Precision Metabolic and Bariatric Medicine: Reconciling Predictive Machine Learning with Explanatory Biostatistics to Validate the Independent Effect of a 10.5% Preoperative TBWL Threshold.

backgroundPreoperative total body weight loss (PreopTBWL%) is widely used before metabolic and bariatric surgery (MBS), yet its long-term prognostic value remains debated. This study integrated machine learning (ML), unsupervised clustering, and linear mixed-effects modeling (LMM) to define data-driven preoperative targets and characterize multi-year postoperative weight trajectories.

methodsWe analyzed prospectively collected data and electronic medical records from 802 adults undergoing primary laparoscopic sleeve gastrectomy, primary Roux-en-Y gastric bypass, or revision/conversion MBS within a structured, protein-focused preoperative lifestyle program. Five machine learning models were trained with nested cross-validation to predict each postoperative year. Each model used only patients with that year's outcome recorded (608 at Year 1 to 83 at Year 6). From Year 2 onward, earlier postoperative measurements for those patients were included as predictors. Priority was given to Years 1-4. Top-performing SHAP profiles guided feature selection for confirmatory longitudinal LMMs. Although the LMM framework spanned the full 6-year horizon (858 repeated measures from 420 active patients), the most robust inferences were restricted to Years 1-4 because of substantial attrition thereafter. Youden Index optimization, segmented regression, and unsupervised k-means clustering on UMAP embeddings were used to define precise PreopTBWL% targets and map distinct multi-year phenotypic subgroups.

resultsMachine learning models demonstrated strong predictive performance, peaking at Year 2 (Random Forest R² = 0.524 ± 0.033, AUC = 0.875 ± 0.029) and remaining robust through Year 4, with PreopTBWL% consistently ranking among the strongest prognostic features. Unsupervised clustering (k = 5) revealed that preoperative behavioral engagement separated high- from low-performing phenotypes far more powerfully than procedure type or demographics. Initial racial disparities in weight-loss trajectories narrowed progressively and were no longer statistically significant by Year 4 (p = 0.1454). Linear mixed-effects modeling confirmed a robust, independent positive association between PreopTBWL% and long-term outcomes (β = 0.447, p < 0.001). Threshold optimization identified 10.5% PreopTBWL% as the optimal clinical target; patients falling below this threshold experienced a significant long-term deficit (β = -4.608, p = 0.017) that persisted after adjustment for diabetes and operative duration. Interaction terms between the threshold and surgical procedure were non-significant (Threshold × Sleeve, p = 0.532; Threshold × Revision/Conversion, p = 0.778). Segmented regression further identified a structural dose-response breakpoint at 19.68% PreopTBWL%.

conclusionsAchieving ≥ 10.5% PreopTBWL% was associated with durable multi-year PostopTBWL% after MBS, and procedure type did not modify that association in this cohort. By integrating predictive machine learning with explanatory longitudinal modeling, this work shifts preoperative evaluation from arbitrary administrative requirements toward structured, phenotype-guided risk stratification and care escalation.