost-effectiveness of the TLE surgery itself, so excluded the MM treatment only option. The sensitivity of the results to several parameter values is tested, including the prevalence of class I/II outcomes; the proportion of video-EEG/MRI results that are indeterminate; and the sensitivity and specificity of the visualisation tests. The calculations underlying the results provided are unclear, and no quality-of-life data are provided to inform the cost-effectiveness. As a result, it appears the ICER results provided for Strategies 3 and 4 against Strategy 2 represent the incremental cost per additional Engel class I/II outcome. Furthermore, it is unclear if these results include the estimation of cost savings over a lifetime, and it is therefore unclear if they represent only the results of the short-term decision tree analysis or the cost-effectiveness of the strategies over the lifetime of the patients.

Population

The decision problem addressed by the decision model is the cost-effectiveness of presurgical visualisation strategies for the lateralisation of seizure focus in medically refractory epileptic patients who have had discordant findings from initial video-EEG and MRI scans. Patients who are deemed eligible for a presurgical evaluation have been defined as medically refractory through their failure to respond suitably to at least two antiepileptic drugs. This definition of patients as medically refractory as a prerequisite of surgical candidacy is commonly used and is consistent with the studies used to populate the model. However, owing to the nature of the clinical data available the decision was made to constrain the analysis of the model to only patients with TLE.

Model structure

The main objective of the decision-analytic model is to estimate the incremental cost-effectiveness of FDG-PET in people with medically refractory TLE who are being considered for surgery, and who have already undergone a video-EEG and MRI that has resulted in an indeterminate result (i.e. the decision to proceed to surgery is uncertain). Rather than providing a complete picture of the cost-effectiveness of the full range of alternative strategies, it is intended that the development of the model serves to provide an illustration of how data from appropriately designed clinical studies can be used to inform a decision model to evaluate the cost-effectiveness of visualisation technologies in the presurgical evaluation of TLE surgery. Through this process the model also highlights the range of additional data that are required to appropriately assess the long-term impact of the initial decisions on costs and outcomes suitable for assessments of cost-effectiveness.

The model itself has two components: a short-term element, which characterises the period over which these localisation strategies are applied, and, if appropriate, surgery is conducted, and a long-term element, which considers the costs and outcomes over the remaining lifetime of the patient. The model is designed to evaluate the impact of the application of further tests on the decision made by a clinician after an initial round of feasibility tests for TLE surgery.

The technologies we could evaluate were conditioned by the availability of data. We thus focused on FDG-PET and iEEG, the technologies evaluated in the only ‘decision study’ identified. 59 Even with this restricted set of technologies, a number of separate strategies can be considered based on how the technologies could be used, i.e. combination of tests and the order in which these are applied. The following strategies, to be applied to patients who have had discordant findings from initial video-EEG and MRI scans, were included in the current evaluation:

• Strategy 1 No additional tests are performed.

  1. – All patients with indeterminate results from MRI/EEG receive MM.

• Strategy 2 FDG-PET is performed; if the result of this test does not lead to a positive (S+) or negative (S–) decision to undertake surgery and still leaves the clinician uncertain whether or not to proceed to surgery (S?) then no further tests are performed.

  1. – If S+ after FDG-PET: patients are offered surgery.

  2. – If S– after FDG-PET: patients are offered MM.

  3. – If S? after FDG-PET: patients are offered MM.

• Strategy 3 FDG-PET is performed; if the result of this test does not lead to a clear decision on whether or not the patient should proceed to surgery (S?), iEEG is offered.

  1. – If S+ after FDG-PET: patients are offered surgery.

  2. – If S– after FDG-PET: patients are offered MM.

  3. – If S? after FDG-PET: patients are offered iEEG and the evaluation of the result of this test determines the provision of treatment (S+ after iEEG implies a decision to proceed to surgery, whereas S– and S? assume MM).

As the model is based directly on the approach and results of the single study by Uijl et al. ,59 the model incorporates a series of assumptions about the sequencing of visualisation technologies that are inherent within the Uijl et al. study:59

• Owing to the invasive nature of the iEEG, it is assumed that it would always be the last preoperative test available to the clinician; therefore, in the situation outlined above, FDG-PET will always precede it.

• iEEG never directly follows the initial MRI and video-EEG. This assumption is made because an uncertain MRI and video-EEG may lack the potential to accurately define a sufficiently specific location for the iEEG to visualise (iEEGs can be carried out over only a small brain section at any time). An additional FDG-PET that results in an ‘S?’ is assumed to provide enough additional evidence to facilitate investigation by iEEG. In clinical practice, further non-invasive imaging may be undertaken; given the absence of evidence, however, it was not possible to explicitly evaluate such strategies in the current assessment.

Non-invasive tests are assumed to have to potential to provide sufficient information to facilitate TLE surgery. This assumption is consistent with our understanding of clinical practice. 82

Short-term model

The short-term model is structured as a decision tree as shown in Figure 3, reflecting the short-term clinical outcomes and adverse events associated with each of three localisation strategies. For modelling purposes the short-term model is assumed to have a time span of 1 year for all of the strategies analysed. This time span was chosen to be consistent with the main study used to inform the longer-term modelling. 83

FIGURE 3.

Structure of the short-term decision tree. DS, experiencing at least one disabling seizure.

The short-term model defined in this section was based on the structure of the available evidence in the study by Uijl et al. ,59 in that it considered the use of MRI and video-EEG followed by the possibility of FDG-PET and iEEG as screening strategies. As a result, the three strategies outlined in Figure 3 are directly informed by the data presented by Uijl et al. 59 as discussed in previous sections.

Within Strategy 1, no further visualisation technologies are used after an indeterminate result of previous tests (video-EEG + MRI ‘S?’). Given that the focus is not localised, surgery is not provided, and the patients are assumed to be treated with MM alone.

Strategy 2 is defined as the provision of FDG-PET to the patient population. The results of FDG-PET define the treatment strategy recommended to the patient. The FDG-PET scan may yield a result that is consistent with clinical requirements to proceed to surgery (represented by S+ after FDG-PET in Figure 3). These patients are offered surgery, to which they may comply, or not. Patients for whom the results of the FDG-PET scan lead to a decision not to proceed to surgery (S– after FDG-PET) or are inconclusive (S? after FDG-PET) are assumed to be provided with MM.

Strategy 3 also includes the provision of FDG-PET; however, if the result of the FDG-PET scan is inconclusive then the patient is a candidate for a further visualisation technology – an iEEG. The iEEG test is assumed to be definitive and has two possible outcomes: surgical candidacy or not.

Both invasive procedures that may occur during the short-term model (iEEG and surgery) are associated with risk of death and complications. Non-fatal complications, which may occur as a result of invasive procedures, are defined as either transient (i.e. only affecting the individual for a year after surgery, e.g. postoperative infections) or permanent [i.e. having a permanent impact on health-related quality of life (HRQoL), e.g. verbal decline, hemianopsia or hemiparesis]. For simplicity of presentation and as complications are not assumed to have additional prognostic implications, neither transient nor permanent complications are represented in the short-term model in Figure 3. However, the cost and impact on QoL are incorporated. It is also assumed that all patients who experience a permanent complication during an iEEG are not offered surgery and receive only MM.

