Neural pathways to bariatric success: What explainable AI reveals that conventional fMRI methods may miss

Metabolic-bariatric surgery (MBS) remains a cornerstone of obesity treatment, yet 15%-30% of patients fail to achieve its intended benefits. Existing clinical and biochemical markers offer limited value in identifying who will respond favourably to this intervention. We hypothesize that the long-term success of MBS is influenced by individual differences in preoperative brain function.

We collected presurgical resting-state fMRI data from 45 patients undergoing MBS, with the aim of identifying neural patterns associated with achieving at least 50% excess weight loss 12 months post-surgery. The data were analysed using both conventional methods and a high-powered machine learning approach. For the latter, we trained five predictive models on functional connectivity, regional brain activity, and clinical variables. We then applied SHapley Additive exPlanations (SHAP) to the best-performing model to interpret its internal logic, thereby revealing the neural features most strongly linked to treatment success.

Conventional methods proved inadequate for this study. A multilayer perceptron model, trained exclusively on functional connectivity data, achieved a noteworthy AUC of 0.85. Its SHAP analysis revealed key neural circuits in the postcentral gyrus, dorsolateral prefrontal cortex, and angular gyrus-regions associated with interoception, executive control, and social-cognitive processes such as theory of mind.

Explainable AI-powered fMRI analysis uncovered subtle neural patterns that conventional methods failed to detect. These findings suggest that a patient's "neural readiness" for MBS may extend beyond self-regulatory circuits. It may also depend on their capacity to perceive and interpret internal bodily signals and to process emotional information-both personal and social.