Construction and clinical validation of a machine learning-based risk prediction model for poor wound healing after anorectal surgery
Abstract
Objective To construct and validate a machine learning-based risk prediction model for poor wound healing after anorectal surgery, enabling early identification and individualized intervention. Methods Clinical data from 246 patients (Jan 2022–Dec 2024) were retrospectively collected. Wound healing status at 4 weeks post-surgery defined outcome (good: n = 191; poor: n = 55). Independent predictors were identified via multivariate Logistic stepwise regression. Patients were split 7:3 into training and internal validation sets. Three models—Logistic regression, gradient boosting machine (GBM), and random forest (RF)—were developed. Performance was assessed via AUC, calibration curves, Brier scores, and decision curve analysis. Temporal validation was performed using 152 patients from the same center (Jan 2025–Dec 2025) as an independent time-based cohort. Results BMI, diabetes, stool consistency, and preoperative perianal infection were common core predictors. In internal validation, AUCs were 0.893 (Logistic), 0.897 (RF), and 0.877 (GBM). In temporal validation, AUCs were 0.874 (Logistic), 0.865 (GBM), and 0.853 (RF). Logistic regression showed the smallest AUC fluctuation across training, internal, and temporal sets (0.864→0.893→0.874) and lowest Brier scores (0.092–0.123), indicating best calibration and generalizability. GBM achieved the highest training AUC (0.925) but significant performance decay in validation (ΔAUC = 0.048), suggesting overfitting. After grid search with 10-fold cross-validation, RF attained internal AUC of 0.897 and sensitivity of 0.800, close to Logistic regression, but temporal performance (AUC=0.853, Brier=0.129) was slightly inferior. Conclusion Logistic regression demonstrated the best overall discriminative ability, calibration, clinical net benefit, and cross-cohort generalizability, and is recommended as the preferred tool for individualized risk assessment. RF, after thorough optimization, shows promise as an alternative. GBM was limited by sample size and exhibited overfitting, warranting further validation in larger studies. This study provides a quantitative framework for early risk stratification in anorectal surgery patients.