In the base-case, short-term model it will be assumed that compliance to surgery and iEEG is 100%, such that all patients who are deemed suitable for surgery or an iEEG subsequently receive them. This assumption, and the sensitivity of results to it, will be considered later in this chapter. In addition compliance with FDG-PET will also be assumed to be 100%.

The short-term tree considers three treatment end points: (1) patients can be deemed unsuitable for surgery and continue to receive MM alone for their epilepsy, (2) the patients can be referred for surgery and survive the procedure or (3) the patients can die as a result of procedure-related mortality due to the use of iEEG (Strategy 3 only) or surgery (Strategies 2 and 3). For all patients in Strategy 1, and those patients who survive the initial procedure for Strategies 2 and 3, there are three possible outcomes: (1) achieving seizure freedom for a year after the treatment option was provided, (2) having a disabling seizure (DS) within that year or (3) dying within that year from epilepsy- or non-epilepsy-related events. In the event of death (procedural or during the course of the year), the patient’s progression through the model stops.

There are a limited number of costs associated with the short-term model:

• cost of FDG-PET and iEEG

• cost of TLE surgery

• additional cost of complications occurring during surgery or iEEG.

No costs other than those directly related to the pathways of the patients through the decision tree are considered.

Similarly, the impact on HRQoL – expressed in terms of heath disutility – associated with these patients is reduced due to factors directly related to the patient pathways considered; these are:

• disutility of transient and permanent complications.

Long-term Markov model

The long-term model aims to characterise the long-term prognosis for patients leaving the short-term model after (1) achieving seizure freedom for a year after the treatment option was provided or (2) having a DS within that year. As we have highlighted previously, imaging strategies allow for the identification of patients eligible for surgery, and thus their benefits are largely dependent on how much better the outcomes attained with surgery are in comparison with MM. The long-term structure of the model will reflect the differences between surgery and MM. Important factors in the evaluation of the long-term prognosis are whether the patients have seizures (or DSs) and whether therapy with antiepileptic drugs is sustained. 84

As the focus of our research project was on the performance and value of imaging technologies as opposed to specifically evaluating the clinical effectiveness and cost-effectiveness of the subsequent treatment strategies, the long-term model was based largely on a previously published study. 83 This study by Choi et al. 83 used a decision model (including a long-term Markov model) to estimate life expectancy and quality-adjusted life expectancy of surgery and MM for patients with refractory epilepsy. Despite implications on costs not having been quantified in Choi et al. ’s study,83 given the aims and methods used we believe this study constitutes an appropriate basis to inform the current model. Also, the study by Choi et al. 83 was informed by a seemingly comprehensive search of the evidence base: a series of detailed reviews of clinical, epidemiological and QoL data were undertaken. Given the constraints of our project, we did not attempt to update or critically appraise these reviews. Instead, we focused on adapting the study for the purposes of assessing long-term costs and QALYs for a UK cost-effectiveness study.

The model we use is presented in Figure 4 and is adapted from the long-term Markov model implemented by Choi et al. 83 In a Markov model, hypothetical individuals reside in one out of a set of mutually exclusive health states at particular points in time. During discrete time intervals of equal length (normally referred to as Markov cycles), individuals can either remain in a particular health state or move to a separate health state (e.g. because of a patient experiencing a particular clinical event). The movements between states represent the potential clinical pathways that a patient may follow at different time points and over his or her remaining lifetime. The likelihood that an individual remains in a particular health state, or moves to a separate state, is specified in terms of transition probabilities. Defining and subsequently estimating these transition probabilities represent key structural and analytical elements of any decision model. Here, patients enter the Markov model directly from the short-term model either having been SF for the first year, having a DS within that year or having died. The probability of entering each of these states is therefore different depending on whether the patients were treated by MM or surgery.

FIGURE 4.

Structure of the long-term Markov model.

Patients who are long-term SF in the Markov model (defined here as having more than 2 years of seizure freedom) may be given the possibility of having the treatment with AEDs stopped, as is consistent with NICE clinical guidelines. 80 We expect that accounting for this will impact mainly on the long-term costs of care for patients, and represents an additional possible benefit of surgery. To incorporate this into the model, the state representing seizure freedom was separated into a set of ‘tunnel states’: SF for 1 year, SF for 2 years, and SF for 3 years and longer. This modelling ‘trick’ allows for tracking the time patients are SF, incorporating time dependency into the model without having to define a more complex structure. Specifically, when in the first two health states, patients face a probability of dying, of having a DS or progressing to the next state. Patients who maintain SF for more than 2 years arrive at the health state ‘SF for 2 or more years’. Patients in this SF state can be either on or off AEDs, such that not all patients who are long-term SF are removed from treatment to allow for the implicit inclusion of other factors that may relate to the decision to discontinue medication. The proportion that is removed is defined as ‘pAED’ and is discussed later in this chapter. 85,86 In the next time cycle, patients can remain in this state, die or have DSs again, in which case patients that were off AEDs are assumed to restart MM. All MM and surgically treated patients are assumed to receive AED treatment in the short-term; however, all patients have the possibility to become long-term AED-free through prolonged seizure freedom. The long-term model developed from Choi et al. 83 also assumes that there is no discontinuation of AEDs in either the MM or the surgical treatment strategies other than discontinuation due to prolonged seizure freedom or death.

Patients who have some form of DS in a year can either remain in this state (such that in the next year they have another DS), die or become SF and as such transit to state ‘SF for 1 year’ as shown in Figure 4. Owing to limitations in the available data the model does not correlate seizure history with the risk of having seizures in the following year.

The patients who have experienced a transient or permanent complication during an invasive procedure are assumed to have the same long-term transition probabilities but have an additional QoL disutility applied in the model.

From all states in the long-term model patients face the risk of death; this can be epilepsy related or due to unrelated reasons. The impact of epilepsy on mortality for any given cycle of the model is assumed to be different for patients who are SF from those who have had a DS in that year.

Patients’ transition probabilities in the long-term model depend on whether they are being treated with MM or have had surgery, but the same structure represented in the figure above applies to both treatment options. The difference between the sets of transition probabilities will be discussed later in this chapter.

The efficacy of the AED used in the MM treatment strategy (and therefore in the NHS) is assumed to be consistent with that used in Choi’s analysis; for simplicity this will be assumed to represent an average of AEDs recommended by NICE. 80 The efficacy of specific individual AEDs or adverse events associated with their use were not considered in the model.

The costs considered in the long-term model are as follows:

• cost of an annual course of AEDs

• non-drug costs of provision of treatment to epilepsy sufferers who are SF and on AEDs (e.g. general practitioner visits)

• non-drug costs of provision of treatment to epilepsy sufferers who have had a DS in that year (e.g. hospital admissions, outpatient visits, etc.).

