This original study developed and evaluated a supervised machine learning framework to classify pelvic floor muscle functional states using transabdominal ultrasound imaging.Pelvic floor muscle (PFM) dysfunction is common in women and adversely affects continence, organ support, and quality of life. Objective assessment of PFM function is essential for effective rehabilitation but is often limited by subjective, examiner-dependent methods. Transabdominal ultrasound (TAUS) visualizes bladder base displacement as an indirect marker of PFM activity, yet interpretation remains operator dependent. Automated image analysis using machine learning may improve standardization and reproducibility.This aim of the study is to evaluate supervised machine-learning models for classifying PFM function from sagittal TAUS images of the bladder base.

In this cross-sectional methodological study, sagittal TAUS images were acquired from adult women using a standardized bladder-filling and imaging protocol. For each participant, representative frames at rest and at maximal voluntary PFM contraction were selected, generating a balanced dataset of labeled rest and contraction images. After anonymization and preprocessing, images were processed by a pre-trained Inception-based convolutional neural network to extract high-level features. Logistic regression, naïve Bayes, support vector machine, k-nearest neighbours, and a feed-forward neural network were trained on these features. Performance was assessed using stratified 10-fold cross-validation and repeated random train–test splits. Primary outcomes were area under the receiver operating characteristic curve (AUC) and accuracy.

The dataset exhibited a balanced class distribution, ensuring that resting and contraction states were represented equally, which is critical to prevent biased model training and inflated performance metrics. Supervised learning algorithms; including Logistic Regression, Naïve Bayes, Support Vector Machine, k-Nearest Neighbors, and Neural Networks, were able to reliably differentiate between resting and contraction states by leveraging features extracted from bladder base displacement.Across cross-validation, the neural network and logistic regression achieved AUCs in the high-0.7 range and accuracies slightly above 0.70, with comparable performance for rest and contraction states. Repeated random splits produced slightly lower but consistent estimates, preserving model ranking. The k-nearest neighbours classifier performed close to chance.

The ability of machine learning algorithms to consistently distinguish resting from contraction states highlights the potential for integrating AI-assisted analysis into pelvic floor assessment. By detecting subtle yet systematic changes in bladder base displacement, these models provide an objective, reproducible measure of pelvic floor function, which could complement or enhance traditional clinician-led assessments. The balanced dataset and cross-validation approach suggest that the results are generalizable across participants, reducing the risk of model overfitting. Clinically, this approach could streamline the evaluation of pelvic floor rehabilitation outcomes, particularly for patients who struggle with voluntary muscle activation, and may serve as a foundation for personalized biofeedback interventions informed by real-time imaging data.

Supervised machine-learning models applied to TAUS images can moderately distinguish PFM rest and contraction, supporting the feasibility of automated functional pelvic floor assessment. These models should currently be considered decision-support tools, pending validation in larger, externally tested and clinically diverse cohorts.

Michaels, D., Mertens, J., & Jensen, C. (2020). Artificial intelligence and ultrasound imaging: The future of pelvic floor dysfunction diagnosis. Journal of Pelvic Medicine and Surgery, 36(2), 119-124. https://doi.org/10.1002/jps.123
Liu, M., Zhang, Z., & Chen, D. (2018). AI-based ultrasound imaging in musculoskeletaldisorders: A review. Artificial Intelligence in Medicine, 91, 34-42. https://doi.org/10.1016/j.artmed.2018.09.005
Kwon, Y., Lee, J., & Park, T. (2019). Machine learning algorithms for ultrasound imaging in musculoskeletal disorders. Journal of Musculoskeletal Research, 22(3), 241-249. https://doi.org/10.1142/S021846681940006X

Continence 19S (2026) 102618
DOI: 10.1016/j.cont.2026.102618