Exploring clinical and sociodemographic patterns of treatment completion and drop-out in pelvic floor muscle training: a deep learning and SHAP-based analysis

Abstract Objective The aim of this study was to identify the clinical and sociodemographic factors that influence Pelvic Floor Muscle Training (PFMT) completion using SHapley Additive exPlanations (SHAP) analysis. Design An exploratory retrospective study. Setting A Women’s Health Unit in Department of Physiotherapy and Rehabilitation. Participants 377 women clinically diagnosed with urinary incontinence (UI) and referred for PFMT. Interventions An 8-week supervised PFMT program. Treatment completion was defined as attending at least 80% of sessions and completing the final assessment. Main outcomes and measures Data included sociodemographic features, UI types, presence of pelvic organ prolapse (POP), POP stages, and various validated scales (ISI, ICIQ-SF, IIQ-7, and PFDI-20 subscales). A deep learning model integrated with regularization techniques (Batch Normalization, Dropout) was developed to predict completion, with SHAP analysis used to interpret non-linear feature contributions. Results Treatment was completed by 29.7% of participants (n = 112), with the deep learning model achieving 85.3% accuracy. Clinical and symptom-related variables were the most influential predictors. High ‘VAS–Wetness’ and ‘Leakage during laughing’ strongly predicted completion, while Stress UI was more favorable than complex types like Mixed UI. Higher ‘POPDI-6’, ‘IIQ-7’, ‘Leakage Amount’, and ‘Leakage during sneezing’ correlated with lower completion; conversely, ‘Leakage during lifting’, ‘Maximum pelvic floor pressure’, and ‘Hypertension’ were positive predictors. Sociodemographically, a higher ‘Living Children Count’ negatively impacted completion, whereas ‘Occupation’ was a consistent positive predictor. Conclusions PFMT completion is primarily driven by clinical symptom perception and disease complexity rather than sociodemographic characteristics. While high perceived wetness and specific triggers like laughing promote adherence, barriers such as mixed UI and a higher number of children significantly hinder persistence. Integrating AI-based predictive models into clinical practice can help identify at-risk patients and enable tailored interventions to improve treatment adherence and outcomes in women with UI.