The impact on QoL was assessed by applying utility decrements to age/sex match estimates of QoL from the UK general population. The HRQoL decrements reflect the disutility of having:

• at least one DS in a given year

• a transient complication (only applied during the first year after surgery)

• a permanent complication following an invasive procedure (applied throughout the remainder of a patient’s lifetime).

Model inputs

The sources of input data and a description of the available evidence used to inform the decision model are reported in this section. In the base case, it is assumed that the age at the start of the model is 35 years old. 83 The Choi et al. study83 (a study based on American data) is used for the base-case age for two reasons: first, a suitable study identifying the relevant distribution of epilepsy suffers’ ages in the UK was not identified, and, second, the long-term model relies on many transition probabilities from Choi et al. ’s study,83 which may be less applicable to patients of different ages. The proportion of male to female patients was assumed to be 0.49, based on data derived from the CIA World Factbook (2011 estimates) for that population age. No data were available on the male–female ratio in epilepsy specifically, which is often claimed to affect women more than men. 87

Assumptions necessitated by available data

Several assumptions were necessary to consider the cost-effectiveness of visualisation strategies in the manner presented in the previous section, relating to both the short- and long-term models. In addition to the model structures being derived from respective studies of Uijl et al. 59 and Choi et al. ,83 a significant number of data were also sourced from these studies. The assumptions necessary to use these data, as well as those needed to link the two parts of the model together, are discussed in the following section.

All patients were assumed to comply with invasive procedures in the base case (also to tests and treatments), with no associated uncertainty. This assumption was applied, as Uijl et al. 59 did not report compliance data. This assumption will be evaluated further in the scenario analysis, reported later in this report.

In addition, in using data from the study by Uijl et al. ,59 several other assumptions had to be made about the nature of the data reported and the patient pathways:

• all patients in the study by Uijl et al. 59 who have an S– result from FDG-PET are assumed in the model to receive the MM treatment strategy

• all patients who have an S+ result from the FDG-PET are assumed to receive the surgery treatment strategy.

Several other assumptions had to be made in informing the short-term model:

• Mortality due to an invasive procedure is assumed to be the same for iEEG and TLE surgery. In Choi et al. ’s study83 it is recognised that, although there is no evidence suggesting a mortality risk associated with the TLE surgery, there is a risk associated with the use of anaesthetic in such procedures.

• In the cases of all visualisation and surgical technologies it is assumed that the technology is already in place, such that the cost considered is equal to the marginal cost as reported in the NHS schedule for reference costs 2009–10. 88 This can be extended to assume that all of the relevant expertise necessary exists as spare capacity, such that there is no ‘learning-curve’ associated with any of the strategies.

The long-term model is also based on a several additional assumptions:

• As clinical recommendations suggest (NICE guidance 20, 201181), the models assumes that a proportion of patients are taken off treatment with AEDs if they are SF for a minimum period of 3 years.

• Owing to limitations in the available data, it is assumed that patients who are taken off AEDs in our model have the same mortality and the same likelihood of having a DS as those who are SF but still on AEDs. The decision to discontinue AEDs is also being considered independent of the probability of a relapse occurring.

• The evidence from Choi et al. 83 is used throughout the time horizon of the model (e.g. the likelihood of relapse). As this evidence is drawn from the wider literature, it is not always clear what time frames were used in the input data.

Similarly, several assumptions have been made to incorporate Choi et al. ’s83 QoL scores into this study’s model:

• Choi et al. 83 assume the uncertainty over the QoL scores to be represented by a uniform distribution. Given that we, here, used disutilities rather than the scores themselves, gamma distributions were used instead. Although this assumption may cause a degree of error from fitting a new distribution using the CIs provided, this effect is likely to be small.

• Quality-of-life decrements due to events in the model are applied against the expected QoL for an individual of a certain age and gender. 89

As with all decision models, by using distinct sources of data these need to be assumed valid for their purpose. The link between the short-term results and the longer-term progression is based on different data sets, which, in their own rights, were considered here to represent the best available evidence to populate the cost-effectiveness model. The sample considered in the Uijl et al. study59 is thus assumed to be representative of the population defined in our decision problem. Similarly, the evidence from Choi et al. 83 is assumed to represent adequately the progression in the disease, conditional on knowing the treatment decision of whether or not to proceed to surgery. The assumptions were necessary, given the limited evidence available to populate both the short- and long-term models and also given the resource constraints of this project, which meant it was not possible to develop our own long-term model. However, the robustness of the results to several key assumptions is explored within later sections.

Base-case analysis

The results of the base case are presented in Table 17 for the three strategies.

The results show that Strategy 2 (use of FDG-PET following an indeterminate result with video-EEG and MRI) represents a more expensive but more effective strategy than Strategy 1 (no further non-invasive or invasive tests). The ICER of Strategy 2 compared with Strategy 1 is £1671 per QALY. Strategy 3 (use of iEEG following an indeterminate result with FDG-PET and video-EEG/MRI) is also more costly and also more effective than Strategy 2. The ICER of Strategy 3 compared with Strategy 2 is £3201 per additional QALY.

At a threshold of £20,000 per QALY, the probability that Strategy 3 is the most cost-effective strategy is 0.83. This probability increases marginally to 0.84 at a threshold of £30,000 per QALY. Figure 5 presents the decision uncertainty surrounding the probability that each strategy is cost-effective using multiple CEACs across a range of potential threshold amounts a decision-maker might be prepared to pay for a QALY.

FIGURE 5.

Cost-effectiveness acceptability curves for the base-case analysis.

Consequently, at conventional thresholds used to determine value for money in the NHS, Strategy 3 – the use of FDG-PET for indeterminate cases following video-EEG and routine MRI, together with iEEG for any cases that are still indeterminate following the FDG-PET – appears to be the most cost-effective strategy, with low uncertainty surrounding this decision.

Scenario analysis

The previous sections show that applying the inputs and assumptions in the base-case analysis, Strategy 3 appears the most cost-effective strategy and decision uncertainty is minimal. However, given the series of assumptions which were required within the base-case analysis, additional analyses were undertaken in order to assess the robustness of the results of the base-case model. These additional scenarios focused on several critical assumptions required to both link the separate sources (Uijl et al. 59 and Choi et al. 83) within the cost-effectiveness modelling framework and to populate the long-term model itself.

Inevitably the additional value of any visualisation strategy is closely related both to the impact on treatment decisions as well as the cost-effectiveness of the treatments themselves (MM or surgery). However, the results from the theoretical ranges for individual parameters considered in the base case indicated that the cost-effectiveness conclusions were insensitive to the short-term efficacy of surgery and MM (i.e. Strategy 3 appeared optimal irrespective of whether the short-term success of surgery in terms of the probability of being SF at 1 year was 0 or 1). This clearly demonstrates the importance of the additional benefits being assumed for surgery within the long-term model in driving the cost-effectiveness results. As previously noted these estimates were informed using a recently published study,83 which, although considered most appropriate to inform our long-term model, includes a number of assumptions, particularly in relation to the maintenance of additional long-term benefits for surgery compared with MM (i.e. over and above the impact that surgery has on the proportion of patients assumed to be SF at the end of the short-term model). As a result, it was considered important to explore the robustness of the base-case results to these additional long-term benefits using a separate scenario analysis.

