Bariatric Surgery and Type 2 Diabetes - Analysis of existing and development of new prediction models for post-surgery remission of type 2 diabetes

Abstract

Background: Type 2 diabetes (T2D) is a current global health issue that needs to be addressed as it causes premature morbidity and mortality. Importantly T2D is an expensive disease to medically manage which places a financial burden on the New Zealand healthcare system. Obesity is a key contributor to the increasing prevalence of T2D worldwide. Conversely, weight loss is linearly associated with the likelihood of T2D remission. Bariatric surgery provides effective and sustainable long term weight loss and diabetes remission. However, due to surgical expenses there is limited capacity to provide bariatric surgery in the public health system. Those with the greatest likelihood of remission of T2D are prioritised for cost-effectiveness. Several T2D remission prediction tools have been developed but these tools have yet to be tested in the New Zealand population. Improved prediction models for the NZ population are yet to be developed for short term (1-year) or long term (5-year) T2D remission post bariatric surgery. Aims: (1) To test the performance of existing T2D remission prediction tools in our New Zealand cohort and identify the most effective existing prediction tool. (2) To develop new prediction models to predict short term (1-year) or long term (5-year) T2D remission following Silastic Ring Laparoscopic Roux-en-Y Gastric Bypass (SR-LRYGB) or Laparoscopic Sleeve Gastrectomy (LSG). Methods: 114 participants aged 20-55-years with obesity (BMI 35-65kg/m2) and T2D who had been randomised to SR-LRYGB or LSG were followed up 1-year (100%) and 5-years post bariatric surgery (95%). T2D remission was defined as an HbA1c of <6% (42mmol/mol) without the need of any glucose-lowering medication, assessed at the 1-year and 5-year post-operative visits. A prediction score was manually calculated for each patient based on the scoring criteria of each existing prediction tool. Six prediction tools were tested: ABCD tool for RYGB (ABCDGBP), ABCD for SG (ABCDSG), DiaRem and Ad-DiaRem for RYGB and Metabolic Surgery Diabetes Remission (MDR) and DiaBetter tools for both RYGB and SG. Discrimination and calibration performance of each existing prediction tool was tested in our cohort using ROC curves to identify the area under the receiver operating characteristic curve (AUC) and calibration plots. The performance of each existing prediction tool at the literature identified cut-off score indicative of T2D remission was tested in our cohort by calculating sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and predictive accuracy. New cut off scores indicative of T2D remission in our cohort were identified from the ROC curves by maximising the Youden index. A new prediction model was developed for 1 and 5 year T2D remission by identifying the statistically significant predictors using univariate analysis followed by multiple logistic regression analysis. These were internally validated using bootstrap resampling methodology based on the AUC as the criterion for predictive accuracy. Calibration performance of the new prediction models were assessed through visual interpretation of the calibration plots. Results: At 1-year post bariatric surgery, 71.9% of our cohort had T2D remission which decreased to 39.8% 5-years post bariatric surgery. AUC ranged from 0.69-0.88 and 0.76-0.82 for the six existing prediction tools applied to our cohort 1 and 5-years post bariatric surgery, respectively. The MDR tool (utilising pre-operative age, HOMA-B, diabetes duration, HbA1c) performed best at predicting T2D remission 1-year post bariatric surgery (AUC 0.88) compared to all other tools. In our cohort the MDR cut-off score of 5.4 had the highest Youden index making it the score with highest sensitivity and specificity for predicting T2D remission at both 1 and 5 years post bariatric surgery. This compares to the literature identified cut-off score of ≥4 for the MDR tool which had the highest predictive accuracy of 81.58% compared to the other tools for the prediction of 1-year T2D remission in our cohort. The DiaRem tool had the highest predictive accuracy of 75.47% at the literature identified cut-off score of <7 for predicting 5-year T2D remission compared to the other tools. The DiaBetter tool (utilising pre-operative HbA1c, diabetes duration and anti-diabetic medication) performed best at predicting T2D remission 5-years post bariatric surgery (AUC 0.82) compared to all other tools. In our cohort the DiaBetter cut off score of 5.5 had the highest Youden index making it the score with highest sensitivity plus specificity for predicting T2D remission at both 1 and 5 years post bariatric surgery. No existing cut-off score had been identified in the literature for the DiaBetter tool. Multiple logistic regression analysis demonstrated weight loss to be a statistically significant independent post-operative predictor of T2D remission 5-years post RYGB and SG in our cohort (p= 0.0006). Diabetes duration (p= 0.0024), insulin use (p= 0.0030) and HbA1c (p <0.0001) were statistically significant independent preoperative predictors of T2D remission 1-year post RYGB or SG. Only diabetes duration (p <0.0001) and HbA1c (p= 0.0315) were independent preoperative predictors of T2D remission 5-years post RYGB or SG. These independent preoperative predictors were used to build the New Zealand Diabetes Remission Model (NZ-DRM). The 1-year and 5 year NZ-DRM AUC was 0.90 and 0.81 respectively with an average optimism from internal validation of 0.032 and 0.035 (bootstrap-corrected predictive accuracy 0.86 and 0.78) respectively. Conclusions: The NZ-DRM offers a model with comparable performance to the DiaBetter tool for prediction of long term (5-year) T2D remission in the New Zealand cohort.