Multiple-frequency bioimpedance devices for fluid management in people with chronic kidney disease receiving dialysis: a systematic review and economic evaluation

Body Composition Monitor

The BCM is a portable, stand-alone device, which uses bioimpedance spectroscopy to estimate a person’s fluid and nutritional status. The person is placed in a supine position and four electrodes are attached: two to the back of one hand and two to the foot on the same side of the body. The electrodes are connected to the BCM device via a cable. The device passes a painless alternating current at 50 different frequencies (5–1000 kHz) through the body and measures the impedance between the hand and foot, giving relative impedance values for each frequency. This range of measurements determines the electrical resistances of the total body water and ECW and allows distinction of ECW and ICW. 27,56 The software also calculates fluid overload using two physiological models. The volume of ECW that should be present based on the identified amounts of lean and adipose tissue is calculated and compared with the measured volume of extracellular fluid. 57,58 The resulting volume difference between predicted and actual extracellular fluid is used as a measure of a person’s overhydration volume and is reported by the device in litres.

The BCM is intended to be used as an objective measure of fluid imbalance, to complement clinical judgement. The associated software uses two validated physiological models to obtain the clinically relevant parameters: overhydration, lean tissue mass and adipose tissue mass. 4,56 There are no restrictions on the age of the person that this device can be used on. Results from the BCM are available within 2 minutes and are stored on a ‘PatientCard’ automatically, from which it can be loaded onto a database. Cards are reusable and can be reprogrammed for a new patient, or can have a patient’s data deleted if they become full, and remain programmed for that patient.

Good agreement has been shown between BCM assessment and current standard methods for measuring ECW and total body volumes, ICW volume, total fat, fat-free mass and fluid overload in adults and urea distribution volume in children. 24,59 The evidence of association between BCM assessment and improved patient outcomes is mixed. The Canadian Agency for Drugs and Technologies in Health (CADTH)’s rapid-response report, published in 2015,53 identified two randomised controlled trials (RCTs) of 131 and 189 participants, respectively60,61 and one observational study of 110 participants, which assessed the use of the BCM in people receiving HD. 62 The report concluded that there was improvement in some patient outcomes, such as decreased blood pressure and reduced fluid overload, with patient management guided by BCM assessments, but that the evidence base was limited. A study of people receiving PD compared assessment of overhydration status using the BCM with assessment using a standard protocol. Results showed that ECW volume and ECW-to-ICW volume ratio decreased steadily over the 3-month follow-up period in the group assessed using the BCM, but increased in the group assessed using standard methods. In addition, systolic blood pressure (SBP) decreased significantly in the group assessed using the BCM, but increased significantly in the group assessed using standard methods. 63

Further information on the BCM is available from the manufacturer’s website. 56

MultiScan 5000

The MultiScan5000 is a portable device that uses bioimpedance spectroscopy to measure at 50 frequencies (ranging from 5 kHz to 1000 kHz), which are used to calculate body composition and hydration by a mathematical model called Cole–Cole analysis (also used in the BCM models). Values for ECW, ICW, total body water, and volume of over/underhydration are obtained from similar physiological models as used in the BCM. 57,58

The volume of overhydration output is recommended for the assessment of hydration status in people aged 18–70 years. Outside this age range, this output can be used to track relative changes over time. In addition, the ratio of total body water to ECW volume calculated by the device (called the ‘prediction marker’) can be used as an additional marker to track hydration status over time in all age groups. The device can measure body segments, depending on the placement of the electrodes,64 and provides a bioelectrical impedance vector analysis. Additional parameters related to body composition, such as fat weight, lean weight, skeletal muscle mass and body cell mass, can also be estimated. These parameters can be used to estimate nutritional status and, therefore, help to identify malnutrition status in people with CKD who are treated with dialysis. Further information on the MultiScan 5000 device can be found on the product webpage. 64

BioScan 920-II

The BioScan 920-II is a portable multiple-frequency bioimpedance analysis device, which measures at 5, 50, 100 and 200 kHz. The eight electrodes allow monitoring of fluid changes in the whole body, thorax, trunk, legs or arms. All data are recorded and displayed immediately for analysis by the system. Alongside the standard output parameters related to hydration status [target water (minimum/maximum), target weight, target weight (minimum/maximum), extracellular fluid, ECW volume, ICW volume, total body water, ECW (%), ICW (%), total body water (%), ECW-to-ICW volume ratio, plasma fluid (intravascular), fat-free mass hydration], the device estimates additional parameters related to body composition [comprising body mass index (BMI), body density, body cell mass, protein mass, fat mass, fat-free mass and glycogen mass] and mineral content.

These parameters can be used to evaluate nutritional status and help to identify malnutrition in people with CKD receiving dialysis. Further information can be found on the product webpage. 65 The use of the BioScan 920-II is recommended for people aged 5–99 years. A version of the BioScan 920-II device (the BioScan 920-II-P) is also available for monitoring hydration status in preterm, neonatal and paediatric patients (for use from 23 weeks’ gestational age up to 18 years).

BioScan touch i8

According to the manufacturer, an updated version of the BioScan 920-II device, the BioScan touch i8 with an updated user interface, is due to be released during the course of this assessment. As with the BioScan 920-II, it is anticipated that there will be two versions: one suitable for people aged 0–18 years and one suitable for people aged 5–99 years.

InBody S10

The InBody S10 is a portable device that uses a direct multiple-frequency bioimpedance analysis method to provide measurements across six different frequencies (1, 5, 50, 250, 500 and 1000 kHz). Measurements of five segments of the body are available: right arm, left arm, trunk, right leg and left leg. Hydration-related outputs include water volumes (ECW, ICW), ratio of extracellular to total body water and history of body water condition.

These parameters are estimated along with a suggested standard range of values to facilitate identification of overhydrated or underhydrated individuals. In addition, the InBody S10 provides estimates related to body composition such as body cell mass, basal metabolic rate, bone mineral content, skeletal muscle mass, fat-free mass, and BMI. These parameters can be used to evaluate nutritional status and help to identify malnutrition in people with CKD who are on dialysis. A full list of outputs can be found on the product webpage. 66 The use of the InBody S10 device is recommended for people aged 3–99 years.

Population

People with CKD treated with HD or PD.

Interventions

The multiple-frequency bioimpedance devices considered in this assessment were:

• BCM

• MultiScan 5000

• BioScan 920-II and BioScan touch i8

• InBody S10.

Comparator

The comparator considered in this assessment was standard clinical assessment, which takes account of the following parameters:

• blood pressure

• presence of oedema

• changes in weight

• residual renal function

• pre-existing CV conditions

• any patient-reported symptoms of overhydration or underhydration, for example cramps, fatigue, nausea, dizziness, breathlessness, decreased appetite or visual disturbances.

Outcomes

The following outcome measures were considered:

• intermediate measures, including –

  • number and length of HD sessions

  • number of unplanned hospital visits/admissions as a result of fluid overload or dehydration

  • use of antihypertensive medication

  • incidence of anaemia

  • blood pressure

  • left ventricular hypertrophy

  • left ventricular mass index (LVMI)

  • arterial stiffness

  • incidence of overhydration or underhydration

  • changes of dialysis modality (from PD to HD) because of fluid overload

  • adherence with recommended fluid intake.

• clinical outcomes, including –

  • incidence of CV events (including stroke and heart attack)

  • mortality

  • residual renal function

  • incidence of oedema

  • incidence of peritonitis

  • adverse effects associated with hypotensive episodes (including cramps, fatigue, diarrhoea, nausea, dizziness and fainting).

• patient-reported outcomes, including –

  • post-dialysis recovery time and fatigue

  • health-related quality of life.

One further relevant outcome not specified in the scope or protocol was also considered because of its clinical importance: achievement of target weight.

Randomised controlled trials

All five included RCTs were available in full-text format. 60,61,63,76,77 The BCM was the multiple-frequency device used in all five trials. One trial was conducted in Romania,60 one trial in Taiwan,76 one in Turkey77 and one in Portugal,61 and the remaining trial did not provide this information. 63 One trial recruited patients from 23 dialysis centres,61 one trial recruited from two dialysis centres,77 another trial from six dialysis centres76 and one trial recruited from a single dialysis centre. 60 In the remaining trial, it was unclear whether patients were recruited from a single dialysis centre or from multiple dialysis centres. 63 Four trials enrolled solely patients who were treated with HD,60,61,76,77 and one trial enrolled continuous ambulatory PD patients. 63 All five trials involved dialysis in a hospital setting. The multiple-frequency bioimpedance device used for assessment of fluid status by all five trials was the BCM. All five trials included only adults aged ≥ 18 years. 60,61,63,76,77 The main exclusion criteria reported in the trials, which assessed patients receiving HD, were coronary stents or pacemakers;76,77 metallic devices in the body, such as joint prostheses;60,61,76,77 limb amputations;60,76,77 and pregnancy. 60,61 One trial, which assessed PD patients,63 excluded those who had been on one or two exchanges per day because of economic limitation and those patients with acute infection and CV events in the month prior to enrolment.

The length of follow-up of the included trials ranged from 3 months61 to 2.5 years,60 with two trials reporting a follow-up period of 12 months. 61,77 In the case of the trial by Luo et al. ,63 the authors decided to terminate follow-up at 3 months rather than at 6 months, as originally planned, as the emerging differences between the groups and the adverse effect of fluid overload led the decision to extend the follow-up period to be considered unethical.

Three of the five trials had links to Fresenius Medical Care, the company that manufactures the BCM,60,61,77 albeit two of these trials reported that Fresenius had no involvement in the design or conduct of the trial. 60,77 Two trials were supported by grants from independent sources. 63,76

Randomised controlled trials

The five RCTs60,61,63,76,77 randomised a total of 939 participants: 469 to bioimpedance measurements and 470 to standard clinical assessment.

The mean age for each intervention group was reported in all five RCTs60,61,63,76,77 and ranged from 50.9 years77 to 65.8 years61 in the bioimpedance intervention group and from 52.5 years77 to 66.7 years61 in the standard clinical assessment group.

The five RCTs each reported the proportion of males and females for each intervention group. 60,61,63,76,77 Study populations tended to involve approximately equal proportions of men and women, with the exception of the studies by Hur et al. 77 (69% men) and Ponce et al. 61 (76.2% men). The proportion of men ranged from 43.6%63 to 71.3%61 in the bioimpedance intervention group, and from 48.8%63 to 81.8%61 in the standard clinical assessment group. The prevalence of diabetes mellitus among participants varied across trials. The proportion of participants with diabetes mellitus was reported by all five trials60,61,63,76,77 and ranged from 10%60 to 39.8%61 in the bioimpedance intervention group, and from 9%60 to 38.6%61 in the standard clinical assessment group. The mean dialysis vintage was reported in three RCTs60,63,77 and ranged from 35.2 months63 to 107 months57 in the bioimpedance assessment group, and from 33.2 months to 104 months60 in the control group.

Non-randomised studies

The eight included non-randomised cohort studies assessed a total of 4915 participants. 30,50,82,83,85–88 The studies were of two main types: in some studies, the BCM was used to classify patients into groups (e.g. overhydrated/non-overhydrated) and then outcomes were compared across the groups;30,88 and in other studies, the BCM was used as a basis for adjustment of dry weight50,82,85,86 or to obtain hydration parameters. 83,87 Six cohort studies reported the mean age of participants, which ranged from 53.8 to 68.2 years. 30,50,85–88 The two remaining cohort studies reported the median ages of participants of 57.9 years83 and 57 years. 82 Three studies reported the mean age for normohydrated and overhydrated groups. 30,50,85 The age range was 55.9 years85 to 66 years30 for the normohydrated groups and 58.4 years85 to 65.6 years50 for the overhydrated groups, respectively. The proportion of men in the seven studies reporting this information50,82,83,85–88 ranged from 52.5%88 to 64.7%87 and was, in general, higher than reported in the included RCTs. The proportion of participants with diabetes mellitus was reported by six of the observational studies30,82,83,86–88 and ranged from 10.4%88 to 37%. 83 The mean dialysis vintage was reported by half of the studies and ranged from 10.7 months87 to 66 months. 85 In the study by Hoppe et al. ,87 participants were split into those with a short-dialysis vintage (≤ 24 months) or a long-dialysis vintage (> 24 months), with mean dialysis vintage being 9.3 weeks and 76.2 weeks, respectively. The trial by Kim et al. 85 reported mean dialysis vintage separately for dehydrated, normohydrated and hyperhydrated participants, which was 6.0, 4.1 and 5.7 years, respectively.