Impact of the long-term benefits associated with surgery

The study by Choi et al. 83 assumes that TLE surgery confers two main advantages over MM: (1) a significantly higher proportion of patients become SF in the first year after surgery compared with MM and (2) surgery is also assumed to further reduce the long-term impact of epilepsy (avoiding relapses, increasing the likelihood of remission and improving mortality). As we were unable to critically appraise the full range of data sources used to inform these assumptions, we considered an alternative scenario in which the impact of surgery is only assumed to alter the proportion of patients who become SF in the first year. That is, the long-term transition probabilities, costs and HRQoL of patients conditional on becoming SF after the first year are assumed to be the same regardless of whether this was achieved via surgery or MM. Within this scenario, the additional benefits from surgery are thus constrained to increasing the probability of becoming SF within the first year.

Table 20 presents the results of this scenario alongside the results of the base-case analysis. The ICER estimates for Strategies 2 and 3 increase significantly compared with the base-case results. Although the ICER of Strategy 2 compared with Strategy 1 remain below conventional thresholds used to determine value for money in the NHS (£20,000–30,000 per QALY), the ICER of Strategy 3 compared with Strategy 2 now exceeds these thresholds. Consequently, the assumptions regarding the additional long-term benefits of surgery appear critical to determining which visualisation strategy appears optimal in terms of cost-effectiveness considerations. Although the use of FDG-PET still appears to be potentially cost-effective (ICER = £11,526 per QALY), the use of an additional test (iEEG) for patients in whom the decision to proceed to surgery is still uncertain following the results of FDG-PET no longer appears to be cost-effective. In other words, if the benefits of surgery do not extend beyond the probability of becoming SF within the short-term model, MM now appears to be the most appropriate management option for patients in whom the surgical decision following FDG-PET remains uncertain.

TABLE 20 - Cost-effectiveness results (base case vs alternative scenario)

N/A, not applicable.

Strategy	Cost (£)	QALYs	ΔCost (£)	ΔQALYs	ICER (£)	Probability of being cost-effective at
						£20,000	£30,000
Base-case analysis
Strategy 1	23,775	12.88	–	–	–	0.14	0.13
Strategy 2	26,621	14.58	2846	1.70	1679	0.03	0.03
Strategy 3	27,696	14.91	1075	0.33	3227	0.83	0.84
Alternative scenario (no long-term benefits of surgery)
Strategy 1	23,726	12.89	–	–	N/A	0.36	0.25
Strategy 2	27,207	13.19	3482	0.30	11,526	0.37	0.36
Strategy 3	28,416	13.23	1208	0.04	32,876	0.27	0.39

Figure 6 presents the probability that each strategy is cost-effective over a range of threshold values. When compared with the base-case analysis it is now evident that at conventional threshold values (of between £20,000 and £30,000) there is considerably more uncertainty surrounding the decision regarding which is the optimal strategy based on cost-effectiveness considerations in this alternative scenario.

FIGURE 6.

Cost-effectiveness acceptability curves for the alternative scenario.

Given the higher uncertainty surrounding the cost-effectiveness results for this alternative scenario, we replicated the approaches to assessing the sensitivity to parameter uncertainty applied to the base-case analysis. Tables 21 and 22 are thus analogous to those reported for the base case and present the theoretical range of parameter values that will lead to the decision to adopt each strategy (see Table 21) and the probability of actually observing values leading to different adoption decisions (see Table 22).

TABLE 21 - Range of values for each parameter that would lead to each of the alternatives being cost-effective (alternative scenario)

max., maximum; min., minimum; TP, transition probability.

Variable	Mean value	Strategy 1	Strategy 2	Strategy 3
Short-term probabilities
Probability of S+ after FDG-PET	0.56	0 to 0.05	0.05 to 1
Probability of S– after FDG-PET, given not S+	0.68		0 to 1
Probability of S+ after iEEG	0.83		0 to 1
Probability of permanent complications, iEEG	0.04		0 to 1
Probability of transient complications, surgery	0.08	0.25 to 1	0 to 0.25
Probability of permanent complications, surgery	0.04	0.15 to 1	0 to 0.15
Probability of mortality after surgery	0.003	0.05 to 1	0 to 0.05
Probability of mortality after iEEG	0.003		0 to 1
TP of SF patients in the first year, MM	0.08	0.35 to 1	0 to 0.35
TP of SF patients in the first year, surgery	0.719	0 to 0.55	0.55 to 0.95	0.95 to 1
Long-term TPs
TP from SF to DS after the first year, MM	0.254	0.35 to 1	0.15 to 0.35	0 to 0.15
TP from SF to DS between first and fifth years, surgery	0.254		0 to 1
TP from SF to DS after the fifth year, surgery	0.254		0 to 1
TP from DS to SF between first and fifth years, MM	0.047	0.15 to 1	0 to 0.15
TP from DS to SF after the fifth year, MM	0.016		0 to 1
TP from DS to SF between first and fifth years, surgery	0.047		0 to 1
TP from DS to SF after the fifth year, surgery	0.016		0 to 1
Long-term mortality ratios (in relation to the general population)
For patients in SF	1.11		0 to max.
For patients in DS, MM	5.64	0 to 1	1 to 15	15 to max.
For patients in DS, surgery	5.64		0 to max.
HRQoL and disutilities
Baseline utility	0.91		0.05 to 1	0 to 0.05
Disutility for SF in MM	0	Min. to –0.1	–0.1 to 0.1	0.1 to 1
Disutility for SF in surgery, no complication	0		Min. to 1
Disutility for SF in surgery, permanent complication	–0.2		Min. to 0
Disutility for SF in surgery, transient complication	–0.01		Min. to 0
Disutility for DS in MM	–0.21	–0.15 to 0	–0.25 to –0.15	Min. to –0.25
Disutility for DS in surgery, no complication	–0.19		Min. to 0
Disutility for DS in surgery, permanent complication	–0.31	Min. to –0.65	–0.65 to 0
Disutility for DS in surgery, transient complication	–0.22		Min. to 0
Other
Proportion of patients off AEDs (SF) after 5 years	0.36		Min. to 0

TABLE 22 - Probability that each parameter might take values that would lead to each of the alternatives being cost-effectivea

TP, transition probability.

The probability of observing certain parameter values was defined as for the PSA (alternative scenario).

Variable	Strategy 1	Strategy 2	Strategy 3
Short-term probabilities
Probability of S+ after FDG–PET	–	1	–
Probability of S– after FDG–PET, given not S+	–	1	–
Probability of S+ after iEEG	–	1	–
Probability of permanent complications, iEEG	–	1	–
Probability of transient complications, surgery	–	1	–
Probability of permanent complications, surgery	–	1	–
Probability of mortality after surgery	–	1	–
Probability of mortality after iEEG	–	1	–
TP of SF patients in the first year, MM	–	1	–
TP of SF patients in the first year, surgery	–	1	–
Long-term TPs
TP from SF to DS after the first year, MM	0.19	0.67	0.14
TP from SF to DS between first and fifth years, surgery	–	1	–
TP from SF to DS after the fifth year, surgery	–	1	–
TP from DS to SF between first and fifth years, MM	–	1	–
TP from DS to SF after the fifth year, MM	–	1	–
TP from DS to SF between first and fifth years, surgery	–	1	–
TP from DS to SF after the fifth year, surgery	–	1	–
Long-term mortality ratios (in relation to the general population)
For patients in SF	–	1	–
For patients in DS, MM	–	1	–
For patients in DS, surgery	–	1	–
HRQoL and disutilities
Baseline utility	–	1	–
Disutility for SF in MM	–	1	–
Disutility for SF in surgery, no complication	–	1	–
Disutility for SF in surgery, permanent complication	–	1	–
Disutility for SF in surgery, transient complication	–	1	–
Disutility for DS in MM	0.33	0.43	0.24
Disutility for DS in surgery, no complication	–	1	–
Disutility for DS in surgery, permanent complication	0.05	0.95	–
Disutility for DS in surgery, transient complication	–	1	–
Other
Proportion of patients off AEDs (SF) after 5 years	–	1	–

The results of the analyses of parameter uncertainty highlight that removal of the additional longer-term benefits of surgery greatly increases the sensitivity of the model to parameter uncertainty. Despite this greater sensitivity of the model, when the probability of actually observing parameter values that would results in changes to the optimal strategy were considered, in the majority of cases changes in the parameter values do not appear to alter the conclusion that Strategy 2 now appears to be the most cost-effective strategy.

Table 21 shows that more parameters within their theoretical range now have an influence in the adoption decision and that, for most of these, the parameter values in which this shift occurs are closer to the mean. Perhaps of most note from these results is the finding that the cost-effectiveness results are now sensitive to the initial success rates of surgery in terms of the probability of patients becoming SF at 1-year (see ‘TP of SF patients in the first year, surgery’ in Table 21). If the success rate of surgery is < 55% then Strategy 1 would appear to be the optimal strategy. Strategy 2 appears optimal for success rates between 55% and 95% and Strategy 3 would only become optimal if the success rate of surgery was > 95%. This finding contrasts with the base-case analysis, which demonstrated that the optimal strategy (Strategy 3) was insensitive to the success rate when the range of additional long-term benefits was assumed for patients who received surgery.

Despite the greater sensitivity of the model to theoretical changes in individual parameters, when the probability of actually observing parameter values that would result in changes to the optimal strategy was considered, in the majority of cases changes in the parameter values do not appear to alter the conclusion that Strategy 2 now appears to be the most cost-effective strategy.

As previously stated, it was assumed in both the base-case analysis and alternative scenario that the compliance to iEEG and surgery was assumed to be 100%. We also explored the robustness of the cost-effectiveness results to a range of alternative estimates. Our results indicated that the conclusions from the base case and alternative scenario would not be altered unless the compliance to iEEG and surgery was < 20%. During the clinical effectiveness review any papers that recorded compliance to surgery and/or to iEEG were recorded, and the relevant compliance values were extracted. Thirteen papers reported compliance: all had sufficient data on compliance to surgery but only four on compliance to iEEG. 42,49,55–58,61–63,65,66,68,69 The compliance rates for surgery ranged from 63% to 100% and for iEEG ranged from 89% to 100%. We therefore considered that the conclusions from the base case and alternative scenario were robust to this assumption.

Discussion

The results of the base-case analysis suggested that the use of FDG-PET for those patients in whom the decision to proceed to surgery was unclear following video-EEG and routine MRI, followed by iEEG for patients with indeterminate FDG-PET results (i.e. Strategy 3) appeared the most cost-effective strategy at conventional cost-effectiveness thresholds of £20,000–30,000. These conclusions were sensitive to the additional benefits that were assumed to be conferred over the longer-term for patients who received surgery compared with MM alone (i.e. the benefits over and above those attributed to the additional success rate of surgery in increasing the probability of patients becoming SF at 1 year). When these additional benefits were excluded, the use of FDG-PET remained cost-effective; however, MM, as opposed to the use of iEEG, now appeared to be the more appropriate management strategy for patients in whom the decision to proceed to surgery was still unclear following the results of FDG-PET.

We consider that these initial results are important in several respects. First, they provide a provisional indication that some form of non-invasive testing (at least with FDG-PET) appears cost-effective. Second, our analysis demonstrates that it is feasible to assess cost-effectiveness based on appropriately designed clinical effectiveness studies. Thirdly, the model structure highlights the range of additional inputs and assumptions required in estimating cost-effectiveness appropriately and the analysis presented here could now serve to provide both a framework and set of inputs/results that could be revised and updated as new evidence emerges. Finally, the model demonstrates that the value of the visualisation strategies is inextricably linked both to their impact on the decision to proceed to surgery or not and the cost-effectiveness of the subsequent treatments themselves. Our results indicate that the long-term effectiveness and cost-effectiveness of surgery is likely to be an important driver of the cost-effectiveness of alternative visualisation strategies and possible sequences. As such, the number of non-invasive tests used as part of a possible sequence will clearly depend on the long-term maintenance of benefits with surgery. Our provisional findings have indicated that the cost-effectiveness of one possible sequence of an additional non-invasive test (after video-EEG and routine MRI) followed by a further invasive test for patients in whom the decision is still unclear, is sensitive to the potential longer-term benefits for patients receiving surgery beyond the impact on the initial success rate in terms of patients becoming SF.

An important limitation of our work, as we noted from the outset, is that these findings need to be considered in light of the limited clinical data identified from the clinical effectiveness review and the assumptions required to link these data to long-term consequences (costs and outcomes) suitable for informing cost-effectiveness analysis. It was not considered feasible to assess the full range of potential strategies that would be required in order to appropriately inform NHS practice concerning the optimal use of the range of existing visualisation strategies that are available. As such, it is not possible to conclude that FDG-PET is more or less cost-effective than other non-invasive testing strategies or compared with a sequence of non-invasive tests until such time that additional clinical studies are undertaken using appropriate methods suitable for informing these assessments.

Interface used: Ovid MEDLINE(R) In-Process & Other Non-Indexed Citations and Ovid MEDLINE(R) – 1950 to present

• Search date: 23 July 2010.

• Records identified: 1936.

• Restrictions: 2003 to present.

