Costs and outcomes of increasing access to bariatric surgery for obesity: Cohort study and cost-effectiveness analysis using electronic health records

Setting

The setting for the study was the general population with obesity, aged ≥ 20 years, in the UK.

Perspective

The research took the perspective of the NHS. Health service costs were included but social care costs were excluded. Wider societal costs, including changes in productivity, which are hard to estimate precisely, were not included. The research adopted a lifetime time horizon. Utilisation of bariatric surgical procedures in the private sector was not explicitly considered, though evidence emerged that a substantial proportion of procedures recorded in primary care electronic records may be performed privately.

Outcomes

The primary outcome was QALYs, after taking into account the incremental costs associated with intervention. Some simulations might be associated with negative incremental costs (where the intervention is cost saving) or negative incremental QALYs (as when standard care dominates), net health benefits40 were, therefore, estimated as:

where the threshold is the maximal acceptable value of cost per QALY; in the UK this is often taken as £30,000 per QALY. 41 Cost-effectiveness acceptability curves were also constructed using a range of values for the threshold.

Data source

The UK CPRD provided the source of EHRs for this study. The CPRD is the world’s largest primary care database, comprising anonymised longitudinal patient records from UK family practices. The CPRD presently holds more than 80 million person-years of research-quality data from 1990 onwards from more than 600 family practices. Data held within CPRD are considered to be broadly representative of the UK population based on the demographic characteristics of patients and the size and distribution of practices. 42,43 Scientific and ethics approval of the protocol for the study were given by the CPRD Independent Scientific Advisory Committee (ISAC 13_089).

Patient and public involvement

We organised a meeting for patients who had undergone bariatric surgery at St George’s Hospital to give an overview of the research, elicit any thoughts or concerns they had over the project and receive any potential research ideas. Interim findings were disseminated through new media as well as through scientific publications and meetings.

Bariatric surgery cohort

The sample comprised a cohort of adult obese patients with first bariatric surgery procedures performed, including all participants with LAGB, GBP or SG recorded before 30 April 2014. The date of the first procedure was taken as the index date. The earliest index event was in 2002. Participants were excluded if they were < 20 years of age. A lower age limit of 20 years was specified in the study protocol because we aimed to evaluate the use of bariatric surgery in adults. Use of bariatric surgery for individuals in their teens may be viewed as an outcome of childhood obesity and requires separate consideration. The mean age of 14 patients excluded was 18.5 years, range 17–19 years. Procedures recorded within 1 year of the participant start date in CPRD were also excluded because such records might refer to procedures performed before the patient’s registration at a CPRD practice. (When a patient joins a general practice, their previous notes are summarised and any significant diagnoses noted. A record of bariatric surgery within 12 months of joining a practice may refer to a procedure performed before the patient joined the practice.) Participants were excluded if they did not have a BMI record for obesity (BMI ≥ 30 kg/m²) prior to surgery. A minimum BMI value of 30 kg/m² was employed to ensure that all participants were obese, but some BMI records dated from several years before operation and might not reflect preoperative BMI. Baseline BMI values were recorded a median of 1.6 years [interquartile range (IQR) 0.6 to 5.4 years] before surgery. Participants with gastric band removal recorded before the index date were also excluded. Additional exclusion criteria were employed in selected analyses, which are described in Figure 2. In the estimation of rates of bariatric surgical procedures, six participants with more than one procedure type coded on the index date were also excluded. In the analysis of T2DM incidence following bariatric surgery, participants with diabetes diagnosed on or before the index date were also excluded. In the analysis of bariatric surgery and diabetes remission, we included only participants diagnosed with T2DM prior to the index date. A diagnosis of T2DM was taken as the earlier of a medical diagnosis of diabetes, a prescription for antidiabetes medicines or a glycated haemoglobin (HbA_1c) value of 48 mmol/mol (> 6.5%). Participants who were diagnosed with polycystic ovary syndrome and prescribed diabetes medicines, but not diagnosed with diabetes, were excluded from a diabetes diagnosis. Participants who were ever recorded as having a medical code for gestational diabetes were also excluded from these analyses. Blood glucose values were not used for diagnosis, nor assessment of remission, because distinctions between fasting or post-meal values are rarely clear in clinical records.

FIGURE 2.

Selection of bariatric surgical cohort. BS, bariatric surgery.

Reliability study

A sample of 102 participants who had bariatric surgery recorded in their CPRD records was selected for a reliability study in which EHR data were compared with general practitioner (GP)-reported information. The sample selected for study included approximately equal numbers of participants with EHR records for LAGB, SG or GBP. LAGB patients who had records of gastric band removal and GBP and SG patients who had repeat procedures were oversampled. Participants specifically selected included all 16 patients with GBP or SG who had repeat procedures recorded; all 43 LAGB patients who had band removal recorded and the initial procedure was in 2008 or later; and a random sample of 40 each with GBP and SG who had index date in 2008 or later. Participants with procedures from 2008 or later were selected because these were more likely to remain registered with the same general practice than patients treated longer ago. This gave 139 participants, of whom one was duplicated between the first and third criterion, and 138 participants remained eligible for the validation study. However, 36 were excluded because their general practice no longer contributed to CPRD, leaving 102 for further evaluation. The GP for each patient was sent a questionnaire (see Appendix 1) which included items concerning whether or not the patient had bariatric surgery, the date of surgery, the type of procedure, complications experienced, gastric band removal, operation reversal and repeat procedures. Multiple reminders were sent in order to optimise the response rate.

General population cohort stratified by body mass index category

We sampled a second cohort of participants from CPRD to act as a general population comparison sample. The sample was drawn from the list of all acceptable patients who were registered with CPRD practices that contributed to the data linkage scheme at any time between 1 January 2008 and 31 December 2014. For each participant, the mean of all of his or her BMI records was estimated. Then a stratified random sample was taken of up to 50,000 participants in each category of BMI: normal weight (18.5–24.9 kg/m²); overweight (25.0–29.9 kg/m²); obese (30.0–34.9 kg/m²); severe obesity (35.0–39.9 kg/m²); morbid obesity (40.0–44.9 kg/m²); and super-obese (≥ 45.0 kg/m²). Full CPRD data records were then extracted for this sample. There were 1819 participants who ever had bariatric surgery recorded who were excluded from the sample, together with one participant with undefined gender, leaving 257,187 patients for further analysis. The distribution of the sample by gender and BMI category is shown in Table 1. For selected analyses, we further restricted this sample to those participants who were eligible for linkage of IMD quintile and linked HES data to CPRD records. There were 250,046 participants eligible for HES data linkage and 247,537 eligible for linkage of IMD 2010 deprivation category.

The start was the later of 1 January 2008, the participant registration date, or the general practice CPRD start date. The end date was the earliest of 31 October 2014, the date death or end of registration, or the last data collection date for the general practice.

Statistical analysis methods

The aim of statistical analyses was to provide empirical inputs requires for the Markov model.

Body mass index probability analysis

In order to estimate the probability of a person transitioning to normal body weight in the absence of bariatric surgery, a longitudinal analysis of BMI records was conducted using a general population sample as outlined above. The start date for each participant was the later of 1 November 2004 or the beginning of the patient’s CPRD record. The end date was the earlier of 31 October 2014 or the end of the patient’s CPRD record. The first BMI record after the participant start date was used as the index BMI and the date of this record was used as the index date. The number of BMI records was evaluated for each BMI category and the number of records showing either an increase or a decrease in BMI category was calculated. For patients who showed a decrease in BMI category, we evaluated whether subsequent changes in BMI category represented increases or further decreases. Data were analysed in a time-to-event framework to evaluate, first, the proportion of patients from each starting BMI category who attained either normal body weight, or, second, a 5% reduction in body weight during 9 years’ follow-up. A 5% reduction in body weight was not envisaged in the original study protocol but was added as a minor amendment because this is a widely recommended target for body weight reduction. 35 In the first analysis, the annual probability of attaining normal body weight was estimated. The number of events (BMI category recorded as < 25 kg/m²) and the person-years of follow-up were used to estimate the annual rates, and their CIs, which were converted to annual probabilities using the formula, 1-exp^-rate. Among participants who reduced BMI category, the direction of the next change in BMI category was evaluated. In the second analysis, to examine the participants who had lost 5% of body weight, the development of a body weight that was more than 95% of the initial body weight was also evaluated in a time-to-event framework. Analyses were conducted in Stata version 13 using the stset, sts list and stcox commands.

Health-care costs in obesity

To assess costs related to obesity in the general population sample, person-time was allocated to BMI category according to the most recent BMI value, combining records ≥ 40 kg/m² into a single group to represent morbid obesity. Comorbidity status was evaluated using medical diagnoses coded into EHRs during general practice consultations. Person-time was classified according to morbidity status using the first diagnosis of T2DM, CHD, stroke or cancer. Depression was re-evaluated in each year of follow-up and patients were considered to have depression if they were diagnosed with depression in year or if they were prescribed antidepressants in year and were ever diagnosed with depression. Morbidities, including T2DM,48 CHD,49 stroke50 and depression,51 were evaluated using medical codes reported previously, while cancer diagnoses were evaluated using codes for malignant neoplasms.

Health-care utilisation was estimated from participants’ EHRs, with linked HES data. Primary and secondary care utilisation was evaluated including primary care consultations at the practice, by telephone, at home, emergency and out-of-hours. Secondary care utilisation included admissions to hospital, outpatient, day-case and emergency visits. All drug prescriptions issued by the practice were evaluated. Utilisation rates were calculated using person-time at risk. Age-standardised rates were estimated using direct standardisation and the European Standard Population for reference.

The costs of health-care utilisation were evaluated for participants by morbidity and depression status within BMI categories for the period 2008 to 2013. The annual costs were estimated by multiplying the health-care utilisation associated with each state by the costs of each unit of health care. Unit costs were derived from reference sources based on 2013 price estimates. The Personal Social Service Research Unit (PSSRU) publication Unit Costs of Health and Social Care 2013 was used as the main reference source (Table 3). 52 The same unit costs were applied across different ages, genders, BMI categories and morbidity status. Primary care GP consultation, emergency or out-of-hours, home visit and telephone consultations were priced at £45, £45, £114 and £27, respectively. Unit costs of secondary care inpatient, outpatient, day-case and emergency visits were priced at £1400, £135, £697 and £135, respectively. To assess prescription costs, drug codes for prescriptions in the EHRs were linked with costs from a dictionary compiled by RESIP UK (RESIP UK, Chertsey, Surrey, UK).

A two-part model53,54 was used to analyse health-care utilisation costs. In the first stage, a probit model was employed to estimate the probability of health-care utilisation being non-zero. In the second stage, a generalised linear model (GLM) with a log link and gamma errors was used to evaluate the distribution of costs in participants who utilised health care. This approach provided estimates of the predicted mean costs for men and women in different BMI and morbidity categories for each year of age. In the final stage of analysis, a linear regression model was employed to estimate the effects of BMI category, comorbidity and depression on predicted health-care costs controlling for patient gender and age. Interaction terms for comorbidity and depression and for comorbidity and BMI category were included. In order to make the data sufficiently concise for presentation, diabetes, CHD, stroke and cancer were combined into a single category of ‘comorbidity’ present or absent for the linear regression analysis. We combined ‘comorbidity’ into a single category in order to facilitate interaction terms with BMI status.

Epidemiology of bariatric surgery

The bariatric surgical cohort was analysed in order to understand the epidemiology of bariatric surgery over the period of study. The rate of utilisation of bariatric surgical procedures was estimated for men and women and for three age groups: 20–34 years, 35–54 years and 55–84 years. These represented young, middle-aged and older adults. The denominator was person-years at risk for the general population registered in CPRD. Participants were classified, according to the procedure recorded on the index date, into LAGB, SG or GBP. Utilisation of the three procedures, as a proportion of all bariatric surgical procedures, was evaluated from 2002 to 2014. Bariatric surgical codes recorded after the index date were evaluated and participants were classified as having a second operation if a procedure of a different type was recorded more than 30 days after the index date. In participants whose initial procedure was LAGB, we evaluated whether or not a code for removal of gastric band was recorded. The occurrence of repeat operations and band removal were evaluated in a time-to-event framework and annual incidence rates were estimated. Records of body weight, height and BMI were identified in order to estimate changes in body weight following the index date.

Type 2 diabetes mellitus incidence

The bariatric surgical cohort was analysed to evaluate the incidence of diabetes following the procedure. Matched controls for comparison were drawn from the general population cohort, which included 103,502 obese non-diabetic individuals sampled from CPRD who did not receive bariatric surgery and were not older than the maximum age of the bariatric surgery participants. Controls were matched for age, BMI, sex, index year and HbA_1c category. The maximum HbA_1c value before the index date was included using the categories < 42 mmol/mol (6%), 42–47 mmol/mol (6.0–6.49%) and not known. Nearest neighbour matching was performed without replacement. 14 The index date for controls was the date of the earliest BMI record on which the patient attained their highest BMI category.

For bariatric surgery participants and controls, new diagnoses of clinical diabetes were identified using medical diagnoses, drug prescriptions and HbA_1c values. Participants were identified as being diagnosed with clinical diabetes if a medical code for diabetes was recorded or if insulin or oral hypoglycaemic drugs were prescribed or, using World Health Organization criteria,55 if a HbA_1c value was 48 mmol/mol (> 6.5%). Oral hypoglycaemic drugs included sulphonylurea drugs, metformin, acarbose, dipeptidyl-peptidase-4 inhibitors, glitazones and glinide drugs. Participants with recorded diagnoses of polycystic ovary syndrome, who were prescribed antidiabetic drugs but never diagnosed with diabetes, were coded as non-diabetic. The date of the earliest medical, therapy or test event was taken as the diabetes diagnosis date. All new diagnoses of diabetes were included because different diabetes phenotypes cannot always be clearly distinguished in clinical practice. We evaluated new cases of diabetes for the recording of codes for type 1 diabetes, for prescription of insulin within 6 months of the diagnosis date and for diagnoses of gestational diabetes. The latter were excluded as a sensitivity analysis. The index BMI was the most recent recorded value prior to the index date. Records for smoking status, blood pressure and cholesterol were identified and the most recently recorded prior to the index date were used to describe baseline values. Depression, CHD and stroke were identified using previously described medical codes. 49–51

Baseline characteristics of the bariatric surgery participants and controls were described. A time-to-event framework was used to assess T2DM onset, using the Cox proportional hazards model. Failure was a new diagnosis of T2DM. Records were censored at the end of participants’ registration, the last date of CPRD data collection or death. Follow-up was censored after 7 years because few participants remained. Models were adjusted for matching variables, prevalent CHD, stroke and depression ever diagnosed, smoking status, whether or not total cholesterol was > 5 mmol/l, blood pressure > 140/90 mmHg, use of statins and antihypertensive drugs before surgery. Quadratic terms for age and BMI did not improve goodness of fit. Indicator variables were used for missing data on blood pressure and cholesterol. However, 591 (27%) of controls and 18 (1%) of bariatric surgery participants had missing values for blood pressure and 1466 (68%) of controls and 557 (26%) of bariatric surgery participants had missing values for cholesterol. The proportional hazards assumption was evaluated with no evidence that this was violated. Robust variance estimates were employed to allow for clustering of responses by family practice. Several sensitivity analyses were performed: using the unmatched cohort of 103,502 obese non-diabetic individuals for comparison; excluding participants diagnosed with diabetes within 12 months of the index date; and excluding participants diagnosed with gestational diabetes. An analysis was performed to evaluate the impact of competing risk from mortality using the method of Fine and Gray. 56

Diabetes treatment

The bariatric surgical cohort was used to evaluate changes in T2DM control and remission after surgery by selecting participants who had diabetes diagnosed prior to the index date. Control participants were selected from the general population cohort who never had bariatric surgery recorded, but were obese and had T2DM diagnosed before the index date. As the distribution of BMI differed greatly between bariatric surgery cases and the CPRD obese population, control participants were individually matched with cases using nearest neighbour matching on BMI, age, sex and index year. The index date for controls was the date of the first BMI record on which they entered their highest recorded BMI category. Participant records ended if they terminated their registration with a CPRD general practice; if their general practice ended participation in CPRD; if the latest data collection date was reached; or if the patient died.

Glycated haemoglobin records and prescriptions for oral hypoglycaemic drugs and insulin were evaluated for bariatric surgery cases and controls. The person-time for each participant was divided into study years from 3 years before the procedure to 6 years after the procedure. This allowed us to conduct an interrupted time-series analysis. Follow-up was censored at 6 years because few cases remained under follow-up for SG and GBP. The highest HbA_1c value and the total number of diabetes prescriptions were evaluated in each study year. For each year of follow-up, participants were classified as being in remission if the maximum HbA_1c value recorded in year was 48 mmol/mol (< 6.5%) and there were no diabetes prescriptions issued in the year. This is consistent with the definition of ‘complete and partial remission’ suggested by Buse et al. 57 Complete remission [HbA_1c 42 mmol/mol (< 6.0%) and not on medication] was also evaluated. Relative rates were estimated for each year following the bariatric surgery procedure by using a Poisson model with person-time as the exposure. Interpretation of these results requires consideration of the relationship between year of procedure and duration of follow-up, with longer duration follow-up being available only for patients operated on longer ago. A model was fitted to evaluate the effect of group (bariatric surgery or control) and time after surgery, included as indicator variables for each postoperative year. 58 Confounders included age, gender, BMI, quartile of diabetes duration before surgery, whether CHD, stroke or depression were diagnosed before the index date, whether blood pressure was ≥ 140/90 mmHg or serum total cholesterol was ≥ 5 mmol/l, and whether or not antihypertensive drugs and lipid-lowering drugs were prescribed before the index date, and smoking status recorded before the index date. Missing values for covariates were accounted for using indicator variables.

Changes in depression

A controlled interrupted time-series study was conducted, drawing on data from the bariatric surgery cohort and matched control participants from the general population cohort, by evaluating multiple time points both before and after the bariatric surgical procedure. EHRs from primary care were evaluated for clinical diagnoses of depression and prescription of antidepressant drugs in the 3 years before, and 7 years after, bariatric surgery.

Clinical depression was identified through medical diagnoses for depression recorded in clinical or referral records as well as through prescriptions for antidepressant drugs. Participant records were divided into person-years before and after surgery. Individuals were classified as having clinical depression in a given person-year if they had a medical diagnosis of depression recorded in that year, or if they were prescribed antidepressant drugs in the year and were ever diagnosed with depression. 51 BMI was categorised using the World Health Organization criteria55 using the most recent record prior to the index date. Comorbidities were evaluated including stroke, CHD and T2DM. Records for hypertension (blood pressure ≥ 140/90 mmHg), high cholesterol (total cholesterol ≥ 5 mmol/l) and current smoking status were evaluated. Participants were also classified according to whether they were treated with antihypertensive or lipid-lowering medications.

Baseline characteristics in the surgery and control groups were compared. Participant records were divided into 1-year periods from up to 3 years before, to a maximum of 7 years after, the index date. The presence or absence of clinical depression was evaluated for each 1-year period. A multiple logistic regression analysis was conducted using person-years as observations and the presence or absence of clinical depression in each year as the outcome. A model was fitted to evaluate the effect of group (bariatric surgery or control), study year from 3 years before to 7 years after surgery, and time after surgery, included, for improved goodness of fit, as indicator variables for each postoperative year. 58 The reference category was all person-time without surgery from the control group and the bariatric surgery group before operation. The model was adjusted for gender, age, baseline BMI, index year, type of bariatric surgery procedure, prevalent diabetes mellitus, CHD and stroke, smoking status, high blood pressure and cholesterol, and treatment with antihypertensive or lipid-lowering drugs. Robust variance estimates were employed to adjust standard errors for clustering of person-years by participant.

Data analysis to inform health economic modelling

Data to populate the health economic model were derived using epidemiological analysis of the general population cohort, among whom obese patients may have received weight management interventions available in routine primary care. The incidence, prevalence, mortality and costs of health-care utilisation in this group were estimated for each state in the model. Incidence and mortality rates were estimated in a time-to-event framework using a Weibull model. Covariates were age, age squared, gender and BMI category. The prevalence of depression was estimated for each state in the model. 51 Health-care utilisation and associated costs were estimated for each state in the model as described above.

Purpose

The purpose of the review was to systematically retrieve, appraise and synthesise available research evidence. We aimed to provide information for the model by determining the projected short- and long-term effects of bariatric surgical procedures on health states included in the model.

Eligibility criteria

We conducted a rapid review of previous systematic reviews that evaluated bariatric surgery as an intervention for obesity. Subsequently, we identified literature not included in the reviews because of later publication dates or restrictive inclusion criteria of reviews. A single literature search was conducted to identify both systematic reviews and primary reports. We included systematic reviews, controlled clinical trials and controlled observational studies published in English. Modelling studies were included for cost-effectiveness evaluations. Participants were obese adults. The intervention was a bariatric surgical procedure; and a comparator of standard care or comparisons between different surgical procedures were accepted. The study outcomes were incidence of T2DM, CHD, stroke and cancer; remission of T2DM and changes in depression; long-term mortality (more than 30 days post surgery); and cost-effectiveness.

Search strategy

We used the following search terms: (obes* or overweight) and (bariatric surgery or weight*loss surgery or gastric band or gastric bypass or sleeve gastrectomy) and (diabetes or coronary heart disease or CHD or stroke or depression or cancer) or (cost-effectiveness or cost-utility). We searched for papers published after 2000 as there were sufficient recent systematic reviews. The searches were implemented in EMBASE (1980–present), MEDLINE (1946–present), MEDLINE In Process & Other Non-Indexed Citations, PubMed and the Cochrane Database of Systematic Reviews. Reference lists of included studies were checked for further publications.

Study selection and data extraction

The search results were checked for eligibility by one reviewer. The full text was consulted when eligibility was not evident from the abstract alone. Data were abstracted by one reviewer and checked by a second reviewer. For disease incidence, T2DM remission and mortality, the summary measure was ideally presented as a risk ratio, although we also considered odds ratios (ORs) and hazard ratios (HRs). We also included publications which presented rates in the intervention and control groups. Changes in continuous measures, such as changes in fasting blood glucose in T2DM or depression scales, were not included. For the cost-effectiveness data the summary measures were incremental cost-effectiveness ratios (ICERs) or costs per QALY. The data were abstracted by one reviewer into a spreadsheet.

Quality assessment

The AMSTAR (Assessing the Methodological Quality of Systematic Reviews) checklist was used to assess the quality of systematic reviews. 59 The assessment of study quality is presented in Table 4.

Data synthesis

A narrative synthesis was conducted summarising results included in systematic reviews and summarising primary research studies of significance.

Model structure

A Markov model was employed to conduct a cost–utility analysis to compare a strategy in which all eligible participants received bariatric surgery, with standard non-surgical weight management. A diagrammatic representation of the Markov model is shown in Figure 3. The model structure was informed by previously reported research86 and enabled us to include the main influences on costs and outcomes of bariatric surgery identified in the literature search. Healthy subjects, referred to as ‘at risk’, may develop one of the disease states of interest, including T2DM, CHD, stroke and cancer. Onset of disease was irreversible except for diabetes which was allowed to remit with individuals returning to the ‘at risk’ state. Individuals who entered remission then had the same risk of diabetes as all others in the ‘at risk’ state. Participants in each state were allowed to progress to depression, with each main state divided into substates representing ‘not depressed’ and ‘depressed’. Depression was included because it occurs frequently in chronic illness and is associated with higher health-care costs. 51 The model was stratified by BMI category, comprising morbid obesity, severe obesity, simple obesity, overweight and normal weight, and allowed participants to transition between BMI categories. There were, therefore, 50 states in the model that represented all potential combinations of the included BMI categories, morbidities and depression. Each state was further stratified by single year of age and gender. All states might lead to death. The perspective of the model is that of health-care services and only health-care costs were included. Inputs to the model for the standard care group were derived from the sources presented in Table 5.

FIGURE 3.

Simplified schematic diagram of Markov model. BS, bariatric surgery; DM, diabetes mellitus.

Model estimation

The probabilistic Markov model was estimated by cohort simulation, implemented through a program written in R software (The R Foundation for Statistical Computing, Vienna, Austria). The initial population had ages ranging from 20 to 74 years, as we observed that there were few bariatric surgical procedures over the age of 74 years in CPRD. The proportion of the start population with morbidity was also informed by analysis of a CPRD cohort.

All simulations were stratified by single year of age with the initial population ageing by 1 year per cycle. Participants exited the model when they died or reached 100 years of age. The model was run for each sex separately. Outcomes and costs were compared for scenarios in which all participants received bariatric surgery in the first annual cycle of the model or no participants received bariatric surgery. Annual transition probabilities for the model were obtained by sampling from the beta-binomial distribution, using CPRD data as inputs. The costs of each state were sampled from the gamma distribution, with the predicted mean value estimated from a two-part model as outlined above. Utilities for each state were obtained from data published in a compendium of values. 87 Utility values for each state were stratified by single year of age but were the same for men and women. Utility values were sampled from the beta distribution. The utility values used in the model are presented in Table 6. Total costs and QALYs were obtained by summing across the 81 cycles of the model included in each simulation. There were 1000 simulations run for each scenario. Results are expressed as rates per 1000 participants entering the model. Mean costs, and the 95% range, were obtained from the data for 1000 simulations. Costs and QALYs were discounted using a rate of 3.5%, but undiscounted values and values discounted at 1.5% are also shown as sensitivity analyses. 88 Net monetary benefits and net health benefits were estimated at threshold values of £20,000 and £30,000 per QALY. 88