Randomised controlled trials

The frequency of measurements using the BCM in the RCTs was at least every 3 months. The most frequent use of the device was twice monthly in the bioimpedance intervention group (and every 3 months in the control group). 77 Three-month assessments were reported by Onofriescu et al. ;60 monthly assessments by Huan-Sheng et al. 76 and Ponce et al. 61 and 6-week assessments by Luo et al. 63

Non-randomised studies

The frequency of BCM assessments in the non-randomised studies varied across studies: one study involved only one assessment within the first week of dialysis;50 two studies involved three assessments per week;30,88 another study involved weekly assessments;87 two other studies involved monthly assessments;85,86 one study involved quarterly assessments;82 and the remaining study did not report the frequency of the BCM use. 83

Randomised controlled trials

Figure 3 presents the summary of risk-of-bias assessments for all included trials. Risk-of-bias assessments of individual studies are presented in Figure 4.

FIGURE 3.

Summary of risk-of-bias assessments for all included trials.

FIGURE 4.

Risk-of-bias assessments of individual studies.

According to the prespecified criteria for the assessment of the overall risk of bias, one of the five RCTs was rated as being at a high risk of bias,63 and the remaining four trials did not provide sufficient information on which to make a robust judgement. 60,61,76,77

Selection bias

Two trials reported sufficient information on which to make a full assessment of selection bias. Full details of allocation concealment were not reported, but the method of generation of sequence (i.e. random generation by computer) implies that the study personnel would be unable to predict the allocation, thus fulfilling the criterion of low risk. The trial by Ponce et al. 61 involved randomisation of centres, as opposed to randomisation of individuals within centres. No details of the randomisation process were reported. The remaining two trials merely stated that they were randomised trials, but provided no details of how randomisation was achieved. 60,63,77

Performance and detection bias

Only one trial reported that participants were blinded. 60 One trial reported that participants were not blinded. 63 In the remaining three trials, both the intervention group and the control group received BCM assessments but the measurements were used to assess the intervention group only. 61,76,77

Two trials reported that outcome assessors were blinded. 61,77 Luo et al. 63 reported that patients, investigators and dialysis staff were not blinded to treatment assignment; the trial was, therefore, rated as being at a high risk of bias for the blinding of outcome assessors domain and for overall risk. In the trials by Huan-Sheng et al. 76 and Onofriescu et al. ,60 it was unclear whether or not outcome assessors had been blinded.

Attrition bias

One trial reported a low number of dropouts63 and was, therefore, rated as being at a low risk of attrition bias. The remaining four studies were rated as being at a high risk of bias because of the high number of participants who dropped out. 60,61,76,77 It is worth noting that, in the Ponce et al. 61 trial, 29 out of 101 (28.7%) and 42 out of 88 (47.7%) discontinuations were observed in the intervention and control groups, respectively. The reasons given for terminating the trial prematurely were as follows: ‘no valid data available within the time frame, death, transplant or transfer to another clinic’. The proportion of participants within each of these categories and distribution of dropouts across centres were, however, not given.

Reporting bias

In four of the five included trials, the outcomes reported were in accordance with those specified in the respective methods section. 60,61,63,76 The trial by Hur et al. 77 was rated as being at a high risk of reporting bias, as some outcome measures that had not been previously specified were reported, such as iron dose, right ventricle end-diastolic diameter, urine output, triglyceride levels and cholesterol.

Other bias

Three RCTs reported links with Fresenius Medical Care (either in the form of funding, as an honorary speaker or through employment) and were rated as being at a high risk of ‘other bias’. 60,61,77 Two trials were supported by grants from independent sources. 63,76 No other sources of bias were apparent in the included trials.

Non-randomised studies

Figure 5 presents a summary of the risk-of-bias assessments for the included non-randomised cohort studies. The results of individual study-level assessments are presented in Appendix 8.

FIGURE 5.

Summary risk of bias for non-randomised cohort studies.

The majority of studies identified important prognostic factors, provided information on non-respondents/dropouts, included a sufficient length of follow-up, used objective outcome measures, considered important outcomes, delivered the intervention in an appropriate setting and by an experienced person, clearly defined the intervention, collected data prospectively, clearly defined the inclusion/exclusion criteria and involved a representative sample. None of the studies involved blinding of participants or study personnel. Two studies enrolled participants who entered the study at varying points in their disease progression. The study by Hoppe et al. 87 compared short- versus long-dialysis vintage and the study by O’Lone et al. 82 compared incident and prevalent patients. The majority of studies failed to provide information on the characteristics of participants who withdrew or did not complete follow-up. 30,50,82,83,85,88

Blood pressure

Five trials (one rated as being at a high risk of bias and four as being at an unclear risk of bias) reported SBP measurements, which were included in a meta-analysis. 60,61,63,76,77 Figure 6 shows that SBP was lower in participants who underwent bioimpedance measurements using the BCM device than in those assessed by standard clinical assessment, but the difference was not statistically significant (mean difference –2.46 mmHg, 95% CI –5.07 to 0.15 mmHg; p = 0.06, I² = 0%).

FIGURE 6.

Meta-analysis for SBP. df, degrees of freedom; IV, inverse variance; SD, standard deviation.

Arterial stiffness

Two trials (both rated as being at an unclear risk of bias) reported arterial stiffness results, which were included in a meta-analysis. 60,77 The measurement of pulse wave velocity (PWV) is generally accepted as the most simple, non-invasive, robust and reproducible method of determining arterial stiffness, with carotid–femoral PWV regarded as the gold standard. The PWV increases from 4–5 m/second in the ascending aorta to 5–6 m/s in the abdominal aorta and 8–9 m/s in the iliac and femoral arteries. 89 Normal values, using standardised calculation of PWV, are a mean of 7.2 m/s in people aged 40–49 years and 8.3 m/s in people aged 50–59 years. 90 Figure 7 shows that arterial stiffness (as assessed by carotid–femoral PWV) was lower, but not statistically significantly lower, in the bioimpedance assessment group than that in the standard clinical assessment group (mean difference –1.18 m/s, 95% CI –3.14 to 0.78 m/s; p = 0.24, I² = 92%). Substantial statistical heterogeneity between trials was observed.

FIGURE 7.

Meta-analysis for arterial stiffness. df, degrees of freedom; IV, inverse variance; SD, standard deviation.

Mortality

Three of the included trials (all rated as being at an unclear risk of bias) reported mortality data. 60,61,76 As mortality was reported with a HR, the log-HR and log-SE for the three trials were input manually (Table 3).

A total of 19 out of 311 (6.1%) participants in the bioimpedance assessment group and 23 out of 307 (7.5%) participants in the standard clinical assessment group died. Figure 8 shows that, compared with standard clinical assessment, the use of the BCM had no significant effects on mortality (HR 0.69, 95% CI 0.23 to 2.08; p = 0.51, I² = 54%). Moderate statistical heterogeneity was evident among trials.

FIGURE 8.

Meta-analysis for mortality. df, degrees of freedom; IV, inverse variance.

Absolute overhydration

Four trials (one rated as being at a high risk of bias and three rated as being at an unclear risk of bias) assessed absolute overhydration,61,63,76,77 which was defined as the difference between expected ECW and actual ECW. No data on underhydration were available. Figure 9 shows that absolute overhydration was significantly lower in the BCM assessment group than in the standard clinical assessment group [weighted mean difference (WMD) –0.44, 95% CI –0.72 to –0.15; p = 0.003, I² = 49%]. Moderate statistical heterogeneity between trials was apparent.

FIGURE 9.

Meta-analysis for absolute overhydration. df, degrees of freedom; IV, inverse variance; SD, standard deviation.

Relative overhydration

Three trials (all rated as being at an unclear risk of bias) reported data on ROH,60,61,76 which was defined as the ratio of absolute fluid overload to ECW volume. Figure 10 shows that ROH was significantly lower in the BCM assessment group than in the standard clinical assessment group (WMD –1.84, 95% CI –3.65 to –0.03; p = 0.05, I² = 52%). ROH was assessed by the BCM in both groups, therefore these results should be interpreted with caution.

FIGURE 10.

Meta-analysis of ROH. df, degrees of freedom; IV, inverse variance; SD, standard deviation.

Intermediate reported outcomes

Incidence of cardiovascular events

One study reported a combination of acute fluid overload or CV-related events, which included hospitalisation related to CV or cerebrovascular events and episodes of acute fluid overload. Huan-Sheng et al. 76 (rated as being at an unclear risk of bias) reported 14 events in the bioimpedance assessment group, with an incidence rate of 0.10 (95% CI 0.06 to 0.17) per patient-year, and 28 events in the control group, with an incidence rate of 0.21 (95% CI 0.15 to 0.29) per patient-year. The overall incidence ratio was 0.50 (95% CI 0.26 to 0.94) per patient-year (p = 0.03).

Residual renal function

No trials reported residual renal function, but two studies reported urinary volume, which could be considered a surrogate measure thereof. Hur et al. 77 (rated as being at an unclear risk of bias) reported a significant increase in the proportion of anuric patients and a significant decrease in urine output in non-anuric patients at 12 months in the bioimpedance assessment group. By contrast, there was no change in the proportion of anuric patients in the control group and the decrease in urine output in non-anuric patients was not significant at follow-up. Luo et al. 63 (rated as being at a high risk of bias) reported non-significant decreases in urine volume in both the BCM group and the standard clinical assessment group at 12 weeks, although the bioimpedance assessment group showed a numerically larger decrease.

Adverse effects associated with hypotensive episodes

The top five intradialytic complications reported by Huan-Sheng et al. 76 (rated as being at an unclear risk of bias) were hypotension, cramping, skin itching, chest tightness and headache. There were significant differences between the bioimpedance assessment group and the standard clinical assessment group for all of these complications, but not in the same direction. In the bioimpedance assessment group, there was significantly more cramping, chest tightness and headaches, but significantly less hypotension and skin itching.

Frequency of intradialytic hypotensive events was reported by Hur et al. 77 (rated as being at an unclear risk of bias); there was no difference between groups at baseline (63.2 events/1000 dialysis sessions in the bioimpedance assessment group and 63.8 events/1000 dialysis sessions in the standard clinical assessment group; p = 0.9) or at 12 months (66.6 and 63.9 events/1000 dialysis sessions, respectively; p = 0.4). Similarly, Onofriescu et al. 60 (rated as being at an unclear risk of bias) reported no difference between groups in hypotension, cramps or patient-year (p = 0.6). Ponce et al. 61 (rated as being at an unclear risk of bias) defined hypotensive events as SBP reduced by at least 30 mmHg during dialysis or intradialytically below 90 mmHg, and reported no significant difference between groups at baseline (39 events in 17 patients in the bioimpedance assessment group, and 38 events in 12 patients in the standard clinical assessment group) or 12 months (48 events in 20 patients and 41 events in 15 patients, respectively).

No data were available on incidence of oedema or incidence of peritonitis.

