Interpretable machine learning enables early identification of financial risk and resource optimization in colorectal cancer surgery

Abstract

Background Colorectal cancer (CRC) imposes major clinical and economic pressure on health systems. Under diagnosis-related group (DRG) payment, fixed reimbursement benchmarks may not capture the heterogeneity of surgical oncology or the resources required by patients with complex disease, exposing hospitals to financial losses and potentially affecting equitable access to high-quality cancer care. We developed an interpretable machine learning framework to estimate DRG-related financial risk in CRC surgery and support perioperative clinical review and local resource planning. Methods We retrospectively reviewed 2,081 patients who underwent CRC surgery between 2014 and 2024. Candidate features were screened for multicollinearity, assessed using Boruta, and refined through clinical review. Synthetic minority oversampling was then applied to reduce class imbalance. Six machine learning algorithms were trained with cross-validation and hyperparameter tuning. The best-performing model was interpreted using SHapley Additive exPlanations (SHAP), and Sequential Forward Selection was used to reduce inputs to a clinically feasible subset. These variables were incorporated into an interactive web-based tool for individualized risk assessment. Calibration plots, Brier score, and decision curve analysis were used to evaluate reliability and clinical net benefit. A prospective cohort of 12 patients tested workflow feasibility during routine perioperative assessment. Results DRG-related financial risk, defined as expenditure above the DRG benchmark, was present in 38.4% of patients in the retrospective cohort. The random forest model achieved the best discrimination, with an area under the receiver operating characteristic curve of 0.852. SHAP analysis identified surgical approach, postoperative complications, length of hospital stay, and comorbidity burden as major contributors to predicted financial risk. The simplified model remained well calibrated and showed meaningful net benefit on decision curve analysis. In the prospective phase, clinicians used the web-based tool to stratify individual risk and simulate modifiable perioperative strategies for pathway optimization in real clinical workflow. Conclusion An interpretable machine-learning framework can provide individualized estimation of DRG-related financial risk in CRC surgery. Its integration into a point-of-care decision-support tool may support perioperative risk review, resource planning, and complex-case management under DRG-based reimbursement.