Various patterns of detrusor overactivity (DO) have been described in the literature, and their clinical relevance has been discussed by the ICI-RS 2018.[1] Previous research stated that the definitions of these patterns were usable after making some assumptions, but that their usability and reproducibility could be improved by objectifying these definitions. Human classification of DO patterns is inherently subjective. