Five-year direct medical costs and comorbidities after bariatric surgery of obese patients with type 2 diabetes mellitus: A population-based propensity score-matched study CURRENT

Background: Bariatric surgery is effective in weight reduction and diabetes remission. This study aimed to estimate direct medical costs and changes of comorbidities after surgery up to five years among obese patients with type 2 diabetes mellitus. Methods: A population-based retrospective cohort of obese type 2 diabetes patients from Hong Kong Hospital Authority between 2006 and 2017 was assembled. One-to-five propensity score matching method was applied to match 401 eligible surgical patients with 1,894 non-surgical patients. Frequency of healthcare service utilization and dispense of diabetes medication were collected for both groups to estimate the direct medical costs from baseline to up to 60 months; Charlson Comorbidity Index (CCI) and number of comorbidities were measured to compare the changes of comorbidities between two groups over the 5 years. Results: Direct medical costs were US$40,889 for surgical patients and US$6,163 for controls in the index year (p<0.001), with incremental costs of US$34,726. Bariatric surgery and hospitalization were the main cost drivers for surgical patients in the year of surgery. Although surgical patients had significantly lower annual costs than control patients in the subsequent four years, five-year cumulative costs incurred by surgical patients were significantly greater than controls (US$60,174 vs US$33,374, p<0.001), regardless of subgroups. Surgical patients had better profile of comorbidities than controls, as they had significantly lower CCI after baseline and fewer percentages of them proceeded to higher CCI categories. Conclusions: Over 5 years, bariatric surgery was associated with increased medical costs in the year of surgery and cumulative costs. Although bariatric surgery is not cost-saving for type 2 diabetes patients at 5 years, it is associated with improved comorbidity profile. health problems, hyperlipidemia, obstructive sleep apnea, gallbladder disease, musculoskeletal chronic orthopedic

the main cost drivers for surgical patients in the year of surgery. Although surgical patients had significantly lower annual costs than control patients in the subsequent four years, five-year cumulative costs incurred by surgical patients were significantly greater than controls (US$60,174 vs US$33,374, p<0.001), regardless of subgroups. Surgical patients had better profile of comorbidities than controls, as they had significantly lower CCI after baseline and fewer percentages of them proceeded to higher CCI categories.
Conclusions: Over 5 years, bariatric surgery was associated with increased medical costs in the year of surgery and cumulative costs. Although bariatric surgery is not cost-saving for type 2 diabetes patients at 5 years, it is associated with improved comorbidity profile.

Background
As populations aging and lifestyles change, the prevalence of obesity and obesity-related diseases, such as type 2 diabetes mellitus, continues to increase worldwide [1]. The prevalence of obesity since 1980 has doubled in over 70 countries [2] and, the global number of patients with type 2 diabetes mellitus is projected to rise sharply from 451 million in 2017 to 693 million in 2045 [3]. The economic burden of both obesity and obesity-related metabolic disorders is consequently exerting great pressure on health care systems everywhere [4,5].
While lifestyle interventions are considered as the first-line treatment for obese patients with type 2 diabetes mellitus [6], these modalities are often limited by unsustainable and modest effects on weight reduction and disease remission [7]. As a consequence, bariatric surgery has been widely accepted as the most effective therapy for the treatment of morbid obesity [8] and as a useful option to resolve type 2 diabetes mellitus [9][10][11] and reduce the risks of cardiovascular diseases (CVD) and mortality. Guidelines for the management of obesity and/or type 2 diabetes mellitus recommended that bariatric surgery should be considered or recommended for obese patients with certain health conditions [12][13][14]. In recent years, there has been a rapid growth in bariatric surgery globally [15].
This increase, however, has inevitably accelerated the need for long-term support and nutritional follow-up, bariatric surgery-related hospitalization and use of other healthcare services [16].
Reported unit cost of bariatric procedures ranged between US$7,423 and US$33,541, with a mean of US$14,389 [17]. However, the mean cost of lifestyle intervention achieving one kilogram of weight loss varied from only US$54 over a 2-month period to US$1,005 over twelve months [18]. Despite the expensive unit cost of bariatric procedure, costs associated with bariatric surgery mainly accrued at the year of the surgery and reduced during the post-surgery period. Patients with bariatric surgery may have reduced costs of prescription, outpatient visits and/or hospitalization in the long-term, since obesity-related complications and loss of productivity averted after surgery [19][20][21][22]. It was found that most economic evaluations of bariatric surgery were conducted in the US [20][21][22], however, it should be carefully performed anywhere else where the healthcare system differs inherently from the US and the burden of obesity and obesity-related diseases is substantial.
In this regard, this study aimed to 1) estimate five-year cumulative and annual direct medical costs of obese type 2 diabetes patients with and without undergoing bariatric surgery in overall and by subgroups; 2) examine the patterns of healthcare service use and dispense of diabetes medications among surgical and non-surgical patients over a five-year period; and 3) compare the profiles of comorbidities between bariatric surgery patients and control patients from baseline to up to 5 years.

