Predicting Success in Women Chosing Firstline Treatment for Urinary Incontinence

Objectives: This study aimed to develop a model for predicting success in non-surgical treatment of urinary incontinence in women within 14 days of treatment initiation.

Methods: We used data from a cohort of patients with symptomatic urinary incontinence who initiated an 8-week regimen of pelvic floor muscle training guided by a motion-based digital therapeutic device between January 2022 and June 2023. Symptoms were assessed using the Urogenital Distress Inventory Short Form (UDI-6) survey. We included patients who were symptomatic at baseline and completed at least one follow-up survey four or more weeks after initiating therapy. We defined treatment success as a final UDI-6 score ≤ 37.5 on the final survey, indicative of Patient Acceptable Symptom State. Predictors included demographics (age, BMI, race/ethnicity), symptom severity (baseline UDI-6 responses), and data captured by the device within 14 days of first use. The cohort was split at random into training (80%) and testing (20%) sets. We fit random forest, decision tree, logistic regression, gradient boosting, k-nearest neighbors, naive Bayes, stochastic gradient descent, and partial least squares models in the training set. We evaluated discrimination (area under receive operator characteristic curves [AUC]) and calibration curves in the test set for each model. We used standardized coefficients to evaluate the relative importance of each predictor for the final model. Finally, we categorized patients in the entire cohort into groups predicted to have low ( < 25%), intermediate (25-60%), and high (>60%) probability of treatment success and estimated the actual success rate in each category by 10-fold cross-validation.

Results: We identified 647 patients, 41.9% of whom achieved treatment success. Mean baseline UDI-6 score was 56.8 +/- 14.5. The models varied significantly in their ability to discriminate between patients by outcome (AUC range 0.53 to 0.75). The random forest (AUC = 0.75) and logistic regression (AUC=0.74) models had the best discrimination, but the logistic model achieved a wider range of predicted probabilities, which were closer to observed probabilities (Figure). Baseline UDI-6 survey responses, frequency of uses, and the mean hold and lift angles achieved during pelvic floor muscle training were the most important predictors of treatment success. The cross-validate probability of treatment success, among 35.0% of patients categorized as highly likely to succeed, was 66%. In contrast, among the 16.3% categorized as unlikely to succeed, the probability of success was only 21.9%.

Conclusions: A combination of demographic, baseline symptoms, and data captured by a digital pelvic floor training device within 14 days of initiating treatment can predict the probability of achieving a Patient Acceptable Symptom State for urinary incontinence with good discrimination and calibration. External validation is required to confirm these results before the model can be put into clinical use.