Dynamic machine learning model integrating resting energy expenditure for predicting postoperative complications after gastrectomy for gastric cancer

To investigate whether perioperative resting energy expenditure (REE) dynamics improve prediction of postoperative complications after gastrectomy for gastric cancer. We retrospectively analyzed 193 patients who underwent elective gastrectomy for gastric cancer. REE was measured by indirect calorimetry preoperatively and on postoperative day 1 (POD1). REE metrics were expressed as the ratio of measured REE to Harris–Benedict predicted REE (preH-B% and D1H-B%). Postoperative complications were defined as Clavien–Dindo grade II or higher. Candidate predictors were prioritized using random forest, support vector machine, and least absolute shrinkage and selection operator regression. A traditional model was compared with an integrated model including metabolic indices, and a parsimonious model was subsequently developed for clinical visualization. Discrimination, reclassification, and calibration were evaluated using AUC, DeLong’s test, integrated discrimination improvement (IDI), net reclassification improvement (NRI), and bootstrap internal validation. Postoperative complications occurred in 23 of 193 patients (11.9%). PreH-B% was associated with BMI and more advanced tumor stage, whereas D1H-B% was associated with the extent of resection (all P < 0.05). Across all feature-selection methods, preH-B% and D1H-B% were consistently prioritized. The 7-predictor integrated model showed higher discrimination than the traditional model alone (AUC 0.803 95% CI 0.704–0.903 vs 0.654 95% CI 0.538–0.771; DeLong P = 0.0049). A parsimonious 3-predictor model including preH-B%, D1H-B%, and BMI showed an apparent AUC of 0.783 and an optimism-corrected AUC of 0.757, with satisfactory calibration. Perioperative REE dynamics may provide complementary information for predicting postoperative complications after gastrectomy. These findings should be considered hypothesis-generating, and require validation in larger, prospective, multicenter cohorts before clinical implementation. The clinical utility of this model should be further evaluated using decision-curve analysis and prospective validation.