Presentation Authors: Edwin Jonathan Aslim*, Balamurali B K, Yun Shu Lynn Ng, Tricia Li Chuen Kuo, Jacob Shihang Chen, Jer-ming Chen, Lay Guat Ng, Singapore, Singapore
Introduction: The current standard of uroflowmetry is equipment-intensive and needs to be performed on site. We attempt to correlate the audio recordings of urinary flows with standard uroflowmetry, with the aim to create a novel audio-uroflow app for use with a smart phone. This study is IRB approved (CIRB 2017/2241) and supported by the SingHealth Surgery ACP-SUTD Technology and Design Multi-Disciplinary Development Programme (grant No. TDMD-2016-1).
Methods: This study prospectively enrolled 25 healthy male volunteers without lower urinary tract symptoms (LUTS), aged 21 to 50 years, from 01 June 2017 to 31 October 2017. Participants were asked to void into a digital standard uroflowmetry machine (MMS version 9.1z, LABORIE, Mississauga , Canada) with a minimum voided volume of 100ml, and urinary flow sounds were simultaneously recorded using a smartphone. Audio recordings were digitally pre-processed to remove background noise, and then paired with the corresponding uroflowmetry readings (UF) to train a machine-learning (ML) algorithm to understand the relationship between them. 70% of the voiding sessions were used to train the ML algorithm, and the remaining 30% sessions were used for testing. The predicted audio-uroflowmetry readings (AF) derived from the acoustic patterns learned by the ML algorithm (not calculated from flow rate vs time plots) were compared against UF parameters such as maximum flow rates (Qmax) and voided volumes (VV). The comparison was done by visually analysing the scatter plot and calculating Pearsonâ€™s correlation coefficient (r).
Results: There were 52 paired uroflow readings and audio recordings, of which 35 datasets were used to train the ML algorithm, and 17 datasets for AF prediction. In the training datasets, the median Qmax and VV were 27.5ml/sec (range 10.7 to 40.1) and 326ml (range 155 to 723), respectively. The trained model was evaluated by comparing UF and AF readings. The median Qmax corresponding to the test UF and predicted AF were 25.6ml/sec (range 8.6 to 39.7) and 27.0ml/sec (range 15.0 to 29.0), with an r value of 0.70. The median VV corresponding to the test UF and predicted AF were 419ml (range 138 to 791) and 360ml (range 242 to 707), with an r value of 0.83. The scatterplots for VV and Qmax, between UF and AF, showed good correlations.
Conclusions: There is good correlation between AI-assisted audio-uroflow predictions with uroflowmetry parameters. Work is ongoing to train the machine-learning algorithm on a larger sample of men with LUTS to improve its prediction capability.
Source of Funding: SingHealth Surgery ACP-SUTD Technology and Design Multi-Disciplinary Development Programme (grant No. TDMD-2016-1)