Fatigue

Only one trial reported details of any specified patient-reported outcomes. Huan-Sheng et al. 76 (rated as being at an unclear risk of bias) reported four events of intradialytic fatigue in the bioimpedance assessment group and five events in the standard clinical assessment group. The difference between groups was not statistically significant (p = 0.7).

Achievement of target weight

Three trials reported achievement of target weight. Huan-Sheng et al. 76 (rated as being at an unclear risk of bias) reported that post-dialysis target weight (PDTW) adjustment was performed in 816 months (out of a total of 1658 monthly assessments across the 148 participants in the intervention group over the 12-month follow-up period). PDTW was achieved in 650 of these months (80%). Of the 816 months, clinical signs and symptoms were comparable with the BCM results in 482 months (59%), of which PDTW was reached in 426 months (88%). The authors further reported that PDTW adjustments based on BCM results were not supported by firm and clear clinical evidence in up to 41% of occasions. Onofriescu et al. 60 (rated as being at an unclear risk of bias) stated that a significantly higher proportion of participants in the bioimpedance assessment group than in the control group maintained dry weight within 1.1 kg of the bioimpedance-recommended level. However, there is some uncertainty around the number of participants at each time point, and replicating the analysis was not possible. Ponce et al. 61 (rated as being at an unclear risk of bias) reported that, at 12 months, target weight was generally less overestimated in the BCM assessment group than in the standard clinical assessment group (0.67 vs. 1.00 kg).

Use of antihypertensive medication

Two studies reported the use of antihypertensive medication in specified patient subgroups. 85,86 Castellano et al. 86 reported significantly higher consumption of antihypertensive medications per month in the group with average ROH not reduced within 6 months than in those for whom average ROH was reduced within 6 months. 86 Kim et al. 85 reported no significant difference in the consumption of antihypertensive drugs between dehydrated and hyperhydrated patients, although the number of drugs used at week 16 was significantly lower than that at baseline or week 8 in the hyperhydrated group.

Blood pressure

Four studies reported blood pressure among specified subgroups. 85–88 There were no statistically significant differences between groups in which average overhydration within 6 months was reduced versus not reduced,86 short- versus long-dialysis groups87 or groups in which relative fluid overload was < 17.4% versus > 17.4%. 88 Kim et al. 85 compared the blood pressure of dehydrated and hyperhydrated patients, and found that SBP was higher in the hyperhydrated group, although the statistical significance of the comparison was not reported.

Left ventricular hypertrophy

One study assessed left ventricular hypertrophy87 and showed that the thickness of the left ventricle wall (in mm) was not significantly different for short- versus long-dialysis vintage subgroups.

Hospitalisation

Two studies reported data on hospitalisation. 50,88 Kim et al. 50 reported a non-significant difference between overhydrated and non-overhydrated patients in the number of hospital-days per event. Onofriescu et al. 88 reported a significantly higher all-cause hospitalisation rate for patients classified as overhydrated according to a 17.4% cut-off point than for those classified as being not overhydrated. The value of 17.4% was proposed by the authors as a threshold for classifying a patient as overhydrated (i.e. relative fluid overload of at least 17.4%), as opposed to the value widely accepted in the literature of 15%. In contrast, there was no significant difference between all-cause hospitalisation rates for patients classified as overhydrated according to the traditional 15% threshold and those patients classified as being not overhydrated.

Hydration status

The majority of studies reported hydration status at follow-up, although not in a consistent way. Subgroups in which higher levels of overhydration at follow-up were reported were the subgroup whose average ROH was not reduced to < 15% in 6 months, as compared with the subgroup whose values were reduced to the desired level;86 the long versus short-dialysis vintage subgroup;87 patients with a cardiac cause of death, as opposed to those with a non-cardiac cause of death;83 and both absolute fluid overload and relative fluid overload in subgroups with a relative fluid overload of > 17.4%, as compared with subgroups with a relative fluid overload of < 17.4%. 88

Some studies reported the effects of hydration status on mortality. Kim et al. 50 reported a significant effect of overhydration as a risk factor for death; O’Lone et al. 82 reported a significant effect of absolute overhydration on mortality; Onofriescu et al. 88 reported that patients assessed as being overhydrated were at significantly increased risk for all-cause mortality; and Wizemann et al. 30 reported a significant risk of relative hydration status on mortality.

Cardiovascular events

Three studies reported data on CV events. 50,87,88 A non-significant difference in the incidence of acute myocardial infarction and stroke was observed between short- and long-dialysis vintage subgroups;87 no differences were found in the number of CV events per year between overhydrated and non-overhydrated subgroups;50 and no significant differences in the incidence of coronary heart disease, peripheral vascular disease, heart failure or stroke were detected between the subgroup with a relative fluid overload of < 17.4% and that with a relative fluid overload of > 17.4%. 88

Mortality

One study reported a significantly higher number of deaths in the long versus short-dialysis vintage subgroup. 87

Relevant patient population(s)

The model compared the alternative fluid management strategies for a prevalent cohort of people with ESRD receiving either HD or PD. The base-case analysis was conducted using the weighted average of the median age and sex distribution for the respective prevalent dialysis cohorts, as reported in the UK Renal Registry report:99 aged 67.2 years, 61% male for those receiving HD, and aged 64.2 years, 61% male for those receiving PD. Thus, the base-case analysis was run for a mixed cohort at the average age of 66 years, 61% male, with 87% receiving HD and 13% receiving PD. Separate subgroup analyses were also conducted for the PD and HD cohorts, applying the median ages for the respective subgroups. In addition, comorbidity burden is also used in the model in the estimation of baseline hospitalisation risks, and this was estimated from UK registry data. 99,100 Based on these sources, 63% of patients aged ≥ 65 years and 36% of patients aged < 65 years are modelled to have at least one comorbidity at baseline. The estimated mean number of comorbidities in those with any comorbidity is 1.6 and 2.0 for the PD and HD cohorts, respectively.

Monitoring strategies evaluated

Bioimpedance monitoring strategies, to help adjust target weight and guide fluid management, were compared with standard clinical assessment, in which target weight is set based on clinical signs and symptoms, including blood pressure, presence of oedema, changes in weight, residual renal function, pre-existing CV conditions and patient-reported symptoms of overhydration or underhydration. For the bioimpedance strategies, it was assumed that all patients would have their hydration status assessed every 3 months (four times per year), and have their target weight modified in line with the results if necessary. The above monitoring strategy is in line with clinical opinion regarding the necessary frequency of bioimpedance testing in an adult dialysis population, and is also consistent with the approach used in two of the trials included in the systematic review of clinical effectiveness. 60,81 It is less intensive than the testing strategies applied in the other RCTs included in the clinical effectiveness review, which varied from once per week101 to once every 6 weeks. 63 It is assumed in the base-case cost-effectiveness scenarios that quarterly testing can deliver effects in keeping with those observed across all the included randomised trials. The impact of increased testing frequency is addressed in sensitivity analysis.

The bioimpedance technologies included in the scope for this assessment were the BCM, MultiScan 5000, BioScan 920-II, BioScan touch i8 and the InBody S10. However, the review of existing literature only uncovered clinical effectiveness evidence relating to the BCM. Therefore, the economic modelling focused on assessing the cost-effectives of bioimpedance testing using the BCM device. For comparison, we include cost-per-test estimates using each of the other competitor devices, and assess the impact of applying these costs in a sensitivity analysis (assuming equivalent effects).

Framework (method of synthesis)

A discrete-time Markov cohort model was developed to assess the clinical effectiveness and cost-effectiveness of using multiple-frequency bioimpedance testing compared with standard clinical practice for guiding fluid management decisions in the dialysis cohort. This state-transition framework was chosen for its ability to capture the evolving disease process and recurrent event risks over time, while being relatively parsimonious in terms of data and computational requirements. Key states included in the model are ‘stable on HD’, ‘stable on PD’, ‘post-incident CV event – haemodialysis’, ‘post-incident CV event – peritoneal dialysis’, ‘post transplant’ ‘post transplant, post CV event’, ‘dialysis post transplant’, ‘dialysis post transplant, post CV event’ and ‘death’. The model also includes an option to dichotomise the ‘stable’ and ‘post CV event’ dialysis states by baseline ROH status into either severely overhydrated (> 15% ROH) or normohydrated (≤ 15% ROH), as measured by the BCM. This is to allow mortality and hospitalisation rates for the severely overhydrated portion of the prevalent cohort to be factored upwards, reflecting the observed adjusted association between hydration status and these outcomes. 26,30,60,82

Modelled transitions between the relative hydration states were then used to drive effects in an alternative scenario analysis (see Further adjustments to baseline risks for further details). States representing underhydration were not included in this alternative model structure because of a dearth of evidence on (1) the prevalence of underhydration, as measured by the BCM, in UK dialysis cohorts; (2) the impact of underhydration, as measured by the BCM, on the risk of adverse events and/or quality of life; and (3) the effectiveness of bioimpedance-guided fluid management on reducing the prevalence of underhydration. If underhydration (as measured by bioimpedance spectroscopy) is associated with adverse outcomes and quality of life, and bioimpedance-guided fluid management can reduce the prevalence of this, then this secondary model may fail to capture the associated benefits.

The model simulates mortality, hospitalisation events and transition to transplant over the lifetime of the modelled cohorts on a constant 3-monthly cycle (in keeping with the BCM testing cycle). All-cause hospitalisation events are disaggregated across CV events and other causes. It is assumed in the model that hospitalisation for incident CV events results in an increased comorbidity burden, which increases the risk of subsequent hospitalisations.

Costs of dialysis (by modality), background medication [blood pressure, erythropoiesis-stimulating agents (ESAs)], transplant, all-cause hospitalisation and outpatient attendances are included in the baseline model. Health state utility multipliers are applied to the dialysis states, and utility decrements are also incorporated for hospitalisations. These decrements are applied for an acute period for all hospitalisations. For hospitalisations caused by CV events, a long-term utility multiplier is also applied. This reflects the lasting impact that these events can have on health-related quality of life. A schematic of the model structure is provided in Figure 13. A simplifying assumption of the model precludes switching between dialysis modes. This is unlikely to have a significant impact on results since an equal baseline mortality rate is applied for patients on dialysis irrespective of modality, and the estimated costs of PD and HD were also found to be similar based on current reference costs (see Costs of renal replacement therapy). Furthermore, the clinical effectiveness evidence was insufficient to estimate bioimpedance effects by dialysis modality.

FIGURE 13.

Schematic of the baseline model structure.

The baseline model is replicated for the strategy of bioimpedance-guided fluid management, and correspondingly incorporates the additional cost of quarterly testing on top of standard practice. The bioimpedance model also allows for the incorporation of effects of bioimpedance monitoring on mortality, hospitalisation rates, background management costs (e.g. blood pressure medications) and within-state health state utility. The incorporation of these hypothesised benefits, in light of the available supporting evidence, is discussed in detail under the relevant headings below. The model can also capture downstream cost-savings and quality-of-life benefits associated with reduced hospitalisation rates and prolonged survival.

