An explainable machine-learning model using baseline gut microbiota, circulating short-chain fatty acids, and clinical features to predict weight-loss response after Roux-en-Y gastric bypass
Juan José Hernández-Morante, María Antonia Martínez-Sánchez, Inmaculada Ros-Madrid, Virginia Esperanza Fernández-Ruiz, Mercedes Ferrer-Gómez, Antonio Silva, Andrés Balaguer-Román, María Dolores Frutos, Horacio Pérez-Sánchez, Bruno Ramos-MolinaBackground:
Roux-en-Y gastric bypass (RYGB) is one of the most effective treatments for severe obesity; however, postoperative weight-loss response remains heterogeneous. To address this variability, we developed an explainable machine-learning (ML) model integrating baseline gut microbiota composition, circulating short-chain fatty acids (SCFAs), and clinical features to predict weight-loss outcomes after RYGB.
Methods:
In this prospective cohort study, 143 patients undergoing RYGB were followed for 12 and 24 months. The participants were classified as high- or low-responders based on achieving ≥30% total weight loss (%TWL). Gut microbiota composition was assessed using 16S rRNA gene sequencing. Circulating SCFAs were quantified by liquid chromatography–mass spectrometry. Microbiome-only, clinical–metabolic, and combined models were developed using LASSO logistic regression. Internal validation was performed using stratified repeated five-fold cross-validation with 20 repetitions, and performance was estimated from participant-level out-of-fold predictions. Shapley Additive Explanations (SHAP) were used descriptively to interpret model behavior.
Results:
At 12 and 24 months after RYGB, 94 and 83 participants achieved ≥30% TWL, respectively. Baseline anthropometric parameters did not differ between the response groups. The baseline alpha diversity was higher in high responders at both 12 and 24 months (
Conclusions:
Baseline combined ML models integrating baseline gut microbiota, circulating SCFAs, and clinical features provided moderate, internally validated discrimination for marked weight-loss response after RYGB, with the strongest performance observed at 12 months. These findings support further evaluation of preoperative ML models for patient stratification. Microbiome and circulating SCFA features may provide complementary biological and interpretive information, but their incremental predictive and clinical value requires further validation.