Intervention effects and costs of bariatric surgery

The effect of bariatric surgery was modelled as a reduction in disease incidence and mortality. The effect of bariatric surgery on the incidence of T2DM was drawn from CPRD data analyses7 that gave very similar results to data from the Swedish Obese Subjects (SOS) study. 9 Effects on incidence of CHD, stroke and cancer were also drawn from the SOS study,9,89 which showed a reduction in cancer incidence in women but not in men. 89 The effect on depression prevalence was drawn from CPRD data analyses90 and is also consistent with other reports. Based on CPRD data analyses (see Chapter 8), 40% of T2DM patients were estimated to enter remission following the procedure. Bariatric surgery was modelled as being associated with a positive impact on patient utility equivalent to a two-unit change in BMI category. 6,91 However, this effect was modelled to decline over time, according to year^–0.25, consistent with the known reduction in the initial quality of life improvement following bariatric surgery. 92 The costs of bariatric surgery were drawn from NHS tariffs and included preoperative weight management, the cost of the procedure, and postoperative reviews (Table 7). Bariatric surgery was assumed to comprise one-third each of gastric banding, GBP and SG. The cost of leaks was included as an average cost across all patients. 93 Two per cent of patients were assumed to require repeat procedures each year, slightly higher than the 1.2% observed in CPRD records. Mortality from surgery was estimated at 0.07% from the NBSR report. 11 Costs of health-care utilisation were estimated from CPRD. Costs of health-care utilisation following bariatric surgery were determined by age, sex and morbidity category, and were not modelled as associated with body weight reduction, consistent with the results of empirical studies. 12,13

Subgroup and sensitivity analyses

In the base case we investigated the effect of bariatric surgery in participants with BMI ≥ 40 kg/m², compared with no participants undergoing surgery. This was because the majority of bariatric surgery procedures are performed for morbid obesity at present. The initial population included 19% with T2DM and 4% with CHD, similar to the distribution observed in CPRD at the start of the study period (see Chapter 6). Additional simulations were performed to estimate costs and outcomes separately for men and women; for separate age groups, including 20–34 years, 35–54 years and 55–64 years; and categories of deprivation, comparing the most- and least-deprived quintiles of deprivation according to the IMD 2010 score. IMD scores were linked at patient lower super-output area level. 94 A summary of the simulations performed is presented in Table 8.

Sensitivity analyses were implemented to explore the effects of varying the unit costs of bariatric surgery, including values 50% and 100% higher than the base case; varying the discount rate including values of 0%, 1.5% and 3.5% following NICE recommendations; and to estimate outcomes assuming that intervention effects following bariatric surgery might diminish with time following the operation. This was implemented by allowing intervention effects from bariatric surgery to diminish by year^–0.25 or year^–0.5. The former implies that the effect of bariatric surgery will decline by 44% over 10 years, while the latter indicates that the intervention effect will decline by 68% over 10 years. Sensitivity analysis was also used to test intervention cost-effectiveness for patients with severe obesity (BMI 35–39 kg/m²) or with morbid obesity and T2DM.

Summary of findings

Analysis of primary care EHRs for a large population based sample of men and women over a 9-year period revealed that the probability of obese patients attaining normal weight was very low. The annual probability of patients with simple obesity attaining a normal body weight was only 1 in 131 for women and 1 in 225 for men. The likelihood of attaining normal body weight declined with increasing BMI category, with the lowest probability observed for patients with morbid obesity. The smaller group of patients with super obesity represented a departure from this trend, but nevertheless showed a low probability of attaining normal body weight. Although the probability of patients achieving a 5% reduction in body weight was considerably higher, the majority of these patients went on to regain lost weight, as evidenced by BMI records of > 95% of the initial value, within 2–5 years of the first record that was lower than 95% of the initial value.

These findings raise questions concerning whether or not current obesity treatment frameworks, grounded in weight management programmes accessed through primary care, may be expected to achieve clinically relevant and sustained reductions in BMI for the vast majority of obese patients and whether or not they could be expected to do so in the future. The lack of sustained BMI reductions could be driven by low intervention uptake rates or their lack of effectiveness. In a previous study, we reported that weight-loss interventions are currently offered only to a minority of patients in primary care. 28 Efforts are under way to improve this situation, with the proportion of patients with obesity offered multicomponent weight-loss interventions included among potential new indicators in the 2016/17 consultation for the Clinical Commissioning Group Indicator Set (CCG OIS). 98 However, even when treatment is accessed, evidence suggests that behavioural weight-loss interventions focusing on caloric restriction and increased physical activity are unlikely to yield clinically significant reductions in body weight. 29,99 A recent series of reviews documented the limited progress in reversing the global obesity epidemic and called for regulatory actions from governments as well as co-ordinated efforts across industry and society to reduce obesity. 100–103 Dietz et al. 104 warn that preventative strategies are unlikely to reduce weight in people living with severe obesity and stress the need for changes in the delivery of care for these patients. In combination with previous research, this study highlights the current failures in combating existing obesity cases at a population level.

Comparison with other results

Reductions in BMI category were observed more frequently among patients with a higher baseline BMI but these decreases were more likely to be followed by subsequent increases rather than further decreases or stability in BMI category. Weight cycling, evidenced by both increases and decreases in BMI category, was most common among men and women with baseline BMIs in the morbidly obese category. Greater instability in weight trajectories among patients with higher BMIs has been reported previously. 105 Weight cycling has been linked to a higher risk of morbidity and mortality compared with stable obesity,106–108 although evidence of causality remains inconclusive. 109

The higher likelihood of decreases in BMI category and of 5% weight loss among the more severely obese participants in this study is consistent with results from clinical trials110 and previous cohort studies111 in which higher BMI predicted greater weight loss. The increased probability of weight reduction among patients with more severe obesity may reflect more accurate perceptions of personal weight status112,113 and higher treatment rates among these patients. It is also possible that BMI decreases in severely obese patients reflect unintentional weight loss resulting from greater comorbidity. The finding that a high proportion of patients in this analysis experienced a period of weight regain following weight loss is also consistent with previous research. At least 50% of patients who achieved 5% weight loss were shown to have regained this weight within 2 years. It has previously been reported that approximately 80% of people who intentionally achieve weight loss ≥ 10% of their body weight will regain that weight within 1 year. 114

Strengths and limitations

This analysis had the strengths of a large population-based cohort with prolonged follow-up.

Slightly lower numbers at the highest levels of BMI might have made these estimates slightly less precise. Data are presented for adults aged > 20 years. Inspection of age-specific values revealed, as expected, greater weight gain at younger ages and a somewhat greater tendency to weight loss at older ages. It was not possible to evaluate intentionality of weight loss. Previous studies have reported that the majority of obese individuals would like to lose weight and a large proportion are actively attempting to reduce their weight,115,116 so a relatively high level of intentionality among obese participants may be assumed. Additionally, monitoring of BMI among obese patients in primary care has been shown to positively predict treatment. 117 Patients in the present study were required to have a minimum of three BMI measurements recorded, suggesting that an inflated proportion of patients in this sample might have been involved in and interested in weight-management interventions. Nevertheless, we acknowledge that unintentional weight loss was also included and might result from physical disorders such as cancer, or psychological concerns such as bereavement. 118–120 Additional in-depth analyses might evaluate patterns of weight change in relation to comorbidity.

Recording of body weight in primary care is generally opportunistic and dependent on patients attending the practice. We acknowledge that weight measurements in EHRs may be associated with error and bias including measurement error; confounding by indication, if weight changes prompt weight measurements; variation between professionals and family practices in measurement recording;44 and weight-management strategies. 28 Higher patient baseline BMI was associated with a higher frequency of BMI measurements recorded over the study period. UK general practices have contractual financial incentives to provide a register of adult patients who have a BMI ≥ 30 kg/m² measured in the last 15 months121 which may lead to more frequent recording of BMI for obese patients. We reported on the recording of BMI in primary care in a previous study. 44 For this study, we selected participants with a minimum of three BMI records. We acknowledge that participants with fewer than three BMI records may show different patterns of weight change and the present results might be biased through their omission. However, we believe that this is one of the largest studies yet reported on body weight changes in the general population. The relatively high levels of comorbidity seen in obese compared with normal weight patients would also likely result in more regular consultations and more frequent recording of BMI. However, it is possible that patients from all BMI categories with three or more BMI measurements recorded over the 9-year study period represent a biased, less healthy sample than the general population. If this is the case then unintentional weight loss, along with comorbidities contributing to weight gain such as mobility impairment, may have influenced BMI changes disproportionately in the current sample.

Summary of findings

This analysis investigated health-care costs in relation to BMI and is the first we are aware of to isolate depression as a cost separate from other common obesity-related morbidities. Previous studies investigating health-care costs relating to obesity using patient cohorts have attempted to estimate the marginal effect of obesity on costs. 123,124 Understanding the causal association between BMI and health-care utilisation is important; however, in the context of treating patients in a clinical environment there is little benefit in divorcing obesity from related morbidities and socioeconomic factors.

