Flow prediction in sound-based uroflowmetry

Abstract

Sound-based uroflowmetry (SU) offers a non-invasive alternative to traditional uroflowmetry (UF) for evaluating lower urinary tract dysfunctions, enabling home-based testing and reducing the need for clinic visits. This study compares SU and UF in estimating urine flow rate and voided volume in 50 male volunteers (aged 18-60), with UF results from a Minze uroflowmeter as the reference standard. Audio signals recorded during voiding were segmented and machine learning algorithms (gradient boosting, random forest, and support vector machine) estimated flow parameters from three devices: Ultramic384k, Mi A1 smartphone, and Oppo smartwatch. The mean absolute error for flow rate estimation were 2.6, 2.5 and 2.9 ml/s, with R² values of 84%, 83%, and 79%, respectively. Analysis of the Ultramic384k's frequency range showed that the 0-8 kHz band contained 83% of significant components, suggesting higher sampling frequencies are unnecessary. A 1000 ms segment size was optimal for balancing computational efficiency and accuracy. Lin's concordance coefficients for urine flow and voided volume using the smartwatch (0-8 kHz, 1000 ms) were 0.9 and 0.85, respectively, demonstrating that SU is a reliable, cost-effective alternative to UF for estimating key uroflowmetry parameters, with added patient convenience.

Keywords: Acoustic voiding signals; Flow prediction; Machine learning; Sound-based uroflowmetry.

Fig. 1

Example of data recorded by the Minze uroflowmeter software during a test.

Fig. 2

Laboratory data collection scenario showing the Minze uroflowmeter with a water volume of 400 ml in the basin and the three recording devices: UM, Phone, and Watch, along with their respective heights.

Fig. 3

Results of evaluating Pearson’s correlation for the three devices to compare the envelop extraction.

Fig. 4

Comparison of Minze flow data with the sound envelope of the signal given by UM, Watch, and Phone, for a selection of eight randomly selected signals.

Fig. 5

Diagram showing the pipeline of the proposed methodology, consisting of 4 main steps: data extraction, audios segmentation, feature extraction, and finally model training and validation.

Fig. 6

Evaluation results of the three regression algorithms for each recording device, in terms of the MAE, measured in ml/s. The MAE is the mean of the absolute differences between the predicted and actual flow values, calculated across all labeled audio segments.

Fig. 7

Predictive power (importance) of each frequency component in the flow prediction task from SU. The frequency band selected in our algorithms is shown in blue, showing the highest values of importance. Importance is calculated using the Gini impurity metric with an RF model.

Fig. 8

Analysis of the MAE for the RF prediction model, comparing different audio segment sizes (ms) and the three different recording devices. The MAE is the mean of the absolute differences between the predicted and actual flow values, calculated across all labeled audio segments.

Fig. 9

Analysis of the RF model MAE value for different frequency bands, using the UM microphone. The MAE is the mean of the absolute differences between the predicted and actual flow values, calculated across all labeled audio segments.

Fig. 10

Comparison of the UF (orange) and SF (blue) flow curves. To obtain the SF curves, we selected the Watch audios, and the model selected was a RF with 20 linear-binned FFT as input features, taking segment sizes of 1000 ms.

Fig. 11

Comparison between the VV given by Minze flowmeter and the predicted volume, calculated as the sum of the estimated flows corresponding to each audio signal from the SU tests.