1. exp Epilepsy/ (110,555)

2. epilep$.ti,ab. (75,417)

3. 1 or 2 (123,219)

Line 3 captures terms for epilepsy

1. exp Magnetic Resonance Imaging/ (232,901)

2. exp Magnetic Resonance Spectroscopy/ (152,076)

3. Magnetoencephalography/ (4460)

4. (magnetic resonance tomograph$ or magnetic resonance scan$ or magnetic resonance imag$ or magnetic resonance spectroscop$ or magnetoencephalograph$ or magnetic source imag$ or high density scalp eeg or high density scalp electroencephalogra$ or high density video eeg or high density video electroencephalogra$ or high density eeg or high density electroencephalogra$).ti,ab. (110,579)

5. (mr tomograph$ or mr scan$ or mr imag$ or mr spectroscop$).ti,ab. (38,154)

6. (fmri or mri or nmr or mrs or mrsi or hmrs or “(1)hmrs” or 1hmrs or ihmrs or hmrsi or meg or msi or hd-eeg or hd eeg or hdeeg).ti,ab. (194,728)

7. exp Tomography, Emission-Computed/ (59,345)

8. (spect or single photon emission computed tomograph$).ti,ab. (18,581)

9. (pet or positron emission tomograph$ or subtraction ictal spect or subtraction ictal single photon emission computed tomograph$ or siscom).ti,ab. (43,592)

10. volumetric exam$.ti,ab. (7)

11. quantitative t2 measure$.ti,ab. (8)

12. voxel based morphomet$.ti,ab. (1214)

13. (diffusion weight$ imag$ or dwi).ti,ab. (3867)

14. (diffusion tensor imag$ or dti).ti,ab. (3280)

15. (fluid attenuated inversion recovery or flair).ti,ab. (1896)

16. or/4-18 (503,707)

Line 19 captures terms for imaging technology

1. 3 and 19 (13,231)

Line 20 captures records containing both epilepsy terms and imaging technology terms

1. Neurosurgical Procedures/ (11,010)

2. Anterior Temporal Lobectomy/ (390)

3. exp Cerebral Decortication/ (951)

4. exp Craniotomy/ (8031)

5. Split-Brain Procedure/ (17)

6. (surger$ or surgic$ or neurosurger$ or neurosurgic$ or lobectom$ or cortical resection or hemispherotom$ or hemispherectom$ or callosotom$ or callotom$).ti,ab. (992,393)

7. or/21-26 (1,000,015)

Line 27 captures surgery terms

1. 20 and 27 (3751)

Line 28 captures records containing epilepsy terms and imaging technology terms and surgery terms

1. exp Animals/ (14,917,330)

2. Humans/ and exp Animals/ (11,394,975)

3. 29 not 30 (3,522,355)

4. 28 not 31 (3734)

Line 32 discards records which are about animals

1. (2003$ or 2004$ or 2005$ or 2006$ or 2007$ or 2008$ or 2009$ or 2010$).ed. (5,599,605)

Line 33 retrieves records which were added to MEDLINE (i.e. entry date) between 2003 and 2010

1. 32 and 33 (1936)

Key:

• / = indexing term (MeSH heading)

• exp = exploded MeSH heading

• $ = truncation

• .ti,ab. = terms in either title or abstract fields

Notes:

The appropriate MeSH heading for HD-EEG is Electroencephalography. However, this heading covers a lot of material irrelevant to this search, so it was discarded in this strategy.

The MeSH heading Tomography, Emission-Computed/ includes the headings Tomography, Emission-Computed, Single-Photon and Positron Emission Tomography.

Interface used: FILE 55: Biosis Previews(R) 1993–2010/August W5 (c) 2010 The Thomson Corporation

• Search date: 7 September 2010.

• Records identified: 623.

• Restrictions: 2008 to present.

1. 66890 EPILEP?/TI,AB,DE

2. 133147 (MAGNETIC(W)RESONANCE(W)TOMOGRAPH? OR MAGNETIC(W)RESONANCE(W)SCAN? OR MAGNETIC(W)RESONANCE(W)IMAG? OR MAGNETIC(W)RESONANCE(W)SPECTROSCOP? OR MAGNETOENCEPHALOGRAPH? OR MAGNETIC(W)SOURCE(W)IMAG? OR HIGH(W)DENSITY(W)SCALP(W)EEG OR HIGH(W)DENSITY(W)SCALP(W)ELECTROENCEPHALOGRA? OR HIGH(W)DENSITY(W)VIDEO(W)EEG OR HIGH(W)DENSITY(W)VIDEO(W)ELECTROENCEPHALOGRA? OR HIGH(W)DENSITY(W)EEG OR HIGH(W)DENSITY(W)ELECTROENCEPHALOGRA?)/TI,AB,DE

3. 16171 (MR(W)TOMOGRAPH? OR MR(W)SCAN? OR MR(W)IMAG? OR MR(W)SPECTROSCOP?)/TI,AB,DE

4. 168752 (FMRI OR MRI OR NMR OR MRS OR MRSI OR HMRS OR “(1)HMRS” OR 1HMRS OR IHMRS OR HMRSI OR MEG OR MSI OR HD-EEG OR HD(W)EEG OR HDEEG)/TI,AB,DE

5. 18879 (SPECT OR SINGLE(W)PHOTON(W)EMISSION(W)COMPUTED(W)TOMOGRAPH?)/TI,AB,DE

6. 43855 (PET OR POSITRON(W)EMISSION(W)TOMOGRAPH? OR SUBTRACTION(W)ICTAL(W)SPECT OR SUBTRACTION(W)ICTAL(W)SINGLE(W)PHOTON(W)EMISSION(W)COMPUTED(W)TOMOGRAPH? OR SISCOM)/TI,AB,DE

7. 3 VOLUMETRIC(W)EXAM?/TI,AB,DE

8. 3 QUANTITATIVE(W)T2(W)MEASURE?/TI,AB,DE 

9. 1344 VOXEL(W)BASED(W)MORPHOMET?/TI,AB,DE

10. 2319 (DIFFUSION(W)WEIGHT?(W)IMAG? OR DWI)/TI,AB,DE

11. 2854 (DIFFUSION(W)TENSOR(W)IMAG? OR DTI)/TI,AB,DE

12. 907 (FLUID(W)ATTENTUATED(W)INVERSION(W)RECOVERY OR FLAIR)/TI,AB,DE

13. 279723 S2:S12

14. 9942 S1 AND S13

15. 630376 (SURGER? OR SURGIC? OR NEUROSURGER? OR NEUROSURGIC? OR LOBECTOM? OR CORTICAL(W)RESECTION OR HEMISPHEROTOM? OR HEMISPHERECTOM? OR CALLOSOTOM? OR CALLOTOM?)/TI,AB,DE

16. 3241 S14 AND S15

17. 1589884 PY = 2008:2011

18. 623 S16 AND S17

Key:

• ? = truncation

• /TI,AB,DE = terms in title, abstract, or descriptor fields

• (W) = terms adjacent to each other (same order)

• PY = publication year

• : = range e.g. PY = 2008:2011 means year = 2008 OR 2009 OR 2010 OR 2011

Interface used: BIOSIS Previews 1969–2008 (via Web of Knowledge)

• Search date: 7 September 2010.

• Records identified: 1164.

• Restrictions: 2003–8.