We found that having a comorbidity was the greatest predictor of increased health-care costs, at £1366 per year, followed by clinical depression at £1044 per year. Of the four obesity-related morbidities we assessed, stroke was found to be associated with the highest absolute costs, which is likely to be a result of acute secondary care required after a stroke. This finding is supported by a previous paper that identified cardiovascular diseases and cardiovascular agents as the biggest drivers of health-care claims costs per unit of BMI. 125

We found that depression was the second greatest driver of costs and its presence increased costs in all states, in line with a previous investigation conducted into depression and health-care utilisation. 51 Overall, having high BMI in addition to a comorbidity appeared to increase costs exponentially, although having depression was an even greater contributor to high health-care costs alongside a comorbidity. In spite of this, being overweight was not associated with higher health-care costs than normal weight, and in fact, participants who were normal weight and diabetic had similar costs as diabetic participants who were morbidly obese. A majority of adults are now either overweight or obese. ‘Normal’ weight may sometimes result from weight loss associated with comorbidity and this might account for elevated health-care utilisation in this group. Therefore, diabetes and not just weight alone appears to be behind cost figures.

Comparison with other results

There is large variation in cost estimates for obesity based on population sample studies, ranging from an additional €160 per year for severe obesity compared with normal weight,124 to US$2741 for any obesity compared with normal weight. 123 Our findings estimated an increase of £146 per year for mild obesity in comparison with normal weight, and £456 for morbid obesity. These estimates were for the influence of BMI separate from the considerations of obesity-related morbidities and depression.

One unexpected finding was the high cost of health care in normal-weight diabetic patients compared with obese diabetics. In fact, this phenomenon has been reported elsewhere, and the possibility of greater morbidity in diabetics who are normal weight has been suggested as an explanatory factor. 126,127 The lack of statistically significant difference in costs between normal and overweight BMI categories has also been found in other studies. 128,129

Strengths and limitations

This analysis has a number of strengths that make it a valuable addition to the literature on health-care-related costs of obesity. First, it is based on a large, longitudinal, nationally representative data set and uses costs associated directly with provision of care rather than based on insurance claims, which are an indirect measure of morbidity and may be less accurate. It is also one of few analyses to use clinically measured height and weight values in the calculation of BMI rather than self-reported values which are prone to reporting bias. 130 This is also the only analysis we are aware of to use UK sample data to estimate the health-care costs associated with obesity. One possible limitation of the study is selection bias from inclusion of participants with a BMI recorded during a clinical consultation, who may be more frequent users of health-care services and, therefore, less healthy. Participants’ morbidity status was classified using the first diagnosis they received and we have not accounted for additional diagnoses. The costs associated with the conditions we have highlighted may represent costs from multiple morbidities, which are more frequent in obese patients and we know were present in 18% of the observations. 3

The economics literature has moved towards using an instrumental variables approach to examine the causal effect of obesity on medical costs by using the weight of a biological relative as an instrument. 123,124 The argument for this method is that the instrument predicts the participants’ weight, but not their morbidity status, meaning that the effect of weight on costs can be isolated. Such papers have found higher health-care costs for obesity than non-instrumented methods, so it is possible that our models underestimate the magnitude of the relationship.

Conclusions

The findings of this analysis emphasise that health-care costs for obesity are largely driven by a few key obesity-related morbidities and not by high BMI alone. Persistently high obesity rates and evidence from the literature7 suggest that preventative and reactive treatment strategies are currently of limited use. Until greater success is observed in lowering obesity levels, perhaps clinical focus should be on prevention of secondary conditions as a way of improving the health status of obese patients and lowering resulting costs. In particular, increasing uptake of successful diabetes prevention programmes131 may prove beneficial given the high costs observed across BMI categories in this analysis. Similarly, successful interventions to focus on the mental health of obese might also stem resultant health-care and non-health-care costs related to depression.

Reliability study results

Completed questionnaires were received for 78 patients (Table 16). All 78 responses confirmed that bariatric surgery had been performed on the date indicated in EHR data. The type of bariatric surgical procedure was confirmed for all 30 (100%) patients recorded with LAGB, for 24 out of 25 (96%) patients recorded with SG, and for 19 out of 23 (83%) patients recorded with GBP. Gastric band removal was confirmed for 27 out of 30 (90%) of cases. Among nine patients with second procedures recorded in EHRs following GBP or SG, six were confirmed in GP questionnaire responses. The most common complications after surgery reported by GPs were infection and wound site problems (13%), digestive issues including diarrhoea and vomiting (10%), pain (8%) and difficulties with the throat and swallowing (6%). High rates of complication in this group may be expected as patients requiring further procedures were oversampled. GPs reported that the surgery was privately funded in 32 (41%) of cases.

Utilisation of bariatric surgical procedures

The number of procedures recorded increased over time, with only 104 procedures recorded between 2002 and 2005; 607 between 2006 and 2008; 1406 between 2009 and 2011; and 922 between 2012 and April 2014 (Table 17). The rate of surgery was highest in men and women aged 35–54 years. Rates of bariatric surgical procedures by age group and gender are presented in Figure 6. Rates of bariatric surgery were greatest for women in 2010 at 37 per 100,000 population per year, and in 2012 for men at 10 per 100,000 population per year. Disparity between genders was greatest in the youngest patients, aged 20–34 years, with peak rates of 15 per 100,000 per year in women and 3 per 100,000 per year in men. LAGB was the most frequent procedure, accounting for 1297 (43%) of cases, followed by GBP in 1265 (42%) participants and SG in 477 (16%). LAGB accounted for 97% of 104 procedures performed from 2002 to 2005. The use of GBP and SG increased over time while LAGB declined (Figure 7). During 2012–14, GBP accounted for 55% of procedures, while SG accounted for 25% and LAGB accounted for 20% (see Table 17).

FIGURE 6.

Rates of first bariatric surgery procedures in a large primary care population for (a) men; and (b) women. The denominator is the population registered in CPRD. Blue line, 20–34 years; green line, 35–54 years; black line 55–84 years.

FIGURE 7.

(a) Trends in the utilisation of different bariatric surgical procedures from 2002 to April 2014; and (b) total number of procedures per year. Green line, LAGB; blue line, GBP; black line, SG.

Changes in case mix

Patient characteristics at the index date are presented in Table 17. The mean age at operation increased from 43.4 to 46.8 years during the study (p < 0.001), and the proportion of women declined from 86% to 75% (p < 0.001). The mean recorded BMI increased from 40.6 kg/m² to 44.8 kg/m² (p < 0.001). The proportion of participants with diabetes increased from 19% to 33%, while the proportion of patients prescribed antihypertensive drugs and statins also increased (all p < 0.001). More than half of all participants had depression recorded at some time before the procedure. As a consequence of these trends, there were important differences in case mix for patients undergoing LAGB as compared with GBP and SG (Table 18). LAGB patients were generally operated on in an earlier period, were younger, more often female, less obese and less likely to have diabetes, hypertension or hypercholesterolaemia.

Secondary procedures

There were three deaths within 30 days of the date of the initial procedure. Rates of band removal and reoperation following LAGB are presented in Table 19. The most common procedure was removal of a gastric band, found in 82 (6.3%) cases. This was equivalent to a rate of 1.6 (95% CI 1.3 to 2.0) per 100 person-years and the median time between gastric band insertion and removal was 144 weeks (IQR) 69–203 weeks). There were 60 (4.6%) of LAGB patients who had a subsequent medical code recorded indicating a GBP or SG procedure, with a rate of 1.2 (95% CI 0.9 to 1.5) per 100 patient-years. There were 10 patients who received SG, who later had codes for GBP recorded; and six patients with GBP, who later had codes for LAGB (n = 4) or SG recorded (n = 2). The validation study confirmed 86% and 33% of these reoperations, respectively, with two procedures secondary to SG attributed to hiatal hernia repair.

Long-term follow-up

Recording of BMI values into primary care EHRs after bariatric surgical procedures was generally poor. From 3039 participants, 486 had BMI values recorded in the first year following the procedure and 332 and 241 in the second and third years, representing 18%, 15% and 13% of those remaining under observation, respectively. Evaluation of body weight values indicated a mean reduction in BMI of 6.98 kg/m² (95% CI 6.3 to 7.6 kg/m²) in the first postoperative year and 9.99 kg/m² (95% CI 9.1 to 10.9 kg/m²) in the second postoperative year.

Summary of findings

This analysis provides a large-scale population-based evaluation of the utilisation of bariatric surgical procedures in the UK. The study complements a recently published bariatric surgical registry report,133 which presents data reported by bariatric surgeons. The present results demonstrate that EHRs, including those from primary care, represent a valuable resource for evaluating the utilisation and outcomes of bariatric surgery. The reliability study demonstrated high levels of agreement between primary care EHRs and questionnaire responses obtained directly from GPs. We caution that errors may be present in either data source, and we do not have evidence to show which is the more likely to be correct, but high levels of agreement between the two sources of information lend support to the validity of EHR data.

The rate of bariatric surgery recorded in primary care medical records increased rapidly between 2002 and 2014. Initially, LAGB accounted for most procedures, but the use of GBP and, to a lesser extent, SG has increased since 2008. There have been changes in case mix, with procedures now being performed in older patients, with greater BMI and a higher prevalence of diabetes.

