The urodynamic pressure-flow study (PFS) is used to diagnose the properties of the bladder and the outflow tract during voiding. The part of the PFS curve after the maximum flowrate (Qmax) is most clinically relevant, based on the distensible and collapsible tube theory. Based on this model, the passive urethral resistance relation (PURR) was established, describing the expected ideal relation of pressure and flowrate after Qmax. In addition, the dynamic urethral resistance relation (DURR) was defined as the deviation of the measured curve from the PURR.[1] The (clinical) epidemiology of bladder outflow dynamics (PURR versus DURR) is not known. With this in mind, we have used Artificial intelligence (AI) based unsupervised machine learning (UML). AI-UML analysis can be used for the ‘automated’ and objective classification of signal pattern data into clusters with similar properties and is therefore a useful tool for screening a large dataset on pattern similarities. Some urodynamic studies used supervised machine learning (SML) to automatically classify patterns. However, these models are susceptible to human error, as the gold standard is set by expert opinion. AI-UML results in a consistency of analysis that is difficult or impossible to obtain with human-expert evaluation but requires additional steps for implementation of the result in the clinic. Therefore UML is an ideal tool for a first analysis of a large dataset of PFS on still-unknown or undescribed patterns. The aim of this study is to analyze outflow dynamics in a large set of male PFSs, by analyzing, classifying, and clustering the PURR - DURR using UML and to describe the properties of those clusters.

1662 PFS of men (age: mean 59 years (17-93)) with a 2<Qmax<35 mL/s, PFS voided >100mL without major artifacts (e.g., hitting flowmeter-peaks) were included. The detrusor pressure (pdet) and flowrate signals were filtered using a 2-second moving average filter, and a correction for flowrate measurement delay of 0.75 seconds was applied. To allow UML, the flowrate was normalized to a 30-point scale, based on the minimum and maximum flowrate within a PFS, and pdet to a 0-1 scale. Furthermore, the mean pdet around each normalized flowrate point was calculated. UML was applied, based on the K-means learning model, with the dynamic time-warping (DTW) metric. With DTW (or in this case ‘dynamic flowrate warping’ because not the time, but the flowrate was entered as the independent factor), the UML becomes less sensitive to small variations in the p and Q relation, resulting in a more applicable model. UML requires entering the 'requested’ number of resulting clusters and the optimal amount of clusters was determined using the silhouette score. For each cluster, the amount included PFS, basic patient characteristics such as age and prostate size, and PFS parameters as Qmax, the corresponding pressure (pdetQmax), voided volume, and bladder outflow obstruction (BOO) according to linPURR, URA, and BOOI were compared using the Kruskal-Wallis Test.

Based on the silhouette score, classification into four clusters of PFS -DURR types was found to be optimal. The UML-resulting four clusters can be found in figure 1. The largest cluster consists of 1084 PFS (65%), with high pressure at a high flowrate and low pressure at a low flowrate. The second cluster consist of 264 PFS (16%), with a temporary increase in pressure when flowrate decreased, but ending with low pressure at low flowrate. The third cluster consists of 210 PFS (13%), with an initial decrease of pressure, but an end-voiding increase of pressure. The last cluster consists of 104 PFS (6%), with an overall increase in pressure with a decrease in flowrate.
The distribution of age, Qmax, pdetQmax, TRUS, linPURR, URA, and BOOI was significantly different across the four clusters (p<0.006). No significant difference was found for the voided volume (p = 0.171).

This study showed that the PFS pattern could be divided into four clusters by using AI-UML. As significant differences were found in the patient and urodynamic characteristics, the clustering likely resulted in clinically relevant patient categories. Additional classification of BOO has been performed in the past by the CHESS two-point PFS classification. However, the CHESS classification expects a positive relation between pressure and flowrate, as described by PURR, which is only fully true for (the largest) cluster 1. AI-UML ‘discovered’ 3 other clusters probably or potentially clinically relevant DURR -subtypes, not described earlier.

Based on unsupervised machine learning performed on an extensive database of PFS, we determined 4 types of pressure-flow PURR-DURR patterns, associated with different patient and PFS characteristics.

Schäfer, W.  The Contribution of the Bladder Outlet to the Relation Between Pressure and Flow Rate During Micturition. In: Hinman, F., Boyarsky, S. (eds) Benign Prostatic Hypertrophy. Springer, New York, NY. 1983; https://doi.org/10.1007/978-1-4612-5476-8_44.

Continence 7S1 (2023) 100870
DOI: 10.1016/j.cont.2023.100870