Introduction
Bariatric surgery, while not the first-line therapy, was proved to be the most effective means of achieving sustained weight loss in morbidly obese patients in comparison to non-operative treatment [
1]. Although several surgical techniques are currently available, two most commonly performed are sleeve gastrectomy (SG) and Roux-en-Y gastric bypass (RYGB) [
2]. Selection of the proper bariatric procedure is subjective and has always been an area of active debate.
Despite comprehensive preoperative assessment of each candidate, weight loss outcomes after intervention show distinct deviation ranging from 37.6 to 94.4% of excessive weight loss (EWL) [
1]. More importantly, 7 up to 25% of bariatric patients fail to accomplish optimal result, defined as above 50% of EWL [
1].
Understanding of possible weight loss results of bariatric treatment would facilitate preoperative patient’s assessment and decision-making process [
3]. It would optimize selection of candidates more likely to benefit from the surgery and provide them with the appropriate procedure, resulting in better long-term effects on sustained weight loss and obesity-related comorbidities [
1,
4,
5]. Therefore, reasonable estimation of expected outcomes after bariatric surgery, especially weight loss, seems to be crucial not only for surgical candidates but also their physicians.
Numerous studies revealed that weight loss after bariatric treatment depends on various factors including demographic aspects, comorbidity rate, psychological profile, lifestyle, or socioeconomic status [
6‐
11]. However, it is difficult to take all of them into account during preoperative assessment of the patient in the proportion each one contributes to the outcome. It would, therefore, be helpful to have all significant predictors of weight loss integrated into easy-to-use estimation tool.
With constant development of bariatric surgery, there has been increasing focus on inventing adequate tools for outcomes prediction [
12]. As a result, more and more models and scoring systems for the assessment of postoperative weight loss or comorbidities alleviation are being proposed [
12,
13]. Currently the number of published weight loss prediction models becomes overwhelming. Moreover, most of them require external validation and precise statistical tools assessment. Still, there is lack of comprehensive scientific conclusion on the effectiveness of existing tools in predicting weight loss after bariatric surgery and their utility in clinical practice. Thus, we designed a study to perform a systematic review of the literature for the identification of available models and validate them as the predictors of weight loss at 1 year after SG or RYGB as well as compare their accuracy.
Materials and Methods
Study Design
In this retrospective cohort study, we performed systematic review of the literature to identify predictive models for weight loss after bariatric surgery. The predicted postoperative BMI was calculated for each patient according to original equations based on data obtained from medical records. Then, the relationship between predicted and observed BMI was assessed.
Study Population
We included consecutive patients admitted to our department between April 2009 and October 2017 who underwent either SG or RYGB and completed 1 year of postoperative follow-up. Patients with initially incomplete data, body mass index (BMI) under 30 kg/m2, and revisional surgery were excluded from the analysis.
We divided study population into 3 groups: the ALL group including patients after SG and RYGB, the RYGB group including patients after RYGB, and the SG group including patients after SG.
Candidates for bariatric surgery were evaluated by a multidisciplinary team of surgeons, dieticians, psychologists, clinical nurse specialists, and anesthetists. Demographic, anthropometric, and clinical data were recorded pre- and postoperatively. The follow-up schedule comprised appointment at 1 year after surgery.
Surgical Techniques
All participants underwent either laparoscopic SG or laparoscopic RYGB. Each patient was qualified for the appropriate type of procedure in accordance with the Polish Guidelines for Metabolic and Bariatric Surgery [
14]. The surgical techniques used in our department have been described in detail in our previous publications [
15,
16]. During laparoscopic SG, a 34-French gastric bougie was used to calibrate the gastric sleeve. Gastrectomy started 4–5 cm proximal to the pylorus with continuously applied linear staplers straight to the angle of His. The length of alimentary and enzymatic limb during RYGB was standardized in all patients, 150 and 100 cm, respectively.
Data Collection
Sex, age, height, weight, BMI, comorbidities, preoperative weight loss, time of the procedure, and length of hospital stay (LOS) were collected retrospectively from medical histories. Age was calculated as the difference between the date of birth and the date of surgery. BMI was calculated from the weight (in kilograms) and divided by the square of height (in meters). Investigated comorbidities included diagnosis of hypertension (HTN), heart disease (defined as coronary artery disease or past myocardial infarction), type 2 diabetes mellitus (T2DM), metabolic syndrome, hyperlipidemia, kidney disease, liver disease, obstructive sleep apnea (OSA), polycystic ovary syndrome (PCOS), gastroesophageal reflux disease (GERD), and arthritis.
Weight Loss After Bariatric Surgery
Evaluated outcome of bariatric treatment was defined as patient’s weight at 1 year after initial procedure, assessed by postoperative BMI. Weight change was expressed using percentage weight loss (WL), percentage EWL, and percentage excessive body mass index loss (EBMIL) obtained according to the previously described formulas [
17]. Ideal body weight was calculated as equivalent to BMI 25 kg/m
2. Adequate weight loss after intervention was defined as above 50% EWL [
18].