Modelled baseline risks

The baseline risks of mortality were derived from a number of sources. The UK Renal Registry report44 was first consulted as a source of population-based data. However, this report provides detailed data on survival only (by age) for the incident RRT cohort as a whole, without censoring for transplantation. This is not suited to the decision model structure (see Figure 13), in which mortality rates dependent on continuing to receive dialysis and on transitioning to transplant are required. Therefore, the ERA-EDTA annual report was consulted. 102 This report includes adjusted 5-year survival curves with censoring for transplantation in the dialysis survival estimates. The data are reported from day 91, with adjustment based on Cox regression for age, gender and primary diagnosis. The survival estimates on different modalities are expressed for a cohort of people aged 60 years and 60% male, with the following distribution for cause of renal disease: diabetes mellitus (20%), hypertension (17%), glomerulonephritis (15%) and other causes (48%). This distribution of characteristics is reasonably similar to that of the UK dialysis population, although age is slightly higher in the incident UK cohort at 63 years, and diabetes mellitus and hypertension are reported as the primary renal diagnosis in 26% and 6.5% of incident patients, respectively. 99 Although is preferable to reconstruct patient-level data from published Kaplan–Meier curves for the purpose of extrapolating survival in decision models,103 the adjusted nature of the reported data precluded estimation of numbers of events and censoring events. Therefore, a simple regression-based method was used to fit a Weibull distribution to the summary survival curve data. 104 Given limitations in the evidence base to support differences in survival by mode of dialysis, we based extrapolation on the survival curve for all dialysis modalities combined. The scale and shape parameters from the derived Weibull curves (Table 6) were incorporated in the model and used to extrapolate mortality risks out to 10 years. The scale parameter (λ) was further adjusted to reflect the starting age of the modelled cohort (67 years), using a published HR for mortality (beyond 91 days) associated with a 10-year increase in age:99

in which HR_ACM is the HR for all-cause mortality associated with a 10-year increase in age, and ‘start_age’ is the starting age of the modelled dialysis cohort.

For those transitioning to renal transplant, survival data were derived from a combination of sources (see Table 6). In the first year following transplant, survival probabilities by age groups were taken from the ERA-EDTA Registry annual report. 102 The reported 1-year survival probabilities differ by donor type (deceased/living), and were weighted accordingly. Beyond 1 year, we used published 10-year Kaplan–Meier survival data from a UK population-based study of transplant recipients. 105 The individual patient data were reconstructed for 2887 subjects aged 60–69 years following the approach described by Hoyle et al. ,103 using reported numbers at risk and steps in the published Kaplan–Meier curve. Parametric survival models were then fitted using R statistical software, version 3.1 (The R Foundation, Vienna, Austria), and the best-fitting model selected based on the Bayesian information criterion. This was a Weibull model. The scale parameter of the derived Weibull curve is adjusted in the model for the recipient’s age at time of transplant using the HRs for age reported by Karim et al. 105 All of the parameters used to model survival are presented in Table 6.

To minimise uncertainty associated with the use of parametric curves to extrapolate survival beyond 10 years, we applied an alternative approach to model mortality in the longer term. Mortality rates on RRT were estimated by applying reported relative risks of mortality in the RRT population compared with the UK general population99 to general population mortality rates adjusted for age/sex from UK life tables. For those remaining in a post-transplant state beyond 10 years following transplant, an adjusted relative risk106 was applied to the modelled annual mortality rate of age-matched patients on dialysis. Tonelli et al. 106 conducted a systematic review of observational studies reporting an adjusted HR for all-cause mortality with renal transplant versus dialysis. 106 Although a formal meta-analysis of these data was not conducted because of diversity across the included observational studies, the central estimate (0.42) of the reported range (0.16–0.76) across 23 included studies was applied in the base-case model. The reported range was treated as a CI for the purposes of assigning a log-normal distribution to this parameter.

Three-monthly probabilities of renal transplantation for those on dialysis were derived from the percentage of dialysis patients on a waiting list for a transplant (aged < 65 and ≥ 65 years),99 combined with the median duration of time to transplant (1082 days). 107 The graft failure rate for those receiving a transplant was derived from the 5-year graft survival rates reported for grafts from living and deceased donors. 107

All-cause inpatient hospitalisation was modelled using the first part of a published two-part cost model developed by Li et al. 100 Li et al. 100 used a data set for a cohort of patients on the UK Renal Registry who started receiving dialysis or received a kidney transplant in England between 1 April 2003 and 31 December 2006. The data on these patients were linked to Health Episode Statistics (HES) data for inpatient hospital activity (excluding activity for maintenance dialysis or transplant surgery) up to 6 years following initiation of dialysis or transplant. Each hospital event was costed using the appropriate Healthcare Resource Group (HRG) Payment by Results tariff for the admission. The data were then analysed using a two-part model: logistic regression was used to predict the probability of a patient incurring any inpatient hospital costs in a given year on RRT (up to year 6), and a general linear model was used to predict total inpatient costs in those who had at least one hospital episode in a given year. The models were adjusted for age, gender, years receiving dialysis, mode of dialysis, comorbidities, transplant and year of death (to account for increased hospital resource use in the year of death and year preceding death). The published two-part models for dialysis and transplant patients are replicated in Tables 7 and 8.

These models were incorporated in our decision model to predict the annual probability of hospitalisation each year based on the characteristics of the modelled cohort, and then to apply the associated inpatient hospitalisation costs. To keep the approach manageable in the context of a Markov cohort model, the odds ratios and cost coefficients associated with comorbidities were collapsed into a single weighted average for any one comorbidity, based on the reported frequency of each individual comorbidity. We then estimated the risk of hospitalisation at the cohort level by computing the weighted average of the risk for males and females, with and without comorbidities. The expected number of comorbidities among those in the cohort with any comorbidity was derived from the UK Renal Registry report,99 and the weighted average odds of hospitalisation associated with any one comorbidity was raised to this power in the calculation of hospitalisation risk in this segment of the cohort.

To fit the 3-month Markov cycle, the annual probabilities of hospital admission were converted to 3-monthly probabilities, assuming a constant inpatient hospitalisation rate over the year. Furthermore, the underlying rate was disaggregated into CV event- and other cause-related hospitalisation rates. To inform this process, we conducted a focused search of the literature for data on cause of hospitalisation in ESRD patients receiving dialysis. A number of studies were identified, suggesting that CV event-related hospitalisation rates account for ≈ 20% of all hospitalisations in dialysis cohorts. 108–110 The most relevant source of evidence to the UK RRT population reported that CV events made up 17.6% of the annual inpatient event rate in a cohort of 1226 UK HD patients. 110 This value was applied in the base case. We then further disaggregated expected CV hospitalisation events across types of CV events, in line with the reported relative frequency of CV event histories in the dialysis population (see Table 7). Although this is an uncertain assumption, we could not identify any better UK population-based data by which to model the relative frequency of different types of CV event in the dialysis population. A further limitation of the models used to predict annual hospitalisation risk, is the fact that these require extrapolation beyond the period of follow-up in the data sets used to develop them (i.e. beyond 6 years). We therefore had to assume that estimated probabilities of hospitalisation at 6 years on dialysis are generalisable across future years on dialysis.

Further adjustments to baseline risks

To allow for modelled scenarios in which effects are mediated through associations between hydration status and outcomes, the model was structured to enable mortality and hospitalisation rates to be adjusted upwards for proportions of the dialysis cohorts estimated to be severely overhydrated (ROH of > 15%). Modelled reductions in severe overhydration were then used to drive effects in scenarios using this version of the model.

The expected prevalence of severe overhydration (ROH of > 15%) was based on studies taking BCM measures at clinic visits (not necessarily first thing in the morning) for the PD cohort, and pre dialysis for the HD cohort. The threshold of ROH of > 15% was selected because, as noted in Non-randomised evidence, it has been associated with increased rates of mortality and hospitalisation in observational studies. 30,50,82,88 Time-averaged volume overload may give a more accurate estimate of the average exposure to fluid overload, but this measure has not been linked with mortality in observational studies. Limited data were identified regarding the prevalence of ROH of > 15% in UK dialysis cohorts. One observational study of 529 PD patients from a single UK centre82 reported that ≈ 31% of patients had ROH of > 10%. A multicentre European study, which included 734 patients from centres in Belgium, France, Romania, the UK (167 patients from two centres) and Switzerland, reported that 25.2% of the cohort were severely overhydrated (ROH of > 15%). 111 There were fewer published data available on the prevalence of severe overhydration in the UK HD population. However, a further multicentre European study matched PD patients from France, Romania and the UK with HD patients from the corresponding countries. 112 This study showed that pre-dialysis ROH in HD patients was similar to ROH in PD patients – although the time-averaged volume overload in HD patients was lower than the ROH value of PD patients. Based on these available data, the baseline prevalence of ROH of > 15% was set at 25% for both the HD and the PD cohorts.

The mortality rate for the severely overhydrated proportion of the HD cohort was increased using an adjusted HR of 1.87 (95% CI 1.12 to 3.13), as reported by Onofriescu et al. 88 The all-cause hospitalisation rate was also inflated upwards using an adjusted HR of 1.19 (95% CI 0.99 to 1.41), as reported by Onofriescu et al. 88 For the corresponding segment of the PD cohort, all-cause mortality was adjusted upwards using the HR of 1.83 (95% CI 1.19 to 2.82) reported by O’Lone et al. 82 No data were identified reporting the increased risk of all-cause hospitalisation in severely overhydrated PD patients, and so the same value as that used for HD patients was applied. It is plausible that any mortality/morbidity benefits associated with bioimpedance testing are also partly attributable to the avoidance of underhydration. However, no studies were identified linking underhydration, as measured using bioimpedance spectroscopy, to mortality and adverse events. Therefore, an underhydration state was not included in the model. As mentioned in Framework (method of synthesis), this could potentially underestimate the benefits if bioimpedance-guided fluid management can simultaneously reduce the proportion of patients that are seriously over- and underhydrated. Conversely, if the use of bioimpedance testing to guide fluid management decreases the proportion of patients who are overhydrated at the expense of increasing the proportion who are underhydrated, this model could potentially overestimate the benefits.

Incorporation of relative treatment effects

Alternative approaches to modelling effects of bioimpedance-guided fluid management on the baseline event rates were considered. Given the limitations in the existing evidence base for the clinical effectiveness of bioimpedance testing, combined with further limitations in the evidence base to inform certain baseline events, the modelled cost-effectiveness scenarios are subject to a significant degree of uncertainty.

With the availability of some trial evidence for the technology, the application of direct evidence for effects on final health outcomes was considered the preferred approach for modelling benefits. However, given the limitations in the trial evidence base, this was only possible for all-cause mortality. Of the three available BCM trials that included all-cause hospitalisation rates,61,76,77 these showed inconsistent and insignificant effects on this outcome. Therefore, we did not incorporate an effect on the overall hospitalisation rate in scenarios applying direct estimates of effects. Alternative approaches were explored in further scenario analyses to model plausible effects on CV event-related and non-CV event-related hospitalisation rates.

As a number of the trials reported effects on surrogate end points, including LVMI and PWV, we conducted a focused literature search to identify appropriate published sources of evidence to link changes in these surrogates to final health outcomes in the relevant patient population. A hierarchical approach was adopted to identify suitable sources of evidence, with priority given in descending order to the following types of evidence:

1. evidence linking intervention-induced changes in available surrogate outcomes to changes in the risk of final health outcomes

2. evidence linking non-intervention-induced longitudinal changes in surrogate outcomes to changes in the risk of final health outcomes

3. evidence from large UK or European cohort studies assessing the prognostic value of baseline measures of the surrogate measures for final health outcomes.

