Objective urine flow assessment is clinically valuable for screening and follow-up of lower urinary tract disorders, yet standard uroflowmetry typically requires dedicated equipment and clinic-based measurement. In this study, the feasibility of estimating urine flow rate from audio recordings captured by everyday devices was investigated using a lightweight and interpretable machine-learning pipeline.

A publicly available simulated voiding sound dataset recorded with three device types, including a smartphone, a smartwatch, and an external microphone, was used, covering flow rates from 1 to 50 ml/s. In the proposed approach, engineered time- and frequency-domain acoustic descriptors were extracted from each recording, and multiple regression models were trained and evaluated separately for each device using nested cross-validation.

When both the model family and its hyperparameters were selected within the inner loop, the complete nested pipeline yielded outer-fold mean absolute errors (MAEs) of 1.15 ± 0.34, 1.63 ± 0.82, and 1.62 ± 1.05 ml/s for the smartphone, external microphone, and smartwatch recordings, respectively. Across ten repeated nested partitions, the corresponding mean MAEs were 1.47, 2.01, and 1.43 ml/s.

Errors were largest at the lowest flows (1–5 ml/s), increased when contiguous flow ranges had to be extrapolated, and a model based on the single most informative descriptor reached 2.2–2.8 ml/s. SHAP-based feature analysis highlighted the temporal variability of spectral spread and centroid and the band-energy ratio among the most influential predictors, although the feature rankings depended partly on the model type.

These findings suggest that the proposed pipeline provides an accessible, non-invasive, and interpretable framework for preliminary urine flow estimation. Its structure may also support future development of mobile health applications for longitudinal monitoring and decision support systems.