Can unsupervised machine learning gain new insights into urodynamic pressure flow pattern analysis?

Abstract

To explore the use of unsupervised machine learning (UML) to analyse segments of the pressure flow study (PFS) curve after maximum flow, and subsequently to analyse the urodynamic and patient characteristics of men in the detected clusters. In this study, we considered 1650 PFSs of men with lower urinary tract symptoms, without relevant interventions in the past. After datapoint reduction and normalisation of the PFS curve segments, the k-Shape clustering algorithm was used to identify different pattern clusters. Differences in patient and urodynamic characteristics among those clusters were explored. The UML approach identified four prominent clusters, with significantly different patient and urodynamic characteristics. Two pairs of these clusters were visually similar, and included similar urethral resistance values; however, they differed with regard to detrusor voiding contraction (DVC) and prostate size. In two clusters, the PFS curve pattern was significantly different from the commonly assumed 'normal' urethral resistance pattern in elderly men. In males, PFS patterns are considered to be uniform in shape. However, this study shows that UML can help to identify clusters of pressure-flow urethral resistance subtype patterns in men. We found that these subtype patterns were associated with DVC strength and prostate size. This feasibility study has shown that UML clustering of urodynamic PFSs in men holds promise for improving the diagnosis of urethral resistance and DVC properties and dynamics.