Model Selection
Literature search was performed in accordance with Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) recommendations [
19]. Search of PubMed, Embase, and Cochrane Library databases were performed on September 18, 2020. The following search terms were used: bariatric surgery, postoperative weight loss, weight loss prediction, and prediction model. Additionally, the articles’ reference lists were searched manually for further studies. At first titles and abstracts of each identified study were evaluated, and then full texts for potentially relevant articles were assessed. We included English written studies, investigating all types of bariatric surgery, with prospective and retrospective design which attempted to create an individualized prediction model for postoperative weight loss. Papers presenting models based on non-individualized or postoperative factors as well as variables not routinely checked in our daily practice were excluded.
Statistical Analysis
Continuous variables are presented as mean and standard deviation (SD) or median and interquartile range (IQR) for normally and non-normally distributed variables, respectively. Categorical variables are presented as numbers and percentages. To confirm the normality of the distribution of the continuous variables, we used the Shapiro-Wilk and the Kolmogorov-Smirnov with the Lilliefors correction tests. Equality of variances was assessed using the Brown-Forsythe test. Comparison between RYGB and SG groups was established using an independent t test or Mann–Whitney U test as appropriate for continuous variables and Chi-square test with Fisher’s correction for categorical variables. Comparison of clinical data before and after surgery was performed with the use of a paired sample t test or Wilcoxon test adequately.
The predicted postoperative BMI was calculated for each patient according to the original equations. If model was designed to predict weight change measures other than BMI, it was obtained with mathematical conversions. The relationship between predicted and observed BMI was assessed with linear regression method. Correlation parameters included regression coefficient (B) with 95% confidence interval (95% Cl). To evaluate the diagnostic accuracy of each model, adjusted squared Pearson’s correlation coefficient (R2) was used. Calibration was assessed by standard error of the estimate (SE) and root mean square error (RMSE) as well as paired sample t test between mean predicted and mean observed BMI. Good calibration was indicated by p>0.05. Additionally, the difference between mean predicted and mean observed BMI was obtained.
For all inferential statistics, statistical significance was defined as p≤0.05. All calculations were done with STATISTICA 13.0 software (StatSoft Inc., Tulsa, OK, USA).
Discussion
Our study identified 12 models for the prediction of weight loss after bariatric surgery. Validation on independent cohort of 760 patients revealed significant correlation between predicted and observed BMI in all models. According to the accuracy, examined tools were able to explain from 17 up to 25% of the variation of weight loss outcome at 1 year. The majority of models significantly overestimated effect of bariatric treatment with the exception of Cottam_2. On average the predicted BMI was 4.70 to 5.49 kg/m2 lower than the actual.
Obtained findings confirmed bariatric surgery to be effective method of obesity treatment, reemphasized in numerous papers [
1,
28,
29]. Implemented procedures resulted in significant postoperative weight loss followed by BMI reduction. More importantly, majority of patients achieved adequate effect of treatment with median EWL reaching 62.56%. This stays consistent with previously published studies reporting from 56 to 68% EWL depending on surgical procedure [
30,
31].
The current scientific reports point that there is no homogenous weight loss curve after bariatric procedure for all patients [
32]. Nevertheless, in all identified weight loss curves, distinct trajectories of initial weight change are followed by varied patterns of weight fluctuation over the longer term follow-up [
32]. These findings suggest that the longer term outcomes in weight may be determined by the magnitude and direction or slope of the initial weight reduction. Therefore, preoperative estimation of 1-year outcome based on prediction models might enable to optimize early weight loss trajectory directions, consequently, providing better long-term outcome of bariatric treatment.
While there is an increasing awareness of the importance of prediction models and they are being published in increasing amount, there were hardly any attempts to provide their external validation and comparison of predictive performance. To our knowledge, only Sharples et al. performed such evaluation in 2017 including four models for weight loss prediction [
14]. Since our study explored as many as 12 equations, it could provide more comprehensive, reliable, and up-to-date assessment of currently available weight loss prediction models.