One systematic review, conducted in 2016, considered the value of LVMI as a treatment target in the area of ESRD, and concluded that there was no clear and consistent association between intervention-induced LVM change and all-cause or CV event-related mortality. 113 Furthermore, as only one of the BCM trials included this as an outcome, LVMI was considered no further. The search of available evidence did not identify any existing data showing a clear link between intervention-induced changes in PWV and final health outcomes in ESRD, but a large European observational study was identified. 114 This study assessed the prognostic value of baseline PWV on all-cause mortality and non-fatal CV events in a cohort of 1084 patients recruited from 47 European dialysis centres over a period of 2 years. It highlighted the importance of simultaneously considering abdominal aortic calcification (AAC) when assessing the prognostic value of PWV. Based on a multivariate Cox regression, both variables were found to be significant predictors of mortality and non-fatal CV events, but the effect of PWV was ameliorated at higher levels of aortic calcification (incorporated as tertiles), as a result of a significant negative interaction. The relevant HRs from the published Cox regression are provided in Table 9. Based on these estimates, and assuming that the UK dialysis cohort is similarly distributed across aortic calcification tertiles, we estimated an average effect on all-cause mortality and non-fatal CV events of a unit change in PWV, accounting for the interaction. This yielded a HR of 0.942 (95% CI 0.879 to 1.009) per unit reduction in PWV. We then explored the impact of scaling this effect to the magnitude of the pooled mean reduction in PWV (1.18 m/s) across the included BCM trials (see Figure 7), and applying it to the modelled proportion of all-cause hospitalisation events estimated to be attributable to CV events (assumes that a 1.18 m/s reduction in PWV is generalisable to the UK dialysis cohort). We also explored the impact of applying it to the all-cause mortality rate in the model. These analyses should be treated with caution, as they rely on cross-sectional associative evidence from an observational study to inform possible effects of bioimpedance monitoring. It should be further noted that the pooled estimate for the effect of bioimpedance monitoring on PWV is non-significant and based on results from only two trials, showing inconsistent results (see Figure 7). However, the point estimate is applied in the base-case model and the uncertainty surrounding it is propagated through the probabilistic analysis. Furthermore, the negative interaction between increasing AAC tertiles and the effect of baseline PWV on mortality and CV event-related hospitalisation, suggests that the relative effect of reductions in PWV may be greater in lower-risk groups (with lower AAC scores). On the other hand, evidence for an interaction in the prognostic value of baseline measures of these two variables does not necessarily mean that the AAC score would modify the effect of an intervention-induced reduction in PWV. Therefore, this model could potentially over- or underestimate the likely effects of the estimated reduction in PWV on final health outcomes. Better evidence on the effects of intervention-induced reductions in PWV are required to inform this issue.

As an alternative approach to indirectly estimate possible effects of bioimpedance-guided fluid management on mortality and CV event-related hospitalisation, we considered linking the estimated pooled reduction in SBP (2.46 mmHg; see Figure 11) to effects on CV events and mortality using a meta-analysis on the effects of blood pressure-lowering medications in dialysis patients. Heerspink et al. 115 estimated pooled relative risks of 0.71 (0.55 to 0.92) for CV events and 0.8 (0.66 to 0.96) for all all-cause mortality across eight trials; corresponding to a mean reduction in SBP of 4.5 mmHg. Assuming a log-linear relationship between SBP reduction and the relative risk of events, these effects can be rescaled to the mean reduction in SBP across included BCM trials (2.46 mmHg):

Relative risk for CV events for a 2.46-mmHg reduction in SBP:

Relative risk for all-cause mortality for a 3.44-mmHg reduction in SBP:

These effects are substantially larger than the estimated effects using PWV above, and suggest a potentially larger effect on CV events than on all-cause mortality. However, it is uncertain if effects on SBP induced by blood pressure medication can be generalised to potential reductions in SBP induced by the management of fluid status, that is, some blood pressure medications are thought to have effects on CV events that are independent of their blood pressure-lowering effects. 116 Furthermore, there is a complex relationship between fluid management and blood pressure,117 which makes it difficult to generalise. Nevertheless, the effect of bioimpedance-guided fluid management on SBP (bordering on significance), suggests a possible beneficial effect on both CV events and mortality. Therefore, we explored the impact of applying larger and differential relative effects on these outcomes in further scenario analyses.

Finally, we also explored the impact of using associations between overhydration and all-cause mortality and hospitalisation rates to drive effects in the model. For this analysis, we used data from Chen et al. ,80 to estimate the proportion of patients who could be shifted from the overhydrated to the normally hydrated states in the bioimpedance assessment and standard care arms of the model. This analysis assumed that for everyone who is moved from the overhydrated (relative fluid overload > 15%) to the normally hydrated state, the increased risks associated with overhydration are completely reversed. This was an optimistic assumption, as, again, cross-sectional associations between baseline measures and final outcomes were used to drive the effects of bioimpedance-guided fluid management in the model. The increased risk associated with baseline overhydration may not be fully reversible for those that can be returned to normal hydration status (≤ 15%).

A further problem with this approach is the lack of reporting in the RCTs on the effect of bioimpedance-guided fluid management on the proportion of patients with pre-dialysis ROH of > 15% at baseline and follow-up. Onofriescu et al. 88 did report proportions of patients within, and > 1.1 kg above and below, the BCM-guided target weight, and this study suggested no real change in the average percentage of patients who were > 1.1 kg above target weight throughout the follow-up period. Yet, the study did demonstrate a significant effect on PWV and mortality, leading the authors to speculate that the mechanism for effect may be as much a result of the avoidance of chronic underhydration as overhydration. Huan-Sheng et al. 76 reported a significantly larger reduction in mean overhydration (per litre) in patients with ROH of > 15% at baseline. We used these data to approximate percentage reductions in ROH of > 15% (absolute overhydration of > 2.5 litres) over the follow-up period by (1) assuming normal distributions for absolute overhydration at follow-up up and (2) subtracting mean reported reductions in overhydration (per litre) from simulated gamma distributions of baseline overhydration of > 2.5 litres. This yielded plausible percentage reductions in ROH of > 15% from 28% to 38% with bioimpedance-guided management relative to control. These were applied in model scenarios utilising the change in ROH status to drive effects on all-cause mortality and all-cause hospitalisation.

Further hypothesised benefits of bioimpedance-guided fluid management that were not incorporated in the main analyses included changes in quality of life (independent of effects on hospitalisation and CV events), maintenance of residual renal function and effects on dialysis requirements (number and duration of sessions).

None of the identified BCM trials reported on health-related quality of life, and only one included any patient-reported outcomes. 80 One observational study was identified that reported an association between hydration status and quality of life in Korean PD patients, as measured by the Kidney Disease Quality of Life-Short Form (KDQOL-SF) questionnaire. This showed that reductions in absolute overhydration (per litre) between baseline and 12 months were associated with improvements in the physical component score (1.81 points, 95% CI 0.78 to 2.84 points), mental component score (0.92 points, 95% CI 0.2 to 1.65 points) and the kidney disease component score (0.9 points, 95% CI 0.36 to 144 points), as measured by the KDQOL-SF questionnaire. These analyses were adjusted for various potential confounders, including age, sex, dialysis vintage, haemoglobin level, baseline overhydration status and comorbidities (as measured by the Charlson Comorbidity Index). Although this study suggests that use of the BCM could lead to improvements in heath-related quality of life (independent of effects on adverse events), it is not clear how generalisable the reported changes are to the UK population. In addition, it is not possible to map from changes in the reported aggregate component scores of the KDQOL-SF questionnaire to changes in health state utility values. Furthermore, our model already captures QALY gains associated with prevention of hospitalisation events and increasing comorbidity, and, therefore, including a constant utility increment associated with the use of bioimpedance testing could lead to double-counting of QALY gains. Nevertheless, the impact of including a 2% and 5% improvement in health state utility as a result of improved interdialytic symptoms was assessed in a further scenario analysis.

The omission of residual renal function as an explicitly modelled state, and its knock-on effects on dialysis requirements and outcomes as potential benefits, is justified by a current lack of supporting evidence. The only BCM trial that assessed proxies for residual renal function in HD patients reported a significant increase in the proportion of anuric patients and a significant decrease in urine output in non-anuric patients at 12 months in the bioimpedance assessment group. 77 In the only other bioimpedance trial to report renal output, PD patients randomised to bioimpedance-guided fluid management showed a slightly larger reduction in mean urine volume at 12 weeks, although this was not statistically significant. 88 There is a recognised risk that aggressively pursuing lower target weights using a high ultrafiltration rate may in fact lead to increased morbidity/mortality as a result of hypoperfusion-induced ischaemic injury and accelerated loss of residual renal function. 118,119 Conversely, identifying patients that are severely underhydrated and adjusting their target weight upwards may help to preserve residual renal function. This question is currently being evaluated in two ongoing RCTs, one in PD patients120 and one in HD patients based in the UK (see Table 5). 94 If bioimpedance testing can preserve residual renal function in non-anuric patients, then our model may underestimate the average benefits in the population as a whole. Conversely, if the use of bioimpedance testing leads to decisions that accelerate the loss of residual renal function, this is a disbenefit that could have knock-on effects on dialysis costs and adverse outcomes.

Related to the above issue, there is also a lack of evidence on the effects of bioimpedance monitoring on the number, and the duration, of dialysis sessions required to achieve the prescribed target weight. A potential cost-saving could be achieved through a requirement for fewer dialysis sessions in some incident patients, if it is found to be effective in preserving residual renal function. As discussed above, this question remains currently unanswered. Conversely, the use on bioimpedance spectroscopy could result indirectly in increased dialysis costs through identification of patients that are severely overhydrated and require longer or additional dialysis sessions to achieve their new target weight without exceeding safe ultrafiltration rates and volumes. 118 Our cost-effectiveness model assumes that any effects of bioimpedance-guided fluid management on dialysis requirements are cost neutral.

Resource use estimation

The base-case cost-effectiveness model incorporates health service costs associated with maintenance dialysis, blood pressure medication and ESAs (on dialysis), all-cause inpatient hospitalisation, renal transplantation (including work-up, surgery and follow-up), post-transplantation immunosuppression and outpatient visits (all expressed in 2014–15 pounds sterling).

Costs of renal replacement therapy

It has previously been noted that dialysis treatment for CKD results in high costs to the health service, and that this can undermine the cost-effectiveness of technologies that prolong survival on dialysis. In some circumstances, a technology that prolongs survival for patients receiving dialysis may not be cost-effective at a price of zero. This has led to inconsistency across economic evaluations in the area of ESRD with respect to whether or not dialysis costs are included. Some have argued that they should not be included for interventions that aim to extend survival without impacting the need for dialysis. 121 A further argument for their exclusion is that a decision has already been made to fund dialysis on broader ethics/equity considerations that are not reflected in the cost-effectiveness of dialysis itself. It may then seem unfair to include dialysis costs when they act as an insurmountable barrier to demonstrating the cost-effectiveness of other technologies that prolong survival on dialysis. The alternative argument is that dialysis costs do represent a real opportunity cost associated with ongoing treatment for ESRD, and thus should be included in the analysis. The NICE Decision Support Unit has produced a report on assessing technologies that are not cost-effective at a price of zero, and this suggests that all NHS and Personal Social Services costs that differ between the technology being appraised and the comparator technologies should be included within the incremental cost-effectiveness ratio (ICER), as this provides the ICER that reflects the real opportunity cost of recommending the technology being appraised. 122 However, the report does also note that a case could have been made in a previous technology appraisal of cinacalcet123 for treating secondary hyperparathyroidism in ESRD, to exclude dialysis costs on the grounds that they are unrelated to the treatment of secondary hyperparathyroidism (the condition of interest in the appraisal). A similar argument was adopted in the modelling assessing the cost-effectiveness of alternative phosphate binders for people with stages 4 and 5 CKD with hyperphosphataemia. 98 In this example, the costs associated with dialysis were argued to be unaffected and unrelated to the choice of phosphate binder, with the target condition being hyperphosphataemia.

Considering the above, it is difficult to argue that dialysis costs are unrelated to a technology being used to guide fluid management decisions in patients receiving dialysis, and in theory the use of the technology could have an impact on dialysis costs in survivors, as well as prolong survival on dialysis. Therefore, dialysis costs are included in our base-case cost-effectiveness scenarios. However, it is a plausible argument that the cost-effectiveness of dialysis reflected in our model does not capture its broader value to society, relating to ethics and equity issues. Therefore, we also explored the impact of excluding dialysis costs.

