Radiotherapy-related Gastrointestinal Adverse Events in Rectal Cancer: Risk Factor Analysis and Predictive Modeling Using Clinical and Small Bowel Dosimetric Features

Purpose To identify key risk factors for radiation-induced enteritis and construct a predictive model integrating dosimetric and clinical parameters. Methods We analyzed 206 colorectal cancer patients treated with pelvic radiotherapy (2015–2020). Clinical variables, dosimetric parameters and anatomical factors were evaluated. Toxicity endpoints included acute upper/lower gastrointestinal reactions (CTCAE v5.0) and late proctitis (LENT/SOMA). Statistical analyses included logistic regression and machine learning (Random Forest/XGBoost). Results The incidences of acute upper gastrointestinal reactions, acute lower gastrointestinal reactions, and late proctitis were 33.5% (69/206), 53.9% (109/206), and 38.3% (79/206), respectively. The incidence of acute upper gastrointestinal toxicity was higher in females compared with males (50% vs. 28%, P = 0.0088). A higher fraction number (OR = 0.74, P = 0.0485) and a lower radiation dose (OR = 0.999, P = 0.0293) were associated with a reduced incidence of acute lower gastrointestinal reactions.. For late proctitis, Vsmall intestine was a significant risk factor (AUC = 0.638, P  = 0.006), while tumor-to-anal verge distance was protective (AUC = 0.583, P  = 0.058). Machine learning models showed superior performance: random forest achieved an AUC of 0.72 for acute upper gastrointestinal reactions, XGBoost/AUC = 0.64 for acute lower gastrointestinal reactions, and LightGBM/AUC = 0.76 for late proctitis, outperforming conventional logistic regression (AUC range: 0.49–0.57 for acute endpoints; 0.57 for late proctitis). Conclusion Gender, fraction number, Vsmall intestine, and tumor-to-anal verge distance are associated with an increased risk of radiation-induced gastrointestinal toxicities. Machine learning models, particularly LightGBM for late proctitis (AUC = 0.76) and random forest for acute reactions (AUC = 0.72), provide robust tools for risk stratification. These findings may support gender-specific care, moderate hypofractionation, and stringent small bowel dose constraints to help optimize personalized radiotherapy.