Beyond compliance: Patient risk, ERAS adherence, and postoperative outcomes through explainable machine learning

Area of Science:

• Colorectal Surgery
• Surgical Outcomes
• Machine Learning in Medicine

Background:

• Enhanced Recovery After Surgery (ERAS) protocols are crucial for improving outcomes in colorectal surgery.
• Variable adherence to ERAS protocols and patient-specific risk factors can impact surgical outcomes.
• Traditional compliance metrics lack the granularity to analyze the interplay between ERAS adherence and patient risk.

Purpose of the Study:

• To quantify the impact of individual ERAS components and clinical factors on postoperative complications using interpretable machine learning.
• To identify distinct patient phenotypes based on ERAS adherence and clinical characteristics.
• To develop data-driven insights for optimizing ERAS strategies in colorectal surgery.

Main Methods:

• Secondary analysis of the prospective EuroPOWER cohort (NCT04889798) involving 2,841 adult patients undergoing elective colorectal surgery.
• Development and interpretation of two Extreme Gradient Boosting models using Shapley Additive Explanations (SHAP) to predict in-hospital complications.
• Clustering of SHAP matrices from a comprehensive model (clinical variables + ERAS items) to derive patient phenotypes.

Main Results:

• The comprehensive model demonstrated good predictive performance (AUC 0.627).
• Key predictors of complications included patient frailty, ASA class, BMI, and age, alongside ERAS components like early mobilization and nutritional care.
• Three distinct patient phenotypes (robust, intermediate, frail) were identified with varying complication rates (17.7% to 41.1%).

Conclusions:

• ERAS adherence's effectiveness is influenced by patient baseline vulnerability and protocol implementation.
• SHAP-based models offer transparent risk assessment and enable the identification of patient phenotypes.
• These findings support the development of personalized ERAS strategies and automated quality monitoring tools for surgical care.