Dialysis costs were taken from the current NHS reference costs. 124 For HD costs, we took the weighted average of the reference costs (per HD session) for the HRG codes LD01A to LD10A (at base and away from base), weighted by the relative proportion of overall activity reported against each code. This was multiplied by three sessions per week, and then by 52 to estimate the average annual cost of maintenance HD. For PD, we applied the weighted average of the reference costs (per day) for HRG codes LD11A to LD13A. These costs were multiplied by 365 to estimate total annual maintenance PD costs.

Transplantation costs were also taken from the reference costs, applying the average costs for HRG codes LA01A, LA02A and LA03A (elective inpatient). We also included costs for follow-up post transplantation. For year 1 this was derived from Treharne et al. 97 To this we added the costs of immunosuppressant in year 1, based on an initiation regimen in the first 2 weeks, and a maintenance regimen thereafter. For the initiation period, we costed basiliximab [(Simulect^®, Novartis Pharmaceuticals UK Ltd, Frimley, UK) day 0 and day 4], prednisolone [(Concordia International, Oakville, ON, Canada) 60 mg per day], mycophenolate mofetil [(Cellcept^®, Roche Products Limited, Welwyn Garden City, UK) 1 g twice daily] and tacrolimus [5 mg twice daily (we used the average list price of Adoport^®, Sandoz Ltd, Frimley, Camberley, Surrey, UK; Prograf^®, Astellas Pharma Ltd, Chertsey, UK; and Advagraf^®, Astellas Pharma Ltd, Chertsey, UK)]. For the maintenance period, we costed prednisolone (7.5 mg per day), mycophenolate mofetil (1 g twice per day) and tacrolimus (5 mg twice daily). Beyond year 1 post transplant, we applied maintenance immunosuppression costs and added average outpatient costs observed in the transplant cohort (see Outpatient costs below). The maintenance dialysis and transplantation costs are provided in Table 10.

Costs of unplanned inpatient hospitalisations

Annual inpatient hospital costs for patients on maintenance dialysis and post transplant were estimated using the two-part model developed by Li et al. ,100 as described above (see Modelled baseline risks and Tables 7 and 8). The second part of this model predicts annual inpatient costs conditional on experiencing any hospitalisation event(s) within a given year. Estimates are based on age, sex, years on dialysis (or years following transplant), dialysis modality and the comorbidity status of the modelled cohort. The model also accounts for increased costs in the year of death and the year preceding deaths that occur in the first half of the following year. This reflects the increasing level of morbidity experienced by patients towards the end of life.

To generate estimates of the cost incurred per hospitalisation event rather than costs per year, the predicted annual cost is divided by the expected number of events per patient-year in our model. The expected event rate is derived from the estimated annual probability of experiencing any hospital inpatient event, assuming a constant event rate across the year. This results in a cost per hospitalisation event that varies by the underlying characteristics of the modelled cohort, but comes to ≈ £4500 per event for patients receiving HD and £4300 per event for patients receiving PD. These estimates are substantially higher than costs per hospitalisation event that have been applied in previous models in the area of ESRD. 17,97,98 However, these previous models generally applied averages of aggregate reference costs. The estimates derived from the data reported by Li et al. 100 are based on a large data set of actual hospital episodes costed according to admission code and adjusting for length of stay. Thus, the estimates derived from Li et al. 100 are applied in the base-case analysis. Alternative estimates are applied in a sensitivity analysis. Furthermore, as the above approach makes a simplifying structural assumption of one hospitalisation event per quarter, it will actually slightly underestimate annual inpatient hospital costs as predicted by the published two-part model of Li et al. 100 Therefore, we also applied a structural sensitivity analysis in which the annual probability of hospitalisation and the conditional annual costs were applied only in the first cycle of whole years.

A further limitation of the inpatient cost models reported by Li et al. 100 is that they predict average costs across all causes of admission. Thus, our base-case model assumes that both CV and other cause-related hospitalisation events incur the same cost on average. A further uncertainty relates to the fact that some inpatient costs will be unrelated to ESRD. However, as Li et al. 100 noted, it is difficult to judge whether individual admissions are related or unrelated, as ESRD is associated with increased risks of hospitalisation across many major causes. The inclusion of all-cause hospitalisation as an outcome in a number of the bioimpedance trials further justifies the inclusion of all-cause hospitalisation events in the baseline model.

Outpatient costs

Total outpatient costs for dialysis and transplant patients were also included in the base-case model. These were taken simply as the observed annual outpatient costs on dialysis and transplant as reported by Li et al. :100 £1202 per year for dialysis patients, and £2388 per year for transplant patients. These were divided by four and applied per quarterly cycle in the model.

Costs of background medications for dialysis patients

Unit costs and the proportion of patients taking blood pressure medicine have been applied to provide the total cost of blood pressure medicine (Table 11). The percentage of patients taking different types of blood pressure medicine was taken from the baseline data of a RCT,127 which recruited dialysis patients from three UK dialysis centres: Stoke-on-Trent, Leeds and Sheffield. For the different classifications of drugs, prices for specific drug names commonly prescribed under each classification (informed by the clinical advisor of the assessment group) were taken from the British National Formulary (BNF). 126 Drugs were costed at the recommended dose as described in the BNF.

We further considered the potential impact of incorporating an effect of bioimpedance testing on the use/cost of blood pressure medication. Only two of the existing BCM trials reported on this outcome. 60,63 Luo et al. 63 reported no significant changes in the dose of blood pressure medication in either the control or BCM groups of their trial. Onofriescu et al. 60 reported that there was no statistically significant change from baseline to end of study in the percentage of patients taking blood pressure medication in the control group of their trial, but they did report a significant within-group reduction in the BCM arm (from 66% to 55%). No formal comparison was reported for this outcome. However, as an exploratory analysis, we assessed the impact of assuming a 10% reduction in blood pressure medication use in the bioimpedance assessment arm of our model. To estimate the associated cost reduction, we assumed 68.4% of the cohort would be taking at least one blood pressure medication,128 at an average cost of £129.81 (£88.76/£0.684) per year. The average cost reduction associated with an absolute 10% reduction in the proportion of patients on any blood pressure medication was then estimated:

The unit costs, units per week and proportion of patients taking ESAs were applied to provide an estimate of the total annual cost for ESAs for dialysis patients (Table 12). The proportion of patients taking an ESA was taken from the UK Renal Registry report,99 that is, 87% of those receiving HD and 68% of those receiving PD. The median dose for the corresponding population receiving HD and PD was 7400 international units (IUs) and 4500 IU per week, respectively. Based on opinion obtained from the clinical advisor of the assessment group, the unit cost per IU was derived as the average of the unit costs for epoetin beta (NeoRecormon^®, Roche Diagnostics, Hertford, UK) and darbepoetin alfa (Aranesp^®, Amgen, Thousand Oaks, CA, USA) as reported in the BNF (£0.00718 per IU). Thus, the total annual cost of ESAs was estimated to be £2403.69 ( = 0.87 × 7400 × 52 × 0.00718) for people receiving HD and £1142 ( = 0.68 × 4500 × 52 × 0.00718) for people receiving PD (see Table 12).

Costs of bioimpedance testing/monitoring

The costs of the devices, provided by the companies (Table 13), were annuitised over 5 years using an annual depreciation rate of 3.5%. The cost of the BCM was applied in the base-case analysis because of a lack of clinical effectiveness evidence for the alternative devices. For comparison, we also estimated the costs per patient-year and cost per test for the alternative devices, with identical assumptions about numbers of tests and staff time requirements per patient. Estimated costs of BCM equipment maintenance were provided at two levels: £250 for an annual maintenance contract, and £600 for annual maintenance, including parts and labour. We included the higher-cost maintenance contract in the base-case scenarios, but also assessed the impact of removing the maintenance costs in a sensitivity analysis.

The unit costs of staff involved in bioimpedance testing were taken from the Unit Costs of Health and Social Care 2015 (Table 14). 129 These were applied to estimates of staff time required to conduct and interpret tests. They were also used to place a cost on staff time invested in training in the use of bioimpedance testing. The company responsible for the BCM device indicated that it takes, on average, 5–10 minutes to conduct a test. 4 In the base-case analysis, the time to perform the measurement was assumed to be 7 minutes. The company responsible for the BCM device indicated that they provide free training on its use, taking a half-day to attend. In the base-case analysis, the training was assumed to take 3.5 hours.

To gain a better understanding of the number of bioimpedance devices required to cover quarterly testing of the dialysis population, a brief questionnaire (see Appendix 11) was sent to the specialist members of the appraisal committee. Six different members responded, although three were from a single large centre covering adults and children. Therefore, we had information on testing practices from three centres covering adults, and two centres covering children (one exclusively). The adult centres were relatively large, with 538–942 dialysis patients in total, compared with the average in England of 456. 99 The impact of adopting lower extremes of device throughput was assessed in the sensitivity analysis.

The questionnaire included questions about centre size (number of HD and PD patients), number of satellite units, current practice with respect to fluid management decisions, and current practice regarding the use of bioimpedance testing. Questions were also included about the estimated level of resource that would be required to conduct quarterly testing of all HD and PD patients for whom the individual’s centre was responsible. Respondents from two of the centres described a situation in which the majority of their patients were already being monitored using bioimpedance testing at least every 3 months. For the third centre, it was noted that bioimpedance testing was not currently performed systematically, but was rather used for selected patients. Consequently, only the anticipated resource use required for quarterly bioimpedence testing was used for this centre.

Details of relevant resources and costs required for quarterly testing, based on the responses from the three adult centres, are summarised in Table 15. Total equipment costs were estimated by multiplying the equivalent annual cost per device by the estimated number of devices required for quarterly monitoring of all dialysis patients across the centres. This was then divided by the total number of patients to estimate the cost of equipment per patient per year, and then further divided by four to estimate the equipment costs per test performed. For example, centre A reported 15 bioimpedance devices to cover a total of 585 patients [(£1273.52 × 15)/585 = £32.65]. The maintenance costs also depended on the reported number of devices required by the centre to cover quarterly testing of its dialysis population. The total estimated annual maintenance cost, with and without parts and labour, was allocated across patients using the same approach as for equivalent annual costs of equipment. The larger centre latterly reported that it did not take out a maintenance contract on its machines, and so we also explored the impact of removing these costs completely.

Staff costs associated with the time required to conduct each test were estimated based on 7 minutes of direct patient contact with a band 6 nurse [7 × (£109/60) = £12.72]. This was further multiplied by four to estimate the staff costs per patient per year (£12.72 × 4 = £51). The added consultant time required to interpret the findings of each bioimpedance test was assumed to be 5 minutes in the base-case analysis [5 × (105/60) = £8.75]. Total training costs for each centre were estimated based on the number of different grades of staff trained, multiplied by their costs per contract hour and the number of hours of training attended. This total initial investment was spread over 5 years, and the equivalent annual cost was divided by the number of patients in the centre to give a cost per patient per year. For example, for centre A the total training costs were estimated to be £11,171. Annuitised over 5 years, this comes to £2474 and £4.23 per patient per year (2474/585).

Finally, to estimate the total annual cost of adding bioimpedance testing to standard practice, the total cost of consumables (electrodes and patient cards), also based on quarterly testing (Table 16), was added to the estimated device, maintenance, staff time and training costs. The total estimated cost per patient-year for each adult centre, and the average cost per patient across centres, is reported in Table 17 for each device. For the base-case cost-effectiveness analysis, we applied the average cost per patient per year using the BCM, based on the higher maintenance costs (£101.41), and applied a distribution incorporating the lower and upper 95% confidence limits of £85 to £125. Given some uncertainty regarding the ongoing maintenance costs for each device, Table 17 also presents estimated costs per patient-year for each device, excluding all maintenance costs. The costs are very similar across the different devices. In addition to the estimates presented in Table 17, based on responses from dialysis units, we also estimated a cost per patient-year based on responses from two paediatric units. As a result of substantially lower throughput, the costs were significantly higher: £149–349 based on four tests per year and £243–451 based on 12 tests per year. However, these may be overestimated in situations where devices can be shared between adults and children.