1. # 11 1,164 #10 AND #9

2. # 10 > 100,000 TS = (surger* OR surgic* OR neurosurger* OR neurosurgic* OR lobectom* OR hemispherotom* OR hemispherectom* OR callosotom* OR callotom*) OR TS = (“cortical resection*”)

3. # 9 9,168 #8 AND #1

4. # 8 > 100,000 #7 OR #6 OR #5 OR #4 OR #3 OR #2

5. # 7 5,297 TS = (“volumetric exam*”) OR TS = (“quantitative t2 measure*”) OR TS = (“voxel based morphomet*”) OR TS = (“diffusion weight* imag*”) OR TS = (dwi) OR TS = (“diffusion tensor imag*”) OR TS = (dti) OR TS = (“fluid attenuated inversion recovery*”) OR TS = (flair) # 6 45,284 TS = (pet) OR TS = (“positron emission tomograph*”) OR TS = (“subtraction ictal spect”) OR TS = (“subtraction ictal single photon emission computed tomograph*”) OR TS = (siscom)

6. # 5 22,822 TS = (spect) OR TS = (“single photon emission computed tomograph*”)

7. # 4 > 100,000 TS = (fmri OR mri OR nmr OR mrs OR mrsi OR hmrs OR 1hmrs OR ihmrs OR hmrsi OR meg OR msi OR hdeeg) OR TS = (“(1)hmrs”) OR TS = (“hd-eeg”) OR TS = (“hd eeg”)

8. # 3 18,462 TS = (“mr tomograph*”) OR TS = (“mr scan*”) OR TS = (“mr imag*”) OR TS = (“mr spectroscop*”)

9. # 2 > 100,000 TS = (“magnetic resonance tomograph*”) OR TS = (“magnetic resonance scan*”) OR TS = (“magnetic resonance imag*”) OR TS = (“magnetic resonance spectroscop*”) OR TS = (magnetoencephalogra*”) OR TS = (“magnetic source imag*”) OR TS = (“high density scalp eeg”) OR TS = (“high density scalp electroencephalogra*”) OR TS = (“high density video eeg”) OR TS = (“high density electroencephalogra*”) OR TS = (“high density eeg”) OR TS = (high density electroencephalogra*”)

10. # 1 86,724 TS = epilep*

Key:

• TS = topic tag; searches terms in Title, Abstract, Author Keywords and Keywords Plus fields

• * = truncation

• ? = embedded truncation

• “ “ = phrase search

• near/1 = terms within one word of each other (any order)

• near/2 = terms within two words of each other (any order)

Interface used: www.thecochranelibrary.com/ Wiley, The Cochrane Library website Issue 7 of 12, July 2010

• Search date: 30 July 2010.

• Records identified: 22.

• Restrictions: 2003 to present.

This search covered the following databases:

• CENTRAL (Cochrane Central Register of Controlled Trials)

• CDSR (Cochrane Database of Systematic Reviews)

• DARE (Database of Abstracts of Reviews of Effectiveness)

• HTA (Health Technology Assessment database)

• NHS-EED (NHS Economic Evaluation Database)

1. #1 MeSH descriptor Epilepsy explode all trees (1975)

2. #2 epilep*:ti,ab (3106)

3. #3 (#1 OR #2) (3787)

4. #4 MeSH descriptor Magnetic Resonance Imaging explode all trees (3817)

5. #5 MeSH descriptor Magnetic Resonance Spectroscopy explode all trees (367)

6. #6 MeSH descriptor Magnetoencephalography, this term only (113)

7. #7 ((magnetic NEXT resonance NEXT tomograph*) OR (magnetic NEXT resonance NEXT scan*) OR (magnetic NEXT resonance NEXT imag*) OR (magnetic NEXT resonance NEXT spectroscop*) OR magnetoencephalograph* OR (magnetic NEXT source NEXT imag*) OR (high NEXT density NEXT scalp NEXT eeg) OR (high NEXT density NEXT scalp NEXT electroencephalogra*) OR (high NEXT density NEXT video NEXT eeg) OR (high NEXT density NEXT video NEXT electroencephalogra*) OR (high NEXT density NEXT eeg) OR (high NEXT density NEXT electroencephalogra*)):ti,ab (2734)

8. #8 ((mr NEXT tomograph*) OR (mr NEXT scan*) OR (mr NEXT imag*) OR (mr NEXT spectroscop*)):ti,ab (705)

9. #9 ((fmri OR mri OR mrs OR mrsi OR hmrs OR ihmrs OR hmrsi OR meg OR msi OR hdeeg)):ti,ab (3115)

10. #10 1hmrs:ti,ab (0)

11. #11 “(1)hmrs”:ti,ab (1)

12. #12 “hd-eeg”:ti,ab (0)

13. #13 (hd NEXT eeg):ti,ab (0 )

14. #14 MeSH descriptor Tomography, Emission-Computed explode all trees (2067)

15. #15 (spect OR (single NEXT photon NEXT emission NEXT computed NEXT tomograph*)):ti,ab (948)

16. #16 (pet OR (positron NEXT emission NEXT tomograph*) OR (subtraction NEXT ictal NEXT spect) OR (subtraction NEXT ictal NEXT single NEXT photon NEXT emission NEXT computed NEXT tomography*) OR siscom):ti,ab (1772)

17. #17 (volumetric NEXT exam*):ti,ab (0)

18. #18 (quantitative NEXT t2 NEXT measure*):ti,ab (0)

19. #19 (voxel NEXT based NEXT morphomet*):ti,ab (33)

20. #20 ((diffusion NEXT weight* NEXT imag*) OR dwi):ti,ab (107)

21. #21 ((diffusion NEXT tensor NEXT imag*) OR dti):ti,ab (73)

22. #22 ((fluid NEXT attenuated NEXT inversion NEXT recovery) or flair):ti,ab (29)

23. #23 (#4 OR #5 OR #6 OR #7 OR #8 OR #9 OR #10 OR #11 OR #12 OR #13 OR #14 OR #15 OR #16 OR #17 OR #18 OR #19 OR #20 OR #21 OR #22) (8804)

24. #24 (#3 AND #23) (159)

25. #25 MeSH descriptor Neurosurgical Procedures, this term only (284)

26. #26 MeSH descriptor Anterior Temporal Lobectomy, this term only (14)

27. #27 MeSH descriptor Cerebral Decortication explode all trees (4)

28. #28 MeSH descriptor Craniotomy explode all trees (268)

29. #29 MeSH descriptor Split-Brain Procedure, this term only (1)

30. #30 (surger* OR surgic* OR neurosurger* OR neurosurgic* OR lobectom* OR cortical resection OR hemispherotom* OR hemispherectom* OR callosotom* OR callotom*):ti,ab (59107)

31. #31 (#25 OR #26 OR #27 OR #28 OR #29 OR #30) (59,269)

32. #32  (#24 AND #31), from 2003 to 2010 (22)

Key:

• MeSH descriptor = indexing term (MeSH heading)

• * = truncation

• “ “ = phrase search

• :ti,ab = terms in either title or abstract fields

• near/1 = terms within one word of each other (any order)

• near/2 = terms within two words of each other (any order)

• next = terms are next to each other.

