Machine Learning-Based Prediction of Weight Regain Risk Following Sleeve Gastrectomy: A Multicenter Cohort Study

Abstract

Weight regain (WR) after bariatric surgery is frequent and problematic, with as many as 20–30% of sleeve gastrectomy patients experiencing clinically important weight return after 2 years. We created and validated five machine learning models (XGBoost, Random Forest, Support Vector Machine, Logistic Regression, Neural Network) to predict risk of WR after bariatric surgery from NHANES data from the pre-pandemic 2017–March 2020 cycle. Using these data, we created a BMI-proxy cohort of adults that met post-bariatric clinical criteria (peak BMI ≥35 kg/m², ≥15% weight loss from peak BMI; n=628). Weight regain was defined as weight gain of ≥2 kg within 12 months following peak weight and had a prevalence of 21.5%. Twenty-three clinical, metabolic, behavioural and psychological features served as input variables. SMOTE was used to mitigate issues from class imbalance. On the held-out test set (n=126), XGBoost had the highest performance with an AUC-ROC of 0.769 (95% CI: 0.675–0.857), sensitivity of 0.926, and NPV of 0.968, while Random Forest had comparable AUC of 0.759 (95% CI: 0.644–0.863) but higher specificity of 0.838. Pairwise statistical testing as performed by DeLong tests showed both ensembles were significantly better than SVM, Logistic Regression, and Neural Network (p-value<0.05). SHAP interpretability testing revealed self-reported intention to lose weight, HDL cholesterol level, age, and maximum BMI were the top contributing features. We were able to use ensemble ML methods to predict patients at high risk for weight regain following bariatric surgery early in their care.