Machine Learning–Based Predictive Models For Outcomes After Metabolic Bariatric Surgery: Accuracy, Validation, And Clinical Utility — A Systematic Review And Meta-Analysis

Abstract

Background: To predict clinical outcomes, including weight-loss outcomes, diabetes remission, and postoperative safety, the use of machine learning (ML) predictive models is increasingly popular in metabolic bariatric surgery (MBS). However, these models possess lower predictive accuracy, external validity and clinical utility.

Methods: We divided the PRISMA 2020 guidelines on how to conduct a systematic review and meta-analysis. PubMed, Embase, Scopus, Web of Science, and Cochrane Library (January 2015-September 2024) were detailed searched to identify the literature that developed or validated the ML models to predict the postoperative outcomes in the MBS patients. Data on the nature of the studies (type of surgical procedure, type of model, performance measures [area under the receiver operating characteristic curve [AUROC], accuracy, calibration], method of validation, and risk of bias [ PROBAST] were independently screened and extracted by two reviewers. Random-effects meta-analyses involved the utilization of the AUROC and accuracy along with the key outcomes consideration; I 2 and Cochran’s Q were used to measure heterogeneity whereas funnel plot and Egger test were used to measure publication bias.

Results: 34 studies, with 42185 patients, were considered. The most common were the Roux-en-Y gastric bypass (52) and sleeve gastrectomy (41). The commonly used algorithms were the logistic regression, random forest, gradient boosting, and neural networks. The high-weight-loss-success, diabetes-remission, and prediction of complications had a high pooled discriminative performance: 0.83 (95% CI, 0.80786 I 2 = 58%), 0.81 (95% CI, 0.77884 I 2 = 61%), and 0.79 (95% CI, 0.74884 I 2 = 55%) respectively. Only 35 percent of the studies the external validation was only done in and had a marginally lower AUROC compared to internal validation (0.79 vs. 0.84, p = 0.04). Calibration reporting and clinical utility analysis was not widespread. The total threat of bias was medium, mostly due to the unreported predictors in all cases, and a complete lack of model transparency. The test of Egger and funnel plots showed that there was low publication bias.

Conclusion: ML-based predictive models of MBS have high accuracy in weight-loss and remission of diabetes and moderate predictive accuracy of complication. However, the external validation remains weak and inconsistency in calibration exists which restricts the use in real life. In order to make these models part of the regularity of bariatric practice placing more emphasis on standardized outcome definitions, some transparent reporting and multi-centre validation would be needed.