Health measurement and valuation

Health state utility values for patients on dialysis and post transplant were identified from a focused review of the literature. We first identified two systematic reviews of utility data in the context of ESRD incorporating studies relevant to the NICE reference case [reporting EuroQol-5 Dimensions (EQ-5D) data for UK patients]. 130,131 We focused our searches on identifying any more recent studies published following December 2010 (the end date of the search conducted for the most recent systematic review). This identified no further studies reporting EQ-5D values, specifically for UK patients. The systematic review conducted by Wyld et al. 131 included a random-effects metaregression to predict utility based on several factors: treatment (transplant, dialysis, pre-treatment, conservative management) and method of utility elicitation. This model predicted an EQ-5D utility value of 0.64 for patients receiving dialysis and 0.75 for transplant patients. However, a limitation of this study was that some of the EQ-5D scores were measured from mapping algorithms, and the age to which the mean utility estimates applied was not reported. The earlier systematic review by Liem et al. 130 restricted a meta-analysis to those studies using the EQ-5D index directly for each modality of RRT, and reported the pooled mean age and sex distribution for the corresponding pooled EQ-5D values. These are reported in Table 18. The age- and sex-matched EQ-5D UK population norms were calculated using an equation published by Ara and Brazier132 and used to derive age-/sex-adjusted utility multipliers from the raw pooled estimates. 133 The alternative utility values derived from Wyld et al. 131 were applied in a sensitivity analysis, assuming the same age and sex distributions, as reported by Liem et al. 130 for purposes of adjustment.

A significant proportion of inpatient hospitalisations are associated with CV events in the dialysis population, as assumed in the model. It is reasonable to assume that such events will be associated with short-term and lasting disutility. This is the assumption that is used in CV event models in non-dialysis populations, and the best-recognised source of English EQ-5D data for different CV event histories is the Health Survey for England, as reported by Ara and Brazier. 132 Therefore, these data were used to estimate age-adjusted utility multipliers during the first and subsequent years following different types of CV event. A weighted average of these multipliers for the first and subsequent years was then calculated (based on relative frequency of CV event histories in the dialysis population) and applied to the proportion of the cohort modelled to experience an incident CV event. For example, a cohort of 60-year-old patients, who were stable and receiving HD, would be assigned a utility value of 0.56, whereas a cohort of 60-year-old patients, receiving HD who experienced an incident CV event within 1 year, would be assigned a utility of 0.466 (0.56 × 0.832) and a cohort of 60-year-old patients, who had gone over 1 year since an incident CV event had occurred, would be assigned a utility value of 0.521 (0.56 × 0.931).

Finally, hospitalisations for any other reason were also assumed to incur an acute utility decrement. These were taken from the modelling used to inform the NICE guidelines on PD. 17 In the modelling for these guidelines, a 6% reduction was applied to any dialysis complication. 17 The same 6% reduction is applied in our model for the second half of the 3-month cycle in which complications occur.

Time horizon and discounting of costs and benefits

The modelling was analysed over the lifetime of patients: 30 years for a cohort of 66-year-old patients in the base-case analysis. The time horizon was extended in years for scenario analyses involving younger cohorts. The lifetime horizon was chosen to fully capture any survival or ongoing quality-of-life benefits associated with bioimpedance testing. All future costs and benefits were discounted at a rate of 3.5% per annum.

Analysis

The results of the model are presented in terms of a cost–utility analysis over the lifetime of the simulated cohorts. The bioimpedance-guided fluid management strategy is compared incrementally with standard care, to estimate its incremental costs and QALYs. This is expressed as the ICER. The net benefit framework is used to identify the optimal fluid management strategy at different threshold ratios of willingness to pay per QALY. To characterise the joint uncertainty surrounding point estimates of incremental costs and effects, probabilistic sensitivity analyses were undertaken. All costs were assigned either normal or gamma distributions, utility multipliers were assigned beta distributions and HRs were assigned log-normal distributions using the point estimates and CIs (or SEs) reported in Tables 6, 9, 10 and 18. The parameters of the derived Weibull survival functions were entered deterministically for the dialysis cohort, but as a multivariate normal distribution for post-transplant survival. Distributions for the computed hospitalisation rates and associated costs were assigned SDs set at 10% of the mean. The results of the probabilistic analyses are presented in the form of cost-effectiveness acceptability curves (CEACs). Further deterministic sensitivity analyses were used to address other forms of uncertainty.

The primary analysis was conducted for a mixed cohort of patients receiving HD or PD. Subgroup analyses were conducted to explore any differences in cost-effectiveness by mode of dialysis and, when data allowed, by characteristics of the patient population. The impact of applying different assumptions with respect to testing frequency and throughput was also explored through scenario analyses. Scenario analyses were also used to explore the impact on cost-effectiveness of other sources of uncertainty.

Cost-effectiveness results

The model was first set up to assess the cost-effectiveness of bioimpedance-guided fluid management versus standard care for a mixed cohort of HD (87%) and PD (13%) patients.

The key assumptions of the base model are as follows:

• The starting age of the cohort is 66 years.

• Survival with HD and PD treatment is equivalent, and patients do not switch between dialysis modes.

• Survival to 10 years with dialysis treatment is based on parametric extrapolation of 5-year survival curves, reported for patients in the European renal registry. 102

• Survival beyond 10 years is estimated by applying published, age-specific, relative risks of death to RRT cohorts compared with general population norms. 99

• Fixed proportions of the cohort are on a waiting list for transplant, and wait a median of ≈ 3 years, conditional on survival. 107 No transplants occur beyond the age of 75 years.

• Following graft failure, transplant patients incur costs of dialysis, that is, no further transplants are modelled.

• Probabilities of all-cause inpatient hospitalisation are estimated by age band, time receiving RRT, RRT modality, sex and comorbidity status, using a published regression based on linked UK Renal Registry – HES data. 100

• It is assumed that 17.6% of all inpatient hospitalisations are caused by CV events.

• Health state utility decrements are applied in the acute period for all hospitalisation events, and ongoing health state utility decrements are also applied post CV event-related hospitalisation. 110

• First-incident CV event-related hospitalisations increase the comorbidity burden on the cohort by one, resulting in an increased risk of hospitalisation in subsequent cycles.

• Costs of dialysis, treatment for anaemia (ESA), blood pressure medication, all inpatient hospitalisations, all outpatient attendances, renal transplants, post-transplant hospitalisations, outpatient attendances and post-transplant immunosuppression are included in the base model.

• Costs of CV event-related hospitalisation are assumed to be equal to the average cost across all hospitalisations in dialysis patients (i.e. CV events account for 17.6% of all hospitalisation costs).

• The incremental cost of monitoring patients using bioimpedance testing is added in the bioimpedance assessment arm of the model (assuming four tests per year).

• Effects of bioimpedance monitoring on all-cause mortality are applied for 10 years in the model.

• Effects of bioimpedance monitoring on CV event-related or all-cause hospitalisation are applied over the lifetime of the cohort.

• Costs and QALYs are discounted at a rate of 3.5% per annum.

The following set of results are based on several alternative base-case scenarios with respect to the possible effects of bioimpedance-guided fluid management on mortality, hospitalisation rates and blood pressure medication use. There is significant uncertainty surrounding the clinical effectiveness of bioimpedance monitoring, as highlighted in the clinical effectiveness chapter. Therefore, the point estimates of incremental cost-effectiveness should be treated with caution.

The main clinical effectiveness scenarios explored are described below and summarised in Table 19.

1. Only the pooled HR (0.689, 95% CI 0.228 to 2.084) for the effect of bioimpedance testing on mortality is applied to the base model. It should be noted that this pooled effect from the meta-analysis (see Figure 8) is not statistically significant, but directionally favours bioimpedance-guided fluid management. Given uncertainty regarding long-term effects, this effect is applied over 10 years in the model (up to cycle 40).

2. A possible effect of bioimpedance testing on non-fatal CV events is added to the effect on mortality in scenario 1. The applied HR (0.9318, 95% CI 0.829 to 1.048) was derived as described in Incorporation of relative treatment effects using published observational data on the prognostic value of PWV for CV events and mortality (see Table 9), combined with the pooled mean reduction in PWV (see Figure 7) observed across the bioimpedance trials included in our systematic review. This scenario is heavily caveated by the application of non-significant effects on PWV, combined with observational prognostic evidence, to model possible effects on health outcomes.

3. This scenario applies the same effect, derived through the pooled reduction in PWV, to both mortality and non-fatal CV events in the model, that is, a HR of 0.9318 is applied to both all-cause mortality and the CV event-related hospitalisation rate. This scenario comes with the same caveats as scenario 2.

4. Scenario 4 replicates scenario 3, but adds a possible effect of bioimpedance-guided fluid management on blood pressure medication use. As described under Costs of background medications for dialysis patients, a possible cost reduction of £12.98 per year was derived from existing trial evidence. Note, however, that this was only observed/reported in one of the RCTs,57 and was not based on a formal adjusted comparison.

5. Scenario 5 uses reported observational associations between baseline hydration status (as measured by the BCM) and mortality and all-cause hospitalisation. The effect of bioimpedance testing is modelled through a plausible reduction in the proportion of the cohort (25%) that is severely overhydrated (ROH of > 15%). 80 Using data reported by Huan-Sheng et al. ,76 it was estimated that the proportion of severely overhydrated patients could be reduced proportionally by 28–38% with bioimpedance-guided fluid management relative to control. This scenario applies a 28% proportional reduction in severe overhydration in the bioimpedance assessment arm of the model.

6. Scenario 6 replicates scenario 5, but applies a 38% proportional reduction in severe overhydration in the bioimpedance assessment arm of the model.

Table 20 presents the model-based cost-effectiveness findings for the main clinical effectiveness scenarios 1–6 (described above). Across the scenarios, bioimpedance-guided fluid management comes out as the more costly strategy, resulting in increased costs to the health service between £4519 and £35,680. These increased costs are accompanied by QALY gains under the alternative effectiveness scenarios between 0.07 and 0.58. The ICERs for bioimpedance testing range from £59,551 to £66,013 per QALY gained. It should be noted that the increased costs associated with bioimpedance-guided fluid management are primarily driven by the high dialysis costs during life-years gained. The cost of bioimpedance testing is modest, adding, on average, £101 per patient-year.

As discussed in Costs of renal replacement therapy, others have argued for the exclusion of dialysis costs in the assessment of technologies that aim to extend survival of patients receiving dialysis without influencing the need for dialysis, as these technologies can act as an insurmountable hurdle to demonstrating cost-effectiveness. The results for effectiveness scenarios 1–6 with dialysis costs excluded are therefore provided for comparison in Table 21. It can be noted that this results in a large reduction in the ICERs for bioimpedance testing, ranging between £15,644 and £21,206 per QALY gained. Note, however, that these point estimates are based on uncertain effects incorporated as deterministic point estimates.

Markov traces

Figures 14 and 15 show the Markov traces for the standard care arm and the bioimpedance assessment arm under clinical effectiveness scenario 3. In the standard care arm, the 10-year mortality for the cohort of 66-year-old patients was 78.8%. This is consistent with the observed 10-year mortality in UK patients receiving RRT surviving beyond 90 days (≈ 68% in 56- to 64-year-olds and ≈ 88% in 65- to 74-year-olds). 99 Assuming a constant effect of bioimpedance-guided fluid management on mortality, the 10-year mortality in the bioimpedance assessment arm was 76.6%. Over the lifetime of the modelled cohort, the gain in undiscounted life expectancy was 0.29 years (6.29 vs. 6.0 years). The modelled lifetime cumulative incidence of any CV hospitalisation event was 46.9% in the bioimpedance assessment arm of the model, and 47.1% in the standard care arm. A total of 7.8% of patients in the bioimpedance assessment arm received a transplant during their lifetime, while the corresponding figure was 7.6% in the standard care arm.

