# Reinforcement Learning–Based Digital Therapeutic Intervention for Postprostatectomy Incontinence: Development and Pilot Feasibility Study

## TL;DR

A new digital therapeutic system using reinforcement learning was developed to help patients with postprostatectomy incontinence through personalized rehabilitation.

## Contribution

The first reinforcement learning-powered digital therapeutic system for postprostatectomy incontinence is introduced.

## Key findings

- The system showed high engagement with patients using it an average of 5.2 times per day.
- Preliminary results showed reductions in urinary frequency, incontinence episodes, and symptom burden.
- Objective urine leakage and patient-reported outcomes improved, though effect magnitudes varied across individuals.

## Abstract

Postprostatectomy incontinence (PPI) is a common complication after robot-assisted radical prostatectomy and significantly impairs patients’ quality of life. Although behavioral interventions such as pelvic floor muscle training and bladder diaries are evidence-based, their effectiveness is often limited by poor adherence and lack of personalization.

This study aimed to develop and evaluate a reinforcement learning (RL)–driven clinical behavioral intervention-supporting system (CBISs) for adaptive, personalized rehabilitation in patients with PPI.

The study comprised 2 sequential stages. First, the CBISs was developed through (1) construction of a medical record database from a prospective cohort of PPI patients using standardized 3-day bladder diaries, (2) design of functional modules and user interfaces based on clinical rehabilitation needs, and (3) development of an RL model using XGBoost (extreme gradient boosting) and Bayesian optimization to generate individualized training plans. Second, a separate cohort of 16 patients participated in a single-arm, pre-post pilot study to evaluate feasibility and preliminary outcome trends over a 3-month intervention period, with assessments based on bladder diary parameters and system usage metrics.

The CBISs successfully implemented an adaptive, closed-loop behavioral rehabilitation framework that dynamically tailored training recommendations according to individual voiding patterns, fluid intake behaviors, and adherence signals. Feasibility outcomes were favorable, with high system engagement observed throughout the intervention (mean usage frequency 5.2, SD 1.1 times per day). In exploratory pre-post analyses (n=16), consistent directional improvements were observed across multiple outcomes. Mean daytime urinary frequency decreased from 5.74 (SD 1.21) episodes per day to 4.69 (SD 1.08) episodes per day, while median nighttime urinary frequency declined from 1.8 (IQR 1.6-2.2) episodes per night to 1.0 (IQR 1.0-1.6) episodes per night. Median incontinence episodes were reduced from 7.0 (IQR 6.0-11.0) episodes per day to 4.0 (IQR 2.0-6.0) episodes per day. Objective urine leakage measured by the 1-hour pad test decreased from a median of 8.5 (IQR 4.0-19.0) g to 3.5 (IQR 2.0-9.0) g. Patient-reported symptom burden, assessed using the International Consultation on Incontinence Questionnaire–Short Form (ICIQ-UI SF), showed a median reduction from 14.0 (IQR 12.0-20.0) points to 9.0 (IQR 6.0-16.0) points. Although several within-participant changes were statistically detectable, effect magnitudes varied across individuals. Given the single-arm design, small sample size, and lack of a control group, findings are presented as exploratory and hypothesis-generating rather than confirmatory of clinical efficacy.

The CBISs represents the first RL-powered digital therapeutic system for PPI, enabling adaptive, evidence-based behavioral optimization. By addressing limitations of static rehabilitation protocols and declining adherence, it offers a scalable approach for personalized PPI management. Future multicenter trials are needed to confirm its clinical effectiveness.