In our research Baltasar_1 was able to explain 21% of postoperative weight, while previous validating studies presented better accuracy with the ability to predict outcome in 59% [
33]. Although the prediction properties of Baltasar_3 and Baltasar_4 models were not reported in the original research, they were externally validated in independent paper, which revealed
R2 equal 0.15 for Baltasar_3 and 0.34 for Baltasar_4 [
34]. Present study finds comparable performance of these two models with similar
R2 ranging from 0.19 to 0.21 depending on surgical procedure. Interestingly, by implementing both models into one cohort comprising patients after either RYGB or SG, Sharples et al. managed to obtain much higher accuracy with
R2 value of 0.61 [
33]. The same predictive performance Sharples et al. reported for Wood model [
33]. Nevertheless, it differs greatly in comparison to our results (
R2 0.61 vs 0.22). Wise et al. was the first author who provided measures of fit along with model development [
23]. According to the original study Wise was able to explain 35% of the variability of weight loss with average error of 17.4% EBMIL [
23]. Present analysis demonstrated the worst accuracy of Wise model in all examined groups with the ability to predict only 17 to 20% of the outcome and one of the highest differences in estimation (RMSE from 4.84 to 5.39kg/m
2). Goulard was the first model developed and validated on SG cohort [
24]. Authors established good predictive accuracy of the model with
R2 ranging from 0.61 to 0.76 as well as decent calibration with SE between 3.01 and 3.38 [
24]. Nonetheless, our study revealed distinct outcomes with regard to the precision of Goulard’s estimations (
R2=0.22 and SE=5.09kg/m
2). Seyssel model in our analysis presented one of the highest
R2 in all investigated cohorts ranging from 0.23 to 0.25 along with the lowest RMSE within 4.70 and 5.19 BMI points. Our results stay inconsistent with predictive properties provided in primary study. Not only training but also two independent validation cohorts suggested better accuracy with
R2 equal to 0.45, 0.66, and 0.69, respectively, and RMSE ranging from 9.6 to 11.6 kg [
25]. Cottam_1 and Cottam_2 models were able to explain respectively 39% and 27% of postoperative weight results at the time of formation, which is still better performance than 21% for both models obtained in our analysis [
3]. Interestingly, Cottam_2 in our study presented better calibration than in original research (SE 5.11 kg/m
2 vs 14.9 kg/m
2) [
3]. As RMSE was calculated in different units, we cannot make direct comparison of Wise, Seyssel, and Cottam_1 calibration between our analysis and original studies. During the creation, Janik model had accuracy comparable to Seyssel with ability to explain 67% of outcomes with only 0.12 kg/m
2 difference from the actual values [
26]. Nevertheless, in our population, it was able to predict 24% of the postoperative weight with higher error of 5.03kg/m
2.
It is striking that all models explored in our analysis demonstrated considerably lower goodness-of-fit in comparison to the primary studies or previous validation researches. Possible explanation of such performance may be unstandardized surgical techniques. There are many technical differences in bariatric procedures including volume of the pouch or limb lengths in RYGB and distance from the pylorus, bougie size, or completeness of fundus resection in SG [
35,
36]. As all of these procedural modifications and adjustments have significant impact on clinical outcomes, they may have affected comparison among different bariatric centers [
35‐
37].
According to the comparison of predicted and observed postoperative BMI and differences between them, vast majority of examined models significantly overestimated ultimately achieved weight loss. Observed tendency may stem from baseline characteristic of the study group. The rates of comorbidities among patients seen in our cohort, particularly T2DM and HTN, are higher than those reported in the literature [
38]. Chronic medical conditions are known to have significant impact on weight loss outcome after bariatric surgery [
8]. Thus, they may have contributed to the overestimation of weight loss in our population.
Additionally, analysis of results in patients solely after RYGB or SG revealed better accuracy and calibration of all models in RYGB group than in SG group, even for tools originally dedicated to SG. Data shows that there is much wider variability in weight loss outcomes after SG in comparison to RYGB, which makes them more difficult to predict [
1]. Consequently, this may have led to higher discrepancy in estimations and could explain observed differences.
The model proposed by Seyssel et al. achieved in our study the highest
R2 value along with the lowest SE and modest difference between predicted and observed postoperative BMI, suggesting it to have the closest fit and be most likely to predict weight loss after bariatric surgery. Nevertheless, predictive accuracy of the model still remains at the insufficient level. As there is a wide range of previously mentioned factors influencing effects of bariatric treatment, not included into analyzed equations, it seems essential for further studies to incorporate these additional variables into prediction models so as to improve their accuracy [
6‐
11].
It is worth noticing that there are other components affecting weight loss not easily measurable or impossible to obtain before surgery including eating habits, educational status, impulsivity, genetics, or gut hormone response [
7,
9,
39,
40]. They are unavailable for the development of preoperative mathematical equation, resulting in limited predictive power of any preoperative model.
R2 values presented in our research may be the bound of explanatory value of preoperative predictors as there are many other factors which play important role in the process of body weight reduction.
Limitations
The study has several limitations. Firstly, due to its retrospective design, there is possible inconsistency of collected data. Secondly, as the study was undertaken in single center, the number of participants was relatively low. Although a number of patients in all samples were rather small for data extrapolations, they were adequately powered to provide reliable external validation of the models. Furthermore, our study comprised only Caucasian patients. As there is extensive evidence for weight loss variability among ethnical groups, it is unclear whether similar findings can be transmitted into worldwide population [
7]. Further prospective validation with a larger sample size including more diverse population is needed to fully understand the efficacy of the prediction models and confirm ethnic differences.
Additionally, we only assessed results of 1-year follow-up which is insufficient to provide estimation of sustained weight loss. However, such timescale seems to be more useful, as the majority of weight loss after both SG and RYGB is achieved within the first year [
1]. More importantly application of abovementioned criteria for the length of follow-up was necessary to provide accurate validation comparable with previous outcomes as majority of examined models were developed with the use of 1-year supervision. Moreover, setting 1-year observation enabled to avoid large drop-off in follow-up reported by other authors [
22,
33]. Finally, other factors which may have an effect on weight loss such as development of postoperative complications and modification in physical activity or dietary behavior after surgery as well as alcohol intake, substance abuse, compliance to visits, psychological profiles, genetic background and support groups participation.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.