Comparison with other results

The large increase in number of bariatric surgery procedures identified in CPRD over the last 10 years is consistent with findings reported from analysis of hospital utilisation statistics. 36 The gender disparity, age profile of surgery patients and changing patterns of surgery were also comparable with the trends seen in data for hospital utilisation and the bariatric surgical registry. 36,133,134 Depression was recorded in > 50% of participants at some time prior to surgery, a higher rate than reported elsewhere. 135 This difference may relate to using primary care rather than hospital setting as a data source, with a diagnosis of depression more likely to be recorded in the former.

Following gastric banding, gastric band removal was observed in 1.6% of patients per year and 1.2% per year were recorded as having a further additional procedure of GBP or SG, with a high level of validation from GP responses. These findings confirm in population-based data that there is significant incidence of band slippage or band intolerance requiring removal. The bariatric surgical registry recorded a much lower proportion of patients undergoing revisional bariatric surgery (0.3%) after gastric banding. 133 This discrepency may reflect the short period (3 years) covered by registry data, problems with data linkage occuring when reoperations are performed at different hospitals or under-reporting of reoperations and revisions.

There were 10 patients who underwent SG who later had codes for GBP recorded, and six patients who received GBP with subsequent codes for LAGB (n = 4) or SG recorded (n = 2). Validation confirmed the majority of these secondary procedures, but coding errors may account for some of these, as misclassification can arise through errors of recording in primary care records, especially when primary care physicians may be unfamiliar with different bariatric surgical procedures. Nonetheless, it is clear that surgical reintervention rate following either SG or GBP is lower than that following gastric banding.

Recording of weight data for bariatric surgery patients in primary care was poor. Clinical guidelines suggest that all bariatric surgery patients are expected to undergo lifelong follow-up to monitor their weight, as well as micronutrient monitoring in GBP patients. 35 Recording of weight and BMI in primary care records is known to be generally limited,44 but it is especially concerning that the majority of patients who have undergone bariatric surgery are not being weighed regularly by their GP. The observed maximum weight loss at 2 years is comparable with data from clinical trials,5,6,136 and gastric banding was associated with lower weight loss that other surgery types in accordance with the literature. 31,133 However, the high rate of attrition in this study means that caution is required when interpreting these results. Additionally, long-term follow-up in primary care may be biased by better attendance of patients whose surgery has proved less successful or who have remained overweight. 44

Strengths and limitations

This study had the strengths of a large nationally representative data source with extended periods of longitudinal follow-up. We acknowledge that clinical information has several limitations when used for research purposes, including missing data values due to opportunistic data collection and recording, but a reliability study suggested a high level of agreement between EHR records and GP reports for primary surgeries. Recording of body weight in primary care was poor subsequent to bariatric surgery.

Sensitivity analyses

The association of bariatric surgery with lower incidence of diabetes was found to be robust in a number of sensitivity analyses. When the entire comparison cohort of 103,502 obese non-diabetic individuals, rather than selected matched controls, was used for reference, the adjusted HR associating bariatric surgery with diabetes incidence was 0.16 (95% CI 0.11 to 0.22; p < 0·001). Diagnosis of diabetes might be more likely soon after the index date because of increased medical surveillance. However, when cases of diabetes diagnosed within the first year following the index date were excluded, the adjusted HR was 0.20 (95% CI 0.12 to 0.32; p < 0.001). Exclusion of participants associated with gestational diabetes diagnoses gave an adjusted HR of 0.19 (95% CI 0.13 to 0.30; p < 0.001). In a competing risks analysis, allowing for the competing risk of death, the adjusted HR was 0.20 (95% CI 0.13 to 0.31; p < 0.001). Analyses that used only clinical criteria, and not HbA_1c, for diagnosis of diabetes gave similar results.

Summary of findings

This analysis demonstrates that in a large population-based cohort of obese participants undergoing bariatric surgery using contemporary procedures, the risk of developing T2DM is substantially reduced over a maximum of 7 years of follow-up. The risk of diabetes was reduced by 80%, compared with control participants who did not undergo surgery. This is perhaps the first large-scale pragmatic study to evaluate the impact of current bariatric surgical procedures on diabetes incidence in the context of usual care settings. Even in patients observed in routine clinical practice, our results demonstrate that newer bariatric surgical procedures have particular efficacy for diabetes prevention in obese individuals.

An effect of bariatric surgery on diabetes incidence was observed in both men and women and across all ages. GBP and SG procedures were associated with slightly lower relative hazards for diabetes than LAGB but we caution against drawing firm conclusions concerning the comparative effectiveness of different procedures from a non-randomised study because selection for different procedures may be associated with the underlying risk of developing diabetes and some of these subgroup analyses were based on small numbers of outcome events.

Comparison with other results

Previous studies of bariatric surgery and the prevention of T2DM include the SOS study8 as well as several case series. 137–139 In the SOS study,8 69% of patients underwent vertical banded gastroplasty, 19% underwent banding and 12% received GBP surgery. Both the incidence of diabetes in bariatric surgical patients and controls, and the relative risk reduction associated with bariatric surgery, are very similar to those observed in the present study. The present results, therefore, provide important confirmation from a population-based sample, with intervention using current surgical procedures.

Strengths and limitations

Our analysis had the strengths of a large population-based sample with prolonged follow-up and prospective documentation of clinical diabetes in primary care. However, we acknowledge several limitations. We identified the three major procedures used in bariatric surgical practice in the UK. However, there may be less frequently performed procedures, such as the duodenal switch, which were not included. The control group did not receive standardised non-surgical intervention for obesity, and during the study period intensive multimodal weight-loss programmes in primary care were rare. Ascertainment of diabetes outcomes was comprehensive; diagnoses recorded into CPRD records are generally valid. 42 However, we were not able to document subclinical diabetes that might have been confirmed through testing all participants for evidence of hyperglycaemia. We excluded cases in which antidiabetes drugs were prescribed for polycystic ovary syndrome but it is possible that some prescriptions might be for diabetes prevention rather than treatment. Body weight, and other relevant measures, during the period of study were not recorded consistently. Preoperative BMI category might have been misclassified and it was not possible to relate differences in diabetes incidence to changes in body weight, though it should be noted that the effects of bariatric surgery are not entirely mediated by changes in body weight. 136 Observed weight reductions following surgery were consistent with those of previous reports.

Access to bariatric surgery is presently very limited in the UK and, consequently, those receiving surgery represent a highly selected group. The higher prevalence of previous depression diagnoses in the surgical group suggests that patients with depression were more likely to be referred with the belief that weight loss may improve their depressive symptoms. This belief is likely to be misplaced in the long term. 90 Patients who received surgery might have been more adherent than controls to other diabetes prevention advice including diet or exercise. However, we observed that participants who received surgery were more likely to be prescribed antihypertensive drugs or statins, which may sometimes be associated with diabetes. Control participants more often had missing values for blood pressure and plasma cholesterol, suggesting that medical surveillance was lower in this group. Differential medical surveillance might have made detection of diabetes or pre-diabetes more likely in patients considered for surgery. However, the higher proportion of surgical cases on antihypertensive or lipid-lowering therapy might suggest a selection bias in the opposite direction. We included relevant confounders in the analysis but misclassification and missing values might lead to residual bias. In a non-randomised study, residual confounding from unmeasured genetic, social or environmental variables is a concern. However, we fitted several different models and performed sensitivity analyses that provided evidence that the main findings of the study were robust. Nearest neighbour matching gave a comparison group that was not exactly matched for key variables. However, covariate adjustment had minimal effect and very similar results were obtained when the entire source cohort was used for reference. A method of analysis that did not allow for matching might give slightly wider CIs and larger p-values than a matched analysis,24 but the present results did not raise concerns of statistical error.

Summary of findings

In a population of obese patients with T2DM undergoing bariatric surgery there is evidence of improved blood glucose control, reduced requirement for antidiabetic medicines and an increased proportion of participants in diabetes remission for up to 6 years following the procedure. These benefits appear to be greatest among those with class III (morbid) obesity but procedure type was associated with BMI and this might have introduced confounding by indication. Participants receiving either GBP or SG procedures, rather than gastric banding, showed earlier onset and higher rates of diabetes remission. The SG and GBP groups differed at baseline, with low numbers beyond 3 years for SG, making it difficult to draw conclusions with respect to the comparative effectiveness of these procedures.

Bariatric surgery was associated with a fourfold increase in diabetes remission compared with no bariatric surgery. The proportion of patients in remission declined over time following surgery and there was biochemical evidence that individual patients showed relapse of diabetes after a period of remission. About one-third of all participants who received bariatric surgery were diabetic, indicating positive selection of diabetes patients for this intervention, consistent with current policy recommendations that bariatric surgery should be utilised in the management of obesity-associated comorbidity.

Comparison with other results

Changes in mean HbA_1c in our study were similar to those reported in previous studies,72 but rates of diabetes remission after surgery reported in the present study were generally lower than those noted in published reviews, which ranged from 59% to 92%. 6,32,70 However, case definitions have varied. We used a stringent definition of remission that required a normal HbA_1c value to be recorded in each year. Patient-years with missing values for HbA_1c in the present sample could have led to underestimation of remission rates. When remission rates were re-estimated, including only participants with HbA_1c values recorded in year, the proportion of participants in remission was higher. However, the pattern of missing HbA_1c values may be related to diabetes control, possibly with less frequent recording in participants who are either in remission or in good control. In clinical practice, some patients may be continued on metformin therapy even in the setting of apparent biochemical remission but such participants were not classified as being in remission for this study. Underestimation of remission may also occur if patients receive repeat prescriptions for medicines they no longer require. Conversely, relapse after previous remission may not be detected if blood glucose control is not evaluated. Even allowing for these limitations, this pragmatic study raises questions concerning whether or not the results achieved in routine practice may be as favourable as those reported from research studies. Control participants generally had poorly controlled diabetes, as did bariatric surgery patients before the procedure, raising questions concerning the quality of diabetes care received. We do not have information on whether the primary purpose of surgery was weight reduction or improved diabetic control. We caution against concluding that bariatric surgery is the only method for improving blood glucose control in obese patients.

Previous studies suggest that GBP is more effective for the treatment of diabetes than LAGB. 42,74,142 Fewer data have been reported for SG and data with long follow-up are not available. A randomised trial of 60 participants suggested that GBP might be more effective than SG. 143 A recent randomised trial found biliopancreatic diversion may have greater success than GBP, but was also associated with more nutritional side effects and lower quality of life. 144 Our results suggest that both SG and GBP may be more effective than LAGB and do not support the hypothesis that duodenal exclusion is important. We caution that in a non-randomised study procedures may be used selectively for patients with different baseline risks, and are more susceptible to bias than results from randomised trials. Higher remission rates with greater weight loss have also been reported. 74 Although we did not investigate weight loss as a predictor of diabetes remission due to limited follow-up, a higher baseline BMI category was positively associated with remission and greater baseline BMI is known to be associated with greater weight loss. 6 Our results offer some suggestion that the effect of bariatric surgery may decline over time. However, LAGB comprised a higher proportion of procedures at long follow-up. There is a clear need for longer-term follow-up studies to evaluate the long-term diabetes outcomes of bariatric surgery.

The safety of bariatric surgery also requires careful consideration. The report of the UK National Bariatric Surgery Register11 estimated that postoperative mortality in hospital following bariatric surgery was approximately 0.07%, with rates of complications, including anastomotic leakages, of approximately 2.9% following first procedures. In the longer term, band removal may be required following adjustable gastric banding, and there are risks of intestinal herniation and strangulation, which may have serious consequences. These risks require better quantification.

Strengths and limitations

This analysis was based on primary care EHRs of a large population-based sample of patients, enabling comparison of the benefits of bariatric surgery over standard medical care for the treatment of diabetes. However, we acknowledge several limitations. As the analysis was based on primary care data, there was limited information concerning surgical details including whether operations were open or laparoscopic or whether or not there were surgical complications, and assessment of the safety of bariatric surgery was beyond the scope of this research. The problem of missing and inconsistently recorded data has been referred to above. Weight records in primary care were insufficiently recorded to relate study outcomes to weight loss. Preoperative BMI might have been misclassified through infrequent recording.

Control participants were matched with bariatric surgery patients on BMI, age, sex and index year, but there remained differences between groups with respect to important confounders including length of diagnosis of diabetes, smoking history and hypertension. Analyses were adjusted for these confounders in a multiple regression model. An alternative approach might have employed propensity score matching, but results obtained using this technique are generally similar to those obtained with the regression adjustment approach adopted here. 145 Multiple controls per case could have been sampled but there was no evidence of lack of statistical power. We acknowledge that diabetes care was not standardised and quality of diabetes care might have differed between bariatric surgery patients and controls. The control patients may have had a different length of diagnosis or may have suffered different levels of complications from the surgery group. Given the highly selective access to bariatric surgery in the UK we would anticipate that surgery patients would have diabetes for longer and possibly experience more complications than controls. It is possible that surgery patients had better access to specialist diabetes care in the run-up to their surgery. In this pragmatic study, we used HbA_1c values to evaluate diabetes remission, but we acknowledge that greater sensitivity for diabetes might have been achieved through systematic use of glucose tolerance tests. 146

Cases and controls were matched but there were residual differences for some variables, including BMI. This may be expected as patients who receive surgery differ from those who do not receive surgery in many ways. Analyses were additionally adjusted for BMI, as well as other confounders. Nevertheless, residual confounding might bias the results. The median year of procedure was 2011 and only a minority of participants had operations more than 5 years before the study close in April 2014. We also acknowledge that the available duration of follow-up was generally shorter for more recently adopted procedures, including GBP and SG, than for LAGB. The analysis was on an ‘intention-to-treat’ basis and although small numbers of participants might have had conversion, revisional or reversal procedures, these were not excluded. A randomised trial remains the preferred design for evaluating the outcome of these procedures and results from a non-randomised comparison must be treated with caution, as noted above.

Summary of findings

Patients undergoing bariatric surgery have higher levels of depression than other obese patients with similar BMI and of the same age and sex. Frequent comorbidities, including diabetes mellitus, might be associated with this increased frequency of depression. The results of this analysis indicate that bariatric surgery in obese patients may be associated with a modest reduction in the prevalence of depression, and the use of antidepressant medications in primary care, but these effects do not appear to persist more than 3 years following the procedure. The reasons why patients with depressive illness are disproportionately represented among patients undergoing bariatric surgery in this population are unclear. In the UK, only a very small minority of patients with severe obesity undergo bariatric surgery and it is possible that psychological symptoms may be one of the considerations that influence whether or not an obese patient receives surgery. Obese patients seeking treatment may generally have higher levels of psychological distress. In a study in Germany, Herpertz et al. 154 found that patients undergoing bariatric surgery had similar levels of depression to those receiving non-surgical weight-loss therapy, with both being higher than obese controls. Comparison of patients who were either depressed or not depressed in the preoperative year showed a decline in depression in the former and an increase in the latter after surgery. This is compatible with regression to the mean and is consistent with the episodic nature of depression symptoms. Our results do not provide strong evidence that patients should be prioritised in the hope that bariatric surgery will provide long-term relief of clinical features of depression, even though short-term effects might be judged clinically relevant.

Comparison with other results

Previous studies have generally shown larger effects than the present study but these generally used samples drawn from specialist centres that might be susceptible to bias. One of the largest studies was conducted by the Longitudinal Assessment of Bariatric Surgery-2 (LABS-2) group. The included 2148 participants from 10 US hospitals with Beck Depression Inventory recorded at baseline and at least 1 follow-up within 3 years of the procedure. There were 40% who reported treatment for depression before surgery, while 28% had mild depressive symptoms and 5% moderate or severe depressive symptoms. The proportion with any depressive symptoms was < 10% in the first year after surgery but increased to 16% at 3 years. A similar pattern of change in depressive symptoms over time was reported by Burgmer et al. ,155 in a 4-year study of 148 participants. However, changes in use of antidepressant medication in the LABS-2 study were modest, with 35% using antidepressant medication before operation and 28% in the second and third years following the procedure. These results emphasise the importance of long follow-up in future studies as well as the evaluation of different measures of depression. In the SOS study, a battery of health-related quality of life measures, including the Hospital Anxiety and Depression Scale, showed improvement followed by deterioration. 92 An initial improvement is associated with the period of greatest weight loss in the first 2 years following surgery, sometimes viewed as a ‘honeymoon period’,156 followed by a subsequent deterioration associated with weight regain. 92 However, in the SOS study there was evidence of some improvement in depression symptoms up to 10 years of follow-up. Premorbid depression was less frequent in the SOS study than in the present sample.

Strengths and limitations

This analysis had the strengths of a large, population-based sample of patients undergoing bariatric surgery with prospectively recorded data for depression diagnoses and antidepressant prescribing. The interrupted time-series design is generally considered to be more resistant to bias than other non-randomised designs. 157 However, there were clear differences between the intervention and matched comparison group with respect to the outcome of interest, indicating that individuals undergoing surgery represent a highly selected group that is not typical of all patients with the same BMI, age and sex. Patients undergoing surgery may receive a package of supportive care and improved clinical management to prepare them for surgery, which may confound the effect of the surgical procedure. Patients in the control group received the usual care offered at their general practice, which was unlikely to include standardised management of obesity or depression.

The types of surgical procedure and patient case mix changed over time and, though access remained restricted, the numbers of procedures increased. There were small numbers of patients with long follow-up with reduced statistical power for evaluation of later time points. Furthermore, there is a risk of bias because patients operated on longer ago may have different characteristics from those operated on more recently, with shorter periods of follow-up. Use of clinical data may also introduce bias because patients must consult with their physician, and have their symptoms recognised, before a depression diagnosis may be recorded. Some evidence shows rates of diagnosed clinical depression are lower than those found in epidemiological studies. 158 An important limitation is that there was very limited recording of body weight during the period of follow-up and we cannot determine whether changes in depression were associated with weight loss or regain.

Our assessment of depression was based on clinical diagnoses and antidepressant prescribing; the limited changes observed over time might result from difficulty in stopping antidepressant therapy once initiated. A previous CPRD study found that depression is often treated with short-term courses of antidepressant medications, with only a small proportion of patients being prescribed drugs over a long period for chronic depression. 157 The high rates of antidepressant use observed in this study may represent met need and not merely the result of repeat prescribing to patients who might no longer require clinical treatment for depression. Several previous studies used self-report measures to evaluate depressive symptoms and these measures might be associated with greater sensitivity for depression and responsiveness to change. 159,160 For all of the reasons, it would be desirable to test hypotheses using well-designed randomised trials with prospective documentation of depression.

Disease incidence

Summary of findings

This study drew on recently published evidence for the effects of bariatric surgery on incidence of disease,7,8 mortality10 and health-care costs12,13 following the procedure. This research modelled the lifetime health benefits from bariatric surgery. The results project substantial increases in life-years and reductions in years lived with diabetes. Bariatric surgery is expected to be associated with increased health-care costs arising from the costs of the procedure, as well as increased lifetime health-care costs associated with increased life expectancy. When health benefits and costs are combined into a single metric, using accepted values of cost per QALY, use of bariatric surgery is expected to yield substantial net monetary benefits amounting, over a lifetime, to £49M per 1000 persons. Net health benefits may amount to 1634 QALYs per 1000 persons. Bariatric surgery has similar cost-effectiveness in men and women, at different ages and in different deprivation categories. Bariatric surgery is also expected to be cost-effective in individuals with severe obesity (BMI 35–39 kg/m²). Bariatric surgery will be cost-effective even if the cost of the procedure is twice as high as we have estimated, or if the effect of the procedure declines over time so that only 32% of the initial effect remains after 10 years.

Comparison with other results

There have been several previous cost-effectiveness analyses of bariatric surgery for morbid obesity. One of the most authoritative was the 2009 study by Picot et al. 31 Their study reported a model in which weight loss following surgery was viewed as the main mediator of longer-term changes in health outcomes. The study found that bariatric surgery was a cost-effective intervention for morbid obesity, with ICERs ranging between £2000 and £4000 per QALY gained over a 20-year time horizon. 31 A more recent study suggested that bariatric surgery may be cost saving for health-care systems through reduced morbidity,175 though this conclusion has been disputed. 178 Although bariatric surgery does not appear to generate cost savings, its use is associated with substantial health gains at costs that are well below accepted thresholds for cost-effectiveness.

This analysis updates previous studies by including those surgical procedures that are currently utilised including adjustable gastric banding, GBP and SG. The analysis also updated the costs of surgery to present-day values. The present model incorporated direct evidence concerning the long-term outcomes of bariatric surgery in contrast to previous studies, which have used changes in intermediate measures of surgical effect including body weight, blood pressure and lipid profiles to model the long-term outcomes of bariatric surgery. Our analysis recognised that the effects of surgery, particularly those on diabetes, are not entirely weight dependent and our model therefore did not include weight change as an intermediate outcome. The model was informed by recent measures of the effects of surgery on substantive long-term health outcomes and health-care costs, additionally informed by analysis of CPRD EHRs. In spite of this difference of approach, our analyses are consistent with previous reports in showing that bariatric surgery is likely to be very cost-effective. We do not find that bariatric surgery is cost saving. This would not be expected, however, as it is unlikely for a procedure that reduces mortality in a population that experiences a heavy burden of morbidity to reduce lifetime health expenditures. 177

Strengths and limitations

This research was based on empirical data for disease incidence, mortality and costs of health-care utilisation estimated from the EHRs of a large sample of participants managed in primary care in the UK between 2008 and 2013. We caution that the small proportion of very obese patients who had bariatric surgery means there is a real possibility that they differ from other patients whose characteristics on health records are similar. The study also drew on recently published and authoritative estimates of the effects and costs of bariatric surgery with an emphasis on those bariatric surgical procedures that are currently utilised. Both the empirical and abstracted inputs were consistent with estimates from systematic reviews of the effects of surgery on related morbidities. We have used conservative assumptions including that costs of health-care utilisation following surgery are not associated with weight loss; that any gain in utility associated with BMI reduction declines rapidly over time; and that remission from diabetes following surgery occurs in 40% of patients, as estimated from CPRD, which is lower than levels observed in other reported studies. We acknowledge that any model represents a simplification of reality. There are other forms of morbidity that were not represented in the model. However, the costs of health-care utilisation from such conditions will have been included in cost estimates from CPRD, which encompassed all health-care utilisation. We included the major complications of surgery including operative mortality, costs of leaks following surgery, and reoperations in 2% of patients per year. There may be additional costs associated with surgery but a sensitivity analysis showed that even if total costs of surgery were to be twice as high as estimated in the base case, bariatric surgery would still be cost-effective. We modelled bariatric surgery as having a constant effect in the postoperative period, but we showed that even if the effects declined over time surgery would still be cost-effective.

Media interaction

The findings of this research have been reported widely within the media. The research team gave radio interviews to BBC London Live and the Today programme on BBC Radio 4. A summary of newspaper and online articles is presented in Box 2.

Patient engagement

Members of the research team attended a meeting of post-bariatric surgery patients at St George’s Hospital in London. The team presented information relating to the study and invited the 16 patients present to feedback comments and ideas relating to the project and future research. The study was well received by the group and elicited a good deal of feedback. Areas that the patients commented on included the need for better support both prior and subsequent to the procedure, particularly in relation to dealing with the psychosocial issues that encouraged their overeating. A key point was the need to improve support for patients in addition to increasing access to surgery. The demotivating effect of trying to reach a ‘normal’ BMI rather than focusing on weight lost was also discussed. Some patients reported their anxiety about weight regain and unpleasant side effects from surgery, such as loose skin and hair loss. Favourable comments were also made about the weight loss and diabetes remission. We plan to disseminate findings of the study to patients by producing a short summary leaflet and attending a further patient group meeting.

Presentation of results to professionals

Members of the research team have presented aspects of the study findings at the following conferences and meetings:

• International Federation for the Surgery of Obesity and Metabolic Disorders, Montreal, Canada, 26–30 August 2014

• Social Society for Medicine, Oxford, UK, 10–12 September 2014

• Obesity Week, California, USA, 2–6 November 2015.

Overview of patient and public involvement

It was very informative for the team to attend the patient meeting and hear their thoughts on the research. We took on board their suggestions for future research, and although we were not able to include them in the current analyses we have included them in the research recommendations and have referred to them in the interpretation of our results where possible. We were delighted that the publications produced during the study received attention in the mainstream media. As a result of this coverage we were contacted by health professionals who work with bariatric surgery patients, who said that they would be communicating the findings to their patients.

Methodological

This research demonstrates the value of EHRs for long-term follow-up of participants who have received defined health care interventions. Future research might aim to recruit patients into randomised studies with long-term follow-up of substantive clinical outcomes.

Basic research

Improved understanding of the mechanisms of effect of bariatric surgical procedures may yield insights that can be used to inform more effective interventions for people with morbid obesity.

Health technology assessment

• Research is needed, using efficient study designs, to evaluate the long-term effects of currently used bariatric surgical procedures, across a comprehensive range of outcome measures.

• Estimates of intervention effects from bariatric surgery should be stratified by age group and morbidity category (especially diabetes status). This might enable evaluation of optimal age and morbidity distributions to be prioritised to receive bariatric surgery and inform patient selection.

• The value of CPRD for deriving current clinical data on bariatric surgery use and outcomes has been established during this project. Continued observation of this cohort over time may provide further insights into the evolving use of procedures and related outcomes.

• Research to define interventions to improve mental health outcomes of people with morbid obesity.

Service organisation and delivery research

• Research to map and quantify disparities in access to bariatric surgery in relation to need.

• Multi-method research on organisational and service barriers to accessing bariatric surgery, incorporating views and experiences of policy-makers, commissioners, managers, clinicians and patients.

• Research on patient experience including pre- and post-operative patient support, weight regain after surgery, weight loss targets and how this information is presented to patients.

• Evaluation of strategies for scaling up bariatric surgery for the treatment of obesity, while maintaining a safe service.

• Research to define strategies for organising and delivering care to optimise mental and physical health outcomes of people with morbid obesity. This includes primary disease prevention, in the context of NHS Health Checks, and secondary prevention, in the context of primary care management of multi-morbidity.