Study Design and Study Population
The study design and population were described previously [23]. In brief, a population-based retrospective cohort study of obese patients with type 2 diabetes who had ever utilized public  Table 1). Surgical patients were excluded if they underwent non-bariatric procedural operations, did not have pre-existing type 2 diabetes, had no body mass index (BMI) readings or had BMI reading < 27.5 kg/m 2 [24] at the index date (i.e. the date of bariatric surgery). A one-to-five propensity score matching method was used to match each eligible surgical patient with up to 5 non-surgical patients. The index date of non-surgical patients was same as that of their matched surgical patients. Ethics approval of this study was granted by Institutional

Outcomes
Primary outcomes were the five-year cumulative and annual direct medical costs of obese type 2 diabetes patients with and without bariatric surgery in overall and by subgroups, as well as changes of comorbidity profile of both groups over five years. Secondary outcomes were the frequency and costs of use of healthcare services and diabetes medications.

Estimation of annual direct medical costs
The dates and frequency of the use of HA healthcare services by patients in the surgery and matched control group in the index year and subsequent four years were retrieved. Healthcare services included general outpatient clinics (GOPC), specialist outpatient clinics (SOPCs), accident and emergency (A&E) and allied health professionals (including clinical psychologists, dietitians, occupational therapists, physiotherapists, and smoking counselling and cessation visits), as were as length of hospitalization in a general ward, intensive care unit (ICU), cardiac care unit (CCU) and high dependency unit (HDU). Unit costs of healthcare services were extracted from the public charges to non-eligible persons listed in the 2017 Government Gazette and HA Ordinance (Chap. 113) [25]. Types of anti-diabetic drugs prescribed and days supply of each dispensed anti-diabetic drugs at each clinic visit during the follow-up period were also collected. Included diabetes medications were insulin, metformin, thiazolidinedione (TZD), dipeptidyl peptidase 4 inhibitors (DPP-4i), sodium glucose cotransporter 2 inhibitors (SGLT2i), and glucagon-like peptide-1 (GLP-1) receptor agonists. Costs of diabetes medications per 30-day supply were based on a published paper [27], and the midpoints of the cost range of each class of diabetes medication were used to estimate the medication costs.
The pegged exchange rage of US$1 = HK$7.80 was used to convert Hong Kong dollars to the US dollar. Supplemental table 2 summarizes the unit cost of healthcare services and diabetes medications.

Supplemental table 2
The annual direct medical costs for each patient were calculated based on the formula: Annual direct medical costs = (∑ frequency of the healthcare service use in that year ⋅ unit cost of the respective healthcare service) + (∑ days supply of diabetes medication ⋅ midpoint cost of the cost range of that medication ÷ 30)

Changes of comorbidities
Numbers of patients with different comorbidities (including hypertension, mental health problems, hyperlipidemia, obstructive sleep apnea, gallbladder disease, musculoskeletal and chronic orthopedic disorders, CVD, severe hypoglycemia, chronic lung diseases, chronic renal disease [CRD]) and percentage of patients in different CCI categories (1, 2, 3, 4, 5 or above, dead) were measured from baseline to up to 60 months. Also, changes of mean CCIs of two groups from baseline to 5 years were visualized in line graphs, and trends of mean CCI changes were analyzed for both groups through tests for trends.

Statistical analysis Baseline covariates
The baseline covariates of patients included gender, age, BMI, and pre-existing of complications (including hypertension, mental health problems, hyperlipidemia, obstructive sleep apnea, gallbladder disease, musculoskeletal and chronic orthopedic disorders, CVD, and CRD).

Propensity Score Matching Method
Multiple imputation by chained equations (MICE) [28] was used to address the absence of baseline data. HbA 1c , blood pressure and low-density lipoprotein cholesterol were imputed by gender, BMI, history of comorbidities (i.e. hypertension, mental health problems, hyperlipidemia, obstructive sleep apnea, gallbladder disease, musculoskeletal and chronic orthopedic disorders and CVD), duration of type 2 diabetes, CCI, and use of insulin and oral anti-diabetic drugs. Model parameters were estimated from multiple imputed data and then were used to obtain multiple-imputation linear predictions by applying Rubin's combination rules observation wise to the completed-data predictions [29], where twenty imputed datasets were created for multiple imputations. Obtained predictions were then used in propensity score matching.
Each patient's propensity score was computed by multivariable logistic regression adjusting for baseline covariates of patient. The caliper criteria improved the quality of the nearest neighbor matching by specifying a maximum tolerance of the propensity score distance between patients in the surgical group and in the control group. The propensity score matching was performed by 'calipmatch' command on a one-to-five basis without replacement in STATA.
Baseline socio-demographic, clinical parameters and comorbidities of patients in both groups were presented by frequency with percentages for categorical variables and means with SD for continuous variables. Independent t-tests or chi-square tests were used to assess the difference in the baseline characteristics between the two groups. Balance of baseline covariates between two groups after matching was measured by standardized mean difference (SMD). All SMDs less than 0.2 implied optimal balance between the two groups [30].