FIGURE 14.

Markov cohort trace: standard care arm (one stage equals 3 months).

FIGURE 15.

Markov cohort trace: BCM measurement arm, under clinical effectiveness scenario 3 (one stage equals 3 months).

Table 22 provides a breakdown of the cumulative costs for the standard care and bioimpedance measurement arms, respectively, under clinical effectiveness scenario 3. The costs were higher across all categories in the bioimpedance measurement arm, as a result of the slight increase in survival. However, it can be noted that it was the additional dialysis costs in extra years that made up 74% of the total incremental cost of the bioimpedance-guided strategy. This same pattern was consistent across all the main clinical effectiveness scenarios (1–6). The actual increase in lifetime costs, as a result of bioimpedance testing, was small (£491 per patient in clinical effectiveness scenario 3).

Deterministic sensitivity analysis

Figures 16 and 17 illustrate the effects of a one-way sensitivity analysis on key model input parameters, with dialysis costs included (see Figure 16) and excluded (see Figure 17). The reference ICERs for both these tornado diagrams reflected clinical effectiveness scenario 3, that is, a HR of 0.9318, inferred through the pooled reduction in PWV, applied to both all-cause mortality and CV event-related hospitalisation.

FIGURE 16.

One-way sensitivity analysis: BCM measurement vs. standard care (clinical effectiveness scenario 3, including dialysis costs). ACM, all-cause mortality; Bioimp, bioimpedance; c, cost; EV, expected value; ICHD, ischaemic coronary heart disease; p, probability; u, utility.

FIGURE 17.

One-way sensitivity analysis: BCM vs. standard care (clinical effectiveness scenario 3, excluding dialysis costs). ACM, all-cause mortality; Bioimp, bioimpedance; c, cost; EV, expected value; ICHD, ischaemic coronary heart disease; p, probability; u, utility.

When dialysis costs were included, the ICER for bioimpedance-guided fluid management was most sensitive to changes in the HR for the effect on all-cause mortality. The most favourable ICER (£40,283) occurred when the HR on all-cause mortality was equal to one, as this equalised survival and eliminated the excess dialysis costs incurred in added years.

When dialysis costs were excluded, the ICER remained most sensitive to the HR on all-cause mortality. Results were also moderately sensitive to the utility multiplier for HD, the cost of HD and the HR for CV event-related hospitalisation. However, when dialysis costs were included, the ICER remained well above £30,000 when these parameters were varied within their ranges. Conversely, the ICERs all remained below £30,000 when the parameters were varied individually within their ranges (referent to clinical effectiveness scenario 3) with dialysis costs excluded.

Scenario analyses

Table 23 presents the results of further scenario analyses, referent to clinical effectiveness scenario 3 (HR of 0.9318 applied to all-cause mortality and CV event-related hospitalisation). Unless otherwise stated, these additional scenarios excluded dialysis costs to better illustrate sensitivity (around the cost-effectiveness threshold) when the exclusion of dialysis costs was considered to be appropriate for the purpose of decision-making. Under most of the scenarios with dialysis costs excluded, the ICER for bioimpedance monitoring remained below £30,000, and was most often below £20,000.

Under only a few scenarios did the ICER for bioimpedance monitoring fall close to or below £30,000 when dialysis costs were included, when assuming that bioimpedance testing would result in a 5% or 10% reduction in dialysis costs (scenarios 15 and 16) over the lifetime of patients and when it was assumed that bioimpedance-guided fluid management would result in a 5% increase in health state utility, maintained over the lifetime of all dialysis patients (scenario 13). However, there are very few data available to justify these possible scenarios.

Subgroup analysis

Table 24 presents the results of the analysis that considered key subgroups of the dialysis population.

Separate analyses were considered by comorbidity status (none/at least one), dialysis modality (HD/PD), starting age of the cohort (55 years rather than 64 years) and transplant listing (yes/no). For comparability, all of these analyses were conducted with clinical effectiveness scenario 3 (HR of 0.912 for the effect of bioimpedance monitoring on mortality and CV hospitalisation). Finally, we also conducted a subgroup analysis using the overhydration states in the model (clinical effectiveness scenario 6), with the effect of bioimpedance testing modelled through a plausible proportional reduction in severe overhydration (ROH of > 15%), reducing the risk of all-cause mortality and CV event-related hospitalisation. This analysis focused on the subgroups that were identified as being severely overhydrated at baseline, and assumed a 38% reduction over the follow-up period (see Table 24, scenarios 8 and 9).

These analyses did not reveal any large differences in cost-effectiveness by subgroups. The ICER was slightly higher in the subgroup on a waiting list for a transplant, as the patients spent less time on dialysis and so benefited less from the modelled reduction in all-cause mortality and CV event-related hospitalisation conferred by bioimpedance-guided fluid management. In the scenario focusing on the severely overhydrated subgroup, the ICER was ≈ £5000 lower than in the corresponding base case for that clinical effectiveness scenario, but when dialysis costs are included the ICER remains well above the accepted thresholds (£59,318), as it does for all the subgroups (results not shown).

For comparison with the deterministic results in Tables 20 and 21, Tables 25 and 26 present the results for clinical effectiveness scenarios 1, 3 and 4 based on 1000 probabilistic iterations of the model, with dialysis costs included (see Table 25) and excluded (see Table 26). The point estimates of the ICERs are very similar to the deterministic ICERs. The final columns in Tables 25 and 26 indicate the probability of standard practice or bioimpedance testing being the preferred strategy, given a willingness to pay of £20,000 per QALY gained. With dialysis costs included, the probability of bioimpedance testing being cost-effective is ≈26% in scenario 1 and < 13% in scenarios 3 and 4.

With the dialysis costs excluded, the probability of bioimpedance testing being cost-effective at a threshold of £20,000 increased to ≈61–67% across effectiveness scenarios 1, 3 and 4 (see Table 26). There remains a high degree of uncertainty inherent in the approach required to link possible effects of bioimpedance monitoring on arterial stiffness (PWV) to effects on mortality and non-fatal CV events, which is not fully captured in the probabilistic model. Thus, the probability of cost-effectiveness in scenarios 3 and 4 may give a somewhat unrealistic impression of precision.

For further comparison, the incremental cost-effectiveness scatterplots for bioimpedance testing versus standard practice, and the corresponding CEACs, are presented in Figures 18–21 for scenarios 1 and 3 (including dialysis costs). The corresponding scatterplots and CEACs with dialysis costs excluded are presented in Figures 22–25.

FIGURE 18.

Incremental cost-effectiveness scatterplot: BCM vs. standard care (clinical effectiveness scenario 1, including dialysis costs).

FIGURE 19.

Cost-effectiveness acceptability curves: BCM vs. standard care (clinical effectiveness scenario 1, including dialysis costs). WTP, willingness to pay.

FIGURE 20.

Incremental cost-effectiveness scatterplot: BCM vs. standard care (clinical effectiveness scenario 3, including dialysis costs).

FIGURE 21.

Cost-effectiveness acceptability curves: BCM vs. standard care (clinical effectiveness scenario 3, including dialysis costs). WTP, willingness to pay.

FIGURE 22.

Incremental cost-effectiveness scatterplot: BCM vs. standard care (clinical effectiveness scenario 1, excluding dialysis costs).

FIGURE 23.

Cost-effectiveness acceptability curves: BCM vs. standard care (clinical effectiveness scenario 1, excluding dialysis costs). WTP, willingness to pay.

FIGURE 24.

Incremental cost-effectiveness scatterplot: BCM vs. standard care (clinical effectiveness scenario 3, excluding dialysis costs).

FIGURE 25.

Cost-effectiveness acceptability curves: BCM vs. standard care (clinical effectiveness scenario 3, excluding dialysis costs). WTP, willingness to pay.

EMBASE Classic and EMBASE

Date range searched: 1947 to week 40 2016.

Epub ahead of print, MEDLINE In-Process & Other Non-Indexed Citations, MEDLINE(R) Daily (via Ovid) and MEDLINE(R) (via Ovid)

Date range searched: 1946 to 10 October 2016.

Ovid multifile search, URL: https://shibboleth.ovid.com/

Date searched: 10 October 2016.

Science Citation Index

Date range searched: 1970 to 27 June 2016.

Web of Knowledge URL: http://wok.mimas.ac.uk/

Date searched: 27 June 2016.

The Cochrane Library

Cochrane Database of Systematic Reviews: Issue 6 of 12, June 2016.

Cochrane Central Register of Controlled Trials: Issue 5 of 12, May 2016.

URL: www3.interscience.wiley.com/

Date searched: 27 June 2016.

Database of Abstracts of Reviews of Effects: December 2014

Centre for Reviews and Dissemination, URL: http://nhscrd.york.ac.uk/welcome.htm

Date searched: 27 June 2016.

Additional conference proceedings

ERA-EDTA Congress 2014, Amsterdam, the Netherlands, 31 May to 3 June.

ERA-EDTA Congress 2015, London, UK, 28–31 May.

Kidney Week (Journal of the American Society of Nephrology) American Society of Nephrology 2014, Philadelphia, PA, USA, 11–16 November.

Kidney Week (Journal of the American Society of Nephrology) American Society of Nephrology 2015, San Diego, CA, USA, 3–8 November.

Annual Dialysis Conference 2014, Atlanta, GA, USA, 8–11 February.

Annual Dialysis Conference 2015, New Orleans, LA, USA, 31 January to 3 February.

Annual Dialysis Conference 2016, Seattle, WA, USA, 27 February to 1 March.

Clinical Trials (June 2016)

URL: http://clinicaltrials.gov/ct/gui/c/r

Date searched: 4 July 2016.

World Health Organization’s International Clinical Trials Registry Platform (June 2016)

URL: www.who.int/ictrp/en/

Date searched: 4 July 2016.

European Union Clinical Trials Register (June 2016)

URL: www.clinicaltrialsregister.eu/

Date searched: 4 July 2016.

EMBASE Classic and EMBASE

Date range searched: 1947 to week 39 2016.

Epub ahead of print, MEDLINE In-Process & Other Non-Indexed Citations, MEDLINE(R) Daily (via Ovid) and MEDLINE(R) (via Ovid)

Date range searched: 1946 to 2016.

OVID multifile search, URL: https://shibboleth.ovid.com/

Date searched: 27 September 2016.

EMBASE Classic and EMBASE

Date range searched: 1947 to week 27 2016.

Epub ahead of print, MEDLINE In-Process & Other Non-Indexed Citations, MEDLINE(R) Daily (via Ovid) and MEDLINE(R) (via Ovid)

Date range searched: 1946 to 2016.

OVID multifile search, URL: https://shibboleth.ovid.com/

Date searched: 5 July 2016.

Health Technology Assessment, June 2016 and NHS Economic Evaluation Database, December 2014

Centre for Reviews and Dissemination, URL: http://nhscrd.york.ac.uk/welcome.htm

Date searched: 5 July 2016.

Research Papers in Economics

URL: http://repec.org/

EMBASE Classic and EMBASE

Date range searched: 1947 to week 27 2016.

Epub ahead of print, MEDLINE In-Process & Other Non-Indexed Citations, MEDLINE(R) Daily (via Ovid) and MEDLINE(R) (via Ovid)

Date range searched: 1946 to 2016.

Ovid multifile search, URL: https://shibboleth.ovid.com/

Date searched: 8 July 2016.

Cost-effectiveness Analysis Registry: July 2016

URL: https://research.tufts-nemc.org/cear4/default.asp

Date searched: 8 July 2016.

School of Health and Related Research Health Utilities Database: July 2016

URL: www.scharrhud.org/

Date searched: 8 July 2016.