Interface used: Ovid EMBASE – 1980 to 2010 week 28

• Search date: 23 July 2010.

• Records identified: 3015.

• Restrictions: 2003 to present.

1. exp Epilepsy/ (124,629)

2. epilep$.ti,ab. (95,124)

3. 1 or 2 (139620)

4. exp Nuclear Magnetic Resonance Imaging/ (339,416)

5. exp Nuclear Magnetic Resonance Spectroscopy/ (74,276)

6. Magnetoencephalography/ (5369)

7. (magnetic resonance tomograph$ or magnetic resonance scan$ or magnetic resonance imag$ or magnetic resonance spectroscop$ or magnetoencephalograph$ or magnetic source imag$ or high density scalp eeg or high density scalp electroencephalogra$ or high density video eeg or high density video electroencephalogra$ or high density eeg or high density electroencephalogra$).ti,ab. (128608)

8. (mr tomograph$ or mr scan$ or mr imag$ or mr spectroscop$).ti,ab. (44,000)

9. (fmri or mri or nmr or mrs or mrsi or hmrs or “(1)hmrs” or 1hmrs or ihmrs or hmrsi or meg or msi or hd-eeg or hd eeg or hdeeg).ti,ab. (252,652)

10. Computer Assisted Emission Tomography/ (9435)

11. Positron Emission Tomography/ (54,025)

12. (spect or single photon emission computed tomograph$).ti,ab. (23,752)

13. (pet or positron emission tomograph$ or subtraction ictal spect or subtraction ictal single photon emission computed tomograph$ or siscom).ti,ab. (57,202)

14. volumetric exam$.ti,ab. (11)

15. quantitative t2 measure$.ti,ab. (11)

16. voxel based morphomet$.ti,ab. (1543)

17. (diffusion weight$ imag$ or dwi).ti,ab. (5092)

18. (diffusion tensor imag$ or dti).ti,ab. (4233)

19. (fluid attenuated inversion recovery or flair).ti,ab. (2634)

20. or/4-19 (599,154)

21. 3 and 20 (18,633)

22. Neurosurgery/ (32,497)

23. Temporal Lobectomy/ (2193)

24. Decortication/ (1356)

25. Hemispherectomy/ (825)

26. Craniotomy/ (13,211)

27. Split-Brain/ (286)

28. (surger$ or surgic$ or neurosurger$ or neurosurgic$ or lobectom$ or cortical resection or hemispherotom$ or hemispherectom$ or callosotom$ or callotom$).ti,ab. (1,168,161)

29. or/22-28 (1,189,172)

30. 21 and 29 (5147)

31. exp Animal/ not (Human/ and exp Animal/) (1,235,605)

32. 30 not 31 (5145)

33. limit 32 to yr = “2003 -Current” (3015)

Key:

• / = indexing term (EMTREE heading)

• exp = exploded EMTREE heading

• $ = truncation

• ? = embedded truncation

• .ti,ab. = terms in either title or abstract fields.

Interface used: DialogClassic Web FILE 65: Inside Conferences –1993–2010/September 2006 (c) 2010 BLDSC all rts reserv.

• Search date: 7 September 2010.

• Records identified: 15.

• Restrictions: 2003 to present.

1. 6064 EPILEP?/TI,AB,DE

2. 4100 (MAGNETIC(W)RESONANCE(W)TOMOGRAPH? OR MAGNETIC(W)RESONANCE(W)SCAN? OR MAGNETIC(W)RESONANCE(W)IMAG? OR MAGNETIC(W)RESONANCE(W)SPECTROSCOP? OR MAGNETOENCEPHALOGRAPH? OR MAGNETIC(W)SOURCE(W)IMAG? OR HIGH(W)DENSITY(W)SCALP(W)EEG OR HIGH(W)DENSITY(W)SCALP(W)ELECTROENCEPHALOGRA? OR HIGH(W)DENSITY(W)VIDEO(W)EEG OR HIGH(W)DENSITY(W)VIDEO(W)ELECTROENCEPHALOGRA? OR HIGH(W)DENSITY(W)EEG OR HIGH(W)DENSITY(W)ELECTROENCEPHALOGRA?)/TI,AB,DE

3. 1918 (MR(W)TOMOGRAPH? OR MR(W)SCAN? OR MR(W)IMAG? OR MR(W)SPECTROSCOP?)/TI,AB,DE

4. 58296 (FMRI OR MRI OR NMR OR MRS OR MRSI OR HMRS OR “(1)HMRS” OR 1HMRS OR IHMRS OR HMRSI OR MEG OR MSI OR HD-EEG OR HD(W)EEG OR HDEEG)/TI,AB,DE

5. 2185 (SPECT OR SINGLE(W)PHOTON(W)EMISSION(W)COMPUTED(W)TOMOGRAPH?)/TI,AB,DE

6. 8857 (PET OR POSITRON(W)EMISSION(W)TOMOGRAPH? OR SUBTRACTION(W)ICTAL(W)SPECT OR SUBTRACTION(W)ICTAL(W)SINGLE(W)PHOTON(W)EMISSION(W)COMPUTED(W)TOMOGRAPH? OR SISCOM)/TI,AB,DE

7. 0 VOLUMETRIC(W)EXAM?/TI,AB,DE

8. 0 QUANTITATIVE(W)T2(W)MEASURE?/TI,AB,DE

9. 15 VOXEL(W)BASED(W)MORPHOMET?/TI,AB,DE

10. 236 (DIFFUSION(W)WEIGHT?(W)IMAG? OR DWI)/TI,AB,DE

11. 463 (DIFFUSION(W)TENSOR(W)IMAG? OR DTI)/TI,AB,DE

12. 161 (FLUID(W)ATTENTUATED(W)INVERSION(W)RECOVERY OR FLAIR)/TI,AB,DE

13. 74927 S2:S12

14. 489 S1 AND S13

15. 65503 (SURGER? OR SURGIC? OR NEUROSURGER? OR NEUROSURGIC? OR LOBECTOM? OR CORTICAL(W)RESECTION OR HEMISPHEROTOM? OR HEMISPHERECTOM? OR CALLOSOTOM? OR CALLOTOM?)/TI,AB,DE

16. 58 S14 AND S15

17. 2658053 PY = 2003:2011

18. 15 S16 AND S17

Key:

• ? = truncation

• /TI,AB,DE = terms in title, abstract, or descriptor fields

• (W) = terms adjacent to each other (same order)

• PY = publication year

• : = range e.g. PY = 2008:2011 means year = 2008 OR 2009 OR 2010 OR 2011.

Interface used: http://apps.who.int/trialsearch/AdvSearch.aspx

• Search date: 7 September 2010.

• Records identified: 7.

• Epilepsy [in condition field]

• AND

• (magnetic OR magnetoencephalography OR density OR tomography OR electroencephalography OR scan OR scanning OR image) [in intervention field]

Study details