Generalized linear models
Direct estimation of annual cost using normal Gaussian distribution may be problematic due to the skewness of the data and zero-cost issues [31]. Therefore, generalized linear models (GLMs) with gamma family and log-link were used to accurately estimate the five-year cumulative and annual direct medical costs. GLMs, especially the gamma regression model, perform well on the estimation of non-negative and positively skewed cost data [32], and the model has been widely used in analyses of medical costs [33,34]. The GLM goodness of fit was tested by using the Modified Park Test [35].
According to Modified Park Test [35], a model coefficient approximating to 2 indicated that the use of gamma distribution in GLMs was preferred. All statistical analyses were performed by STATA Version 13.1 (StataCorp LP, College Station, Texas).
All significance tests were two-tailed and a p-value of < 0.05 was considered as statistically significant.

Sample characteristics
A total of 422 obese type 2 diabetes patients met the criteria to take part in the surgery group. After propensity score matching, a cohort of 2,295 patients (401 surgical patients and 1,894 non-matched controls) were included in current analysis (Supplemental Fig. 1).
The baseline characteristics between two groups were well balanced (Supplemental Fig. 2 and Supplemental Table 3). The mean BMI of surgery and control patients were 36.81 kg/m 2 and 36.31 kg/m 2 , respectively (p = 0.121, SMD = 0.090). Of 401 surgical patients, 299 (74.6%) underwent laparoscopic sleeve gastrectomy, 74 (18.5%) underwent laparoscopic gastric bypass, 11 (2.7%) underwent laparoscopic adjustable gastric banding, and 17 (4.2%) underwent laparoscopic sleeve gastrectomy with duodenojejunal bypass Frequency and costs of use of healthcare services The mean number of outpatient visits for both groups was between 6 and 10 visits, and the control group had significantly fewer outpatient visits than surgical patients across five years. The mean number of A&E visits for both groups was smaller than one; and the differences were insignificant  Table 4) Supplemental Table 4 Estimated annual and cumulative direct medical cost   Table 1 lists the number and percentage of patients with different comorbidities over the years. Over the 60 months, less percentage of surgical patients newly developed hypertension, hyperlipidemia, musculoskeletal and chronic orthopedic disorders, CVD, severe hypoglycemia, chronic lung disease and CRD. Figure 3 portrays the mean CCI of both groups from baseline to 60 months. Mean CCI of both groups were similar at the baseline, however, surgical patients had significantly lower mean CCI than control patients in the subsequent years. Although the mean CCI of surgical patients increased from 3.40 to 3.82, the p for trend was 0.443, indicating the change of CCI was stable for surgical patients. In contrast, the p-trend of control patients was smaller than 0.001.  Figure 4 shows that percentages of patients in low CCI categories decreased, as patients proceed towards high CCI categories or death in each year. It was found that more percentages of control patients proceed towards CCI category of ≥ 5 or death than surgical patients. Although more percentages of surgical patients died at 12 months, the growth rate of mortality for the surgery group was slower than that of the control group in the remaining years. Also, it was found that over half of surgical patients died from cancers and none of them died due to bariatric surgery at 12 months.

Changes of comorbidity profiles
Supplemental Table 5 summarizes the causes of death for both groups over the years.

