Non-invasive prediction of detrusor underactivity in benign prostatic hyperplasia: an interpretable machine learning framework to optimize surgical selection.

Objective: To develop and internally validate an interpretable, non-invasive machine learning framework to predict detrusor underactivity (DU) in patients with benign prostatic hyperplasia (BPH). Methods: This retrospective cohort study enrolled 538 urodynamically evaluated BPH patients. A rigorous multidimensional feature selection pipeline (LASSO, Boruta, and Recursive Feature Elimination) distilled 15 baseline clinical, anatomical, and uroflowmetry parameters into a parsimonious five-feature subset. Five supervised machine learning algorithms were trained and systematically compared. Shapley Additive exPlanations (SHAP) analysis was integrated for global and local interpretability. Results: The optimized XGBoost model demonstrated superior discriminatory performance (AUC = 0.958), significantly outperforming traditional multivariable logistic regression (AUC = 0.787). XGBoost consistently exhibited superior calibration and higher net clinical benefit across varied threshold probabilities. Crucially, SHAP global dependence plots revealed non-linear pathological trajectories, notably demonstrating a U-shaped risk profile for bladder wall thickness (BWT) that was not captured by classical linear statistical detection. Local SHAP visualizations effectively translated complex probabilistic outputs into individualized clinical reasoning. Conclusion: The interpretable XGBoost framework serves as a robust non-invasive risk stratification tool for DU, decoding complex non-linear clinical interactions. This algorithm holds significant potential to optimize preoperative patient selection and mitigate surgical failures in borderline clinical scenarios. Clinical trial registration: Identifier 2026-048.