Abstract
BackgroundRoutine assessments of lower urinary tract symptoms (LUTS) include standard uroflowmetry (UF), which is labour and equipment intensive to perform, and stressful and unnatural for patients. An ideal test should be accurate, repeatable, affordable and portable.ObjectiveTo evaluate the accuracy of a machine-learning (ML) augmented audio-uroflowmetry (AF) algorithm in predicting urinary flows.Subjects and methodsThis pilot study enrolled 25 healthy men without LUTS, who were asked to void into a gravimetric uroflowmeter. A smartphone recorded the voiding sounds simultaneously. Paired uroflow and audio parameters were used to train an ensemble ML model to predict urinary flows from voiding sounds. Pearson’s correlation coefficient was used to compare UF with AF values. Statistical significance was defined as p<0.05.ResultsA total of 52 voiding session were captured, of which n=35 were used for training and n=17 for testing the algorithm. Each voiding session was divided into 0.1 s frames, resulting in >300 analysable datapoints per session. Pearson’s coefficients showed strong correlations for flowtimes (r=0.96, p<0.0001), voided volumes (r=0.83, p<0.0001) and average flowrates (r=0.70, p=0.0019), and moderate correlation for maximal flowrate (r=0.69, p=0.0022). AF predicted flow patterns showed good agreement with UF tracings. The main limitations were the small participants sample size and use of a single smartphone type.ConclusionsML augmented AF can predict uroflow parameters with a good accuracy, and can be a viable alternative to standard UF. Further work is needed to develop this platform for use in real-life conditions and across genders.