Discussion
To our knowledge, the present study is the first population-based retrospective cohort study that estimated the five-year direct medical costs and comorbidities after bariatric surgery in Hong Kong.
Above all, findings of this study provided additional information on direct medical costs in an Asian context, and provided basis for further cost-effectiveness analysis of bariatric surgery among obese Chinese patients with type 2 diabetes. Indeed, most existing costing analyses on bariatric surgery were conducted in the US [20][21][22], and those results may be inappropriate to be generalized to other jurisdictions, where mandatory insurance cover is not available and/or population ethnicities are totally different. Therefore, information on annual direct medical costs, healthcare utilization and medication use of bariatric patients provided by this study was especially important for policy makers and bariatric surgeons in Asian countries to further optimize the health resources allocation. Besides, comorbidity profiles of both groups were tracked for 60 months, providing an overview of changes of comorbidities. Also, as most previous comparative cost analyses on bariatric surgery did not include exploratory analyses [20,22,36,37], this study also added current knowledge on medical costs of bariatric patients in subgroups, which allowed easy estimation of annual direct medical costs of surgical patients with different health conditions.
One major finding of the present study was that surgical patients had significantly higher medical costs in the year of surgery but had lower annual costs in the subsequent four years. This was mainly because surgical patients had great expenditures on bariatric procedures and bariatric surgeryrelated hospitalization, while control patients did not have such cost considerations. However, in the remaining years, shorter length of hospital stays were found in the surgery group in year 2 onwards.
Indeed, coefficients in Supplemental Table 6 indicated that bariatric surgery was associated with increased medical costs in year 1 (p < 0.001). However, the values of coefficients were below zero in the following years, suggesting that bariatric surgery was associated with reduced medical costs.
Despite the surgical patients having need to use outpatient and allied health professional services more frequently than non-surgical patients over the years, the unit cost of hospitalization was much higher than that of outpatient visits and allied health professional visits. Consequently, the costs saved from outpatient and allied health services by the matched control group were considerably less than the extra costs of hospitalization, leading to the higher medical spending for the control group in the subsequent years. However, the finding that bariatric surgery was associated with reduced costs of inpatient services was inconsistent with the results of previous comparative studies [21,22,37], where higher hospitalization costs were reported in post-surgery periods. One possible explanation to this discrepancy is that re-operation rates of in patients with bariatric surgery was relatively low in Hong Kong. Indeed, the overall re-operation rate in the present surgical group was 4.74% in the index year and nearly zero in year 3 onwards (Supplemental Table 7). In comparison, reported re-operation rates in a systematic review is around 7% [38]. Besides, decreased frequencies of prescription and saved diabetes medication in the bariatric surgery group were observed in the first 4 years after bariatric surgery. Similar findings were also reported in previous studies, which suggested that the mean number of diabetes medication per patient and number of patients with prescriptions dropped after bariatric surgery over time [20,39,40].
Though bariatric surgery had cost-saving effects in year 2 onwards, the five-year cumulative medical costs of surgical patients were higher than those of matched control patients, as saved costs failed to offset the increased costs due to bariatric surgery. However, the slope of the curve representing the cumulative costs of controls was steeper than that of surgical patients (Fig. 2 Of note, the relative differences of five-year cumulative costs between surgical and non-surgical patients were smaller in subgroups of patients who were female, had baseline age over 60 years old, had BMI over 35 kg/m 2 , had diagnosis of type 2 diabetes over 5 years, had CCI greater than 4, had history of CVD, and had history of CRD. These results implied shorter time to break-even may occur in these subgroups, and therefore obese type 2 diabetes patients in these subgroups were more suggested to undergo bariatric surgery. Notably, patients with CRD in the surgery group had similar 5-year cumulative medical costs with those in the control group. One possible explanation was that bariatric surgery has renal protective effects, even in patients with established renal diseases [38,41,42]. Improved renal function after bariatric surgery largely reduced the annual medical costs of surgical patients in post-surgery years, and facilitated cumulative costs of CRD patients with and without bariatric surgery approximately converge at year 5. Another possible reason was that treating CRD patients, especially those with end-stage renal disease, is costly [43]. Therefore, costs of bariatric operation for those patients accounted for 11.6% of the 5-year total medical costs when compared to 36.3% of total costs in overall.
Our study also supported that bariatric surgery delayed the occurrence of most comorbidities and improved profiles of comorbidity, since fewer percentage of surgical patients proceed towards high CCI categories and surgical patients had lower mean CCI across 5 years. Particularly, bariatric surgery had protective effects on hypertension, hyperlipidemia, musculoskeletal and chronic orthopedic disorders, CVD, severe hypoglycemia, chronic lung disease and CRD. Indeed, previous studies have already reported that bariatric surgery was beneficial towards resolving or delaying above comorbidities [9,38,42].

Limitations
Several potential limitations in the present study should be acknowledged. Firstly, we included patients who had used HA healthcare services from 2006 to 2017. However, patients whose index year was 2014 or later have not yet completed their five-year follow-up. Consequently, the smaller sample size affected the estimation of annual medical costs in later years, and some findings should be interpreted with caution. Secondly, unlike a few previous studies [36], the present one did not differentiate the type of bariatric surgical procedures, but presented results in overall. Lastly, this study was based on the data obtained from HA, which is the public health service provider in Hong Kong. Hence, the annual direct medical costs of obese patients with type 2 diabetes attending private health services were not included.

Conclusions
To conclude, bariatric surgery was associated with significantly higher medical costs in the year of surgery and five-year cumulative costs, but was associated with reduced effects on medical costs in

Consent for publication
Not applicable