Uroflowmetry is a key diagnostic tool for assessing lower urinary tract dysfunction (LUTD), providing non-invasive measurements of urinary flow patterns. The interpretation of the flow-volume curve relies on identifying morphological patterns, with a bell-shaped curve indicating normal flow, while other morphologies may suggest abnormality ​(1,2)​. However, subjective interpretation by clinicians introduces variability. This study aims to develop an AI-based platform for automated classification of uroflowmetry curve morphology.

This is a secondary analysis of an ongoing observational longitudinal study, presenting preliminary findings. We analyzed 50 uroflowmetry graphs interpreted by a single urologist (25 female, 25 male). Voided volumes ranged from 150 to 500 mL, excluding outliers. Curves were labeled as ‘Continuous’ (bell-shaped) or ‘Non-continuous’ (all other morphologies). 

A platform was developed to automatically classify uroflowmetry curves. Images were processed using computer vision algorithms to extract the curve and convert it into a one-dimensional flow signal. Frequency-based analysis, including Fast Fourier Transform (FFT), extracted 112 features per case, serving as input for a linear machine-learning classifier. Model performance was assessed using a hold-out validation set.

The model achieved a macro-averaged F1-score of 75% in distinguishing continuous from non-continuous uroflow curves. This metric combines model precision and recall into a single value, where 100% represents perfect classification. The most relevant information was identified in the low-frequency band (0.02–0.13 Hz), which refers to the slower, broader patterns in the flow signal—essentially, the overall shape and rhythm of the curve rather than rapid, tiny fluctuations. These slower patterns were critical for accurate classification, as they capture the key characteristics that define whether a flow is continuous or non-continuous. Figure 1 illustrates an example output, including the extracted flow signal and the annotated frequency spectrum highlighting the most predictive features.

Findings indicate that low-frequency components (0.02–0.13 Hz) are crucial for identifying non-continuous flow-volume morphologies. The model provides an interpretable and objective tool for uroflowmetry analysis, reducing diagnostic variability and enhancing clinical decision support.

This AI-driven approach improves uroflowmetry interpretation by reducing variability and offering quantitative insights into flow-volume alterations. Limitations include selection bias, reliance on a single urologist, and a small sample size. Future studies should incorporate prospective data and additional uroflowmetry parameters (e.g., maximum flow rate; Qmax) to develop a complete machine-learning-based assessment of uroflowmetry.

View of Age, gender, and voided volume dependency of peak urinary flow rate and uroflowmetry nomogram in a tertiary care centre [Internet]. [cited 2025 Mar 31]. Available from: https://www.apjhs.com/index.php/apjhs/article/view/7/901
Gammie A, Clarkson B, Constantinou C, Damaser M, Drinnan M, Geleijnse G, et al. International continence society guidelines on urodynamic equipment performance. Neurourol Urodyn [Internet]. 2014 Apr 1 [cited 2025 Mar 31];33(4):370–9. Available from: https://onlinelibrary.wiley.com/doi/full/10.1002/nau.22546?