Biomarker-integrated machine learning models for predicting treatment-related bowel dysfunction after rectal cancer surgery: a multicenter development and external validation study

Abstract

Background Low anterior resection syndrome (LARS) is a prevalent treatment-related complication following sphincter-preserving surgery for rectal cancer that severely impairs survivors’ quality of life. In the era of biomarker-driven personalized cancer therapy, reliable tools integrating clinicopathological biomarkers for individualized risk stratification remain lacking. Machine learning (ML) approaches offer the potential to integrate heterogeneous biomarkers—including nutritional, inflammatory, and treatment-related indicators—for precision prediction, yet their comparative performance against conventional nomograms in LARS prediction has not been rigorously evaluated in multicenter settings. Methods This multicenter retrospective cohort included 906 patients undergoing laparoscopic low anterior resection for rectal adenocarcinoma, with 506 in the training cohort and 400 from four hospitals for external validation. Major LARS was defined as a 6-month LARS score ≥ 21. LASSO selected features from 24 candidate variables. Five machine-learning models and a logistic regression nomogram were evaluated using AUC, calibration, decision curve analysis, and SHAP interpretation. Results The incidence of major LARS was 48.0% (243/506) in the training cohort and 52.2% (209/400) in the validation cohort. LASSO regression selected 14 predictive biomarkers, including preoperative serum albumin, neoadjuvant chemotherapy, and preoperative radiotherapy. In external validation, the random forest model achieved the highest AUC of 0.903 (sensitivity 0.871, specificity 0.796), followed by the logistic regression nomogram (AUC 0.888). Multivariate analysis identified tumor–anal verge distance (OR = 0.733, 95% CI 0.679–0.792, p < 0.001), stoma reversal time (OR = 1.240, p < 0.001), anastomotic leakage (OR = 7.027, p = 0.004), and diabetes mellitus (OR = 2.793, p = 0.006) as independent risk factors. SHAP analysis confirmed tumor–anal verge distance as the dominant predictor, with preoperative albumin ranking fifth in feature importance, demonstrating the capacity of ML to extract predictive signal from clinically relevant but statistically marginal biomarkers. Conclusion The random forest model integrating clinicopathological biomarkers and treatment-related variables demonstrated robust predictive performance for postoperative or post-treatment risk stratification of treatment-related bowel dysfunction (LARS). The model is intended to support postoperative monitoring, survivorship counseling, and rehabilitation planning rather than purely preoperative treatment selection. These biomarker-driven prediction tools can facilitate individualized functional-risk assessment in rectal cancer patients undergoing multimodal therapy.