Overactive bladder (OAB) is a chronic, multifactorial condition characterised by urinary urgency, with or without urgency urinary incontinence, usually with increased daytime frequency and nocturia, if there is no proven infection or other obvious pathology [1]. In patients with OAB, low treatment persistence rates can lead to unreliable symptom control and progression of symptom severity. The syndrome exerts a significant impact on worldwide healthcare systems, with epidemiological estimates indicating that approximately 18% of individuals aged ≥40 years in Taiwan and South Korea have OAB [2]. The objective of this study was to develop a prediction model for assessing response to antimuscarinic and mirabegron treatment in patients with OAB using data from the FAITH study.

FAITH was a prospective, longitudinal, observational registry study conducted in Taiwan and South Korea to describe the treatment patterns for OAB therapies and to identify and evaluate factors associated with the effectiveness and persistence of OAB therapies. Adult patients (≥18 years) who had been diagnosed with OAB symptoms at least 3 months prior to study enrolment were eligible for inclusion. Patients were due to initiate monotherapy with mirabegron or any antimuscarinic that had been prescribed as part of routine clinical practice for the treatment of OAB symptoms. Informed consent was required to participate in the study. The objective of this investigation was to use data from the FAITH study to successfully develop an algorithm to predict treatment effectiveness that was based on a composite outcome of safety, efficacy, and treatment change. Safety was defined as no occurrences of moderate or severe adverse events; if this occurred the treatment was deemed “safe” otherwise it was judged to be “less safe”. Efficacy was defined as a decrease in Overactive Bladder Symptom Score (OABSS) of ≥3, which has been identified as the minimal threshold for a meaningful change [3]; if this criterion was met the treatment was deemed “successful” otherwise it was “less successful”. Treatment change was defined as any treatment discontinuation, interruption, switch, or addition; if any of these were present, a “treatment change” was deemed to have occurred, otherwise “no treatment change” was recorded. Patients within the FAITH study were followed for 183 days to establish occurrences of the above parameters. The composite outcome was defined as “safe”, “successful”, and “no treatment change”; which meant the treatment was deemed “more effective”, otherwise it was “less effective”. To explore the composite algorithm, a total of 14 clinical risk factors were included in the initial data set, which included demographic factors (age, sex, body mass index, and Charlson Comorbidity Index), OAB symptoms (frequency, urgency, nocturia, OAB type, and incontinence type), questionnaire data (Baseline OABSS and overactive bladder questionnaire [OAB-q] short form [bother and health-related quality of life scores]), any prior OAB medication, and planned treatment. Using data from the initial data set, a 10-fold cross-validation procedure was performed, and a range of machine learning models were evaluated to determine the most effective algorithm for prediction of treatment effectiveness. In total, six different models were chosen (regularised regression, decision tree [C5.0], boosted tree model, random forest, neural network [multilayer perceptron (MLP)], and support vector machine) and tested using a variety of recipes. Due to the nature of some of the models, it was necessary to format the data in different ways to ensure the optimal performance of each model algorithm. Receiver operating characteristic (ROC) area under the curve (AUC) analyses were used to determine rank performance for each model. The chosen model was optimised to help reduce model error rate and maximise ROC curve stability (method of validation) and model generalisability (minimising overfitting). The parameters used to optimise the model were the input parameters within the model from the original set within the baseline characteristics analysis and the number of patients within each tree “split” or “branch” (min_n). To determine the variables that should be included in the final model, the percentage that each variable appeared following the use of 50 resamples was calculated, as using a single train/test split was considered unstable. The final model performance was evaluated using the chosen variables and with the minimum n parameter set to the optimal level. The model was subsequently fitted over the 50 resamples to enable an average AUC and 95% confidence intervals (CIs) to be calculated. The final algorithm was integrated into an online application for intended use as an educational assessment tool. The online platform was constructed to allow for the prediction of “more effective” or “less effective” treatment results for both mirabegron and antimuscarinics.

In total, 396 patients from the FAITH study were included in this analysis. Of these, 266 (67.2%) initiated treatment with mirabegron and 130 (32.8%) initiated treatment with an antimuscarinic. In terms of the composite outcome, 138 (34.8%) were judged to be in the “more effective” group and 258 (65.2%) were in the “less effective” group (Figure 1). Data from the initial data set showed that incontinence type, OAB type, OAB-q score, and Baseline OABSS were significantly different between the “more effective” and “less effective” groups. The ROC analyses indicated that the best performance was achieved with the random forest (AUC: 0.66) and decision tree (C5.0; AUC: 0.65) models. Given the relative simplicity of the decision tree model and its comparable performance to the more complex random forest algorithm, the decision tree (C5.0) model was chosen as it allows for easier interpretation and implementation. The resample analysis showed that Baseline OABSS, planned treatment, OAB-q health-related quality of life scores, urgency, and incontinence type appeared in the models >75% of the time over the 50 resamples and were therefore included in the final model. For the ROC analysis, an AUC result of 0.70 (95% CI: 0.54, 0.85) was achieved (Figure 2) when the optimal value of 15 was used for the minimum n parameter. The input parameters for the online application were integrated to match the inputs for the final decision tree model, namely the OABSS and OAB-q questionnaires, the urination urgency for the patient, and the presence/absence of urge urinary incontinence or urinary incontinence.

A tool was developed that effectively predicted more effective and less effective treatment responses following antimuscarinic or mirabegron therapy in patients with OAB. The final algorithm was based on an optimised decision tree (C5.0) model. This machine learning algorithm model is currently designed specifically for research purposes. To be utilised in clinical practice to predict treatment responses, further analysis, validation, inclusion of additional data, adjustment of clinical input, and possibly certification are required.

This study used real-world evidence from the FAITH study to successfully develop an algorithm to predict treatment effectiveness for patients with OAB. The algorithm was developed using a composite outcome of safety, efficacy, and treatment change, achieving an AUC of 0.70 with the final model. The design and methodology permitted the integration of the algorithm into a simple, rapid, and easy-to-use interface that could be refined in the future to produce a valuable educational or clinical decision-making aid.

Drake MJ. Neurourol Urodyn 2014; 33: 622-624.
Chuang Y-C, et al. Low Urin Tract Symptoms 2019; 11: 48-55.
Gotoh M, et al. Urology 2011; 78: 768-773.

Continence 2S2 (2022) 100366
DOI: 10.1016/j.cont.2022.100366