Predicting Indwelling Catheter Duration in Spinal Cord Injury Patients With Neurogenic Bladder Using Machine Learning.

OBJECTIVE: To develop and validate interpretable machine learning (ML) survival models for predicting indwelling catheter time (ICT) in early-stage spinal cord injury (SCI) patients with neurogenic bladder (NB), identify key early predictors, and derive a practical prognostic stratification tool. METHODS: We retrospectively analyzed 135 consecutive early-stage SCI-NB patients admitted between November 2017 and January 2026. The terminal event was defined as indwelling catheter conversion (ICC). Five ML survival models were developed to estimate the probability of ICC within 1 year, enabling stratification of patients into favorable-, intermediate-, and unfavorable-prognosis groups. RESULTS: The median ICT was 72 days (IQR 33.65-171.00). 102 patients (75.56%) experienced ICC during follow-up. The final SHAP-informed five-variable Random Survival Forest (RSF) model demonstrated strong discriminative performance, with a validation C-index of 0.8742. Time-dependent area under the receiver operating characteristic curve (AUC) values for the final RSF model were 0.9352, 0.8834, and 0.8803 at 3, 6, and 12 months, respectively. SHapley Additive exPlanations (SHAP) analysis identified five dominant, early-obtainable predictors: H-reflex, Lower extremity motor score (LEMS), Urinary tract infection (UTI), Time from lesion to rehabilitation facility (TLRF), Spinal cord independence measure (SCIM) bladder score. An RSF-derived prognostic output generated by the final five-variable RSF model stratified patients into three distinct prognostic groups: favorable, intermediate, and unfavorable prognosis, with significantly different 3-month cumulative ICC incidences (97.30%, 50.23%, and 2.22%, respectively; log-rank p < 0.0001). CONCLUSIONS: We developed a novel prognostic stratification model for predicting ICC in NB patients after SCI. The model demonstrates high predictive accuracy and effective patient stratification. By leveraging an interpretable ML approach, it shows potential to assist clinicians in formulating personalized treatment strategies. TRIAL REGISTRATION: The study protocol was prospectively registered with the Chinese Clinical Trial Registry (ChiCTR2300073824; registered July 21, 2023).