機械学習を用いた説明可能な患者リスク、ERAS遵守率、および術後転帰の評価

Enhanced Recovery After Surgery (ERAS) protocols improve outcomes after colorectal surgery, but adherence remains variable and may interact with patient risk. Traditional compliance scores lack granularity to explore these dynamics. We aimed to use interpretable machine learning to quantify the contribution of individual ERAS items and clinical features to postoperative complications, and to identify data-driven ERAS phenotypes.

This was a secondary analysis of the EuroPOWER cohort (NCT04889798), a prospective European study including 2,841 adults undergoing elective colorectal surgery. Two Extreme Gradient Boosting models were trained to predict in-hospital complications: a complete model (clinical variables + 23 ERAS items) and an ERAS-only model. Both were interpreted using Shapley Additive Explanations (SHAP). In the complete model, SHAP matrices were clustered to derive phenotypes. Feature importance, adherence, and complication rates were compared descriptively.

The complete model achieved an AUC of 0.627. SHAP analysis identified frailty, ASA class, BMI, and age as leading predictors, followed by early mobilisation, nutritional care, and thromboprophylaxis. Three phenotypes were identified, with complication rates of 17.7%, 27.1%, and 41.1%, corresponding to robust, intermediate, and frail profiles. The ERAS-only model showed similar discrimination (area under the curve 0.642), but reduced interpretability. SHAP redundancy analysis supported inclusion of all ERAS items.

The clinical effect of ERAS adherence appears to be modulated by baseline vulnerability and implementation patterns. SHAP-based models enable transparent risk attribution and phenotype identification, supporting more targeted ERAS strategies and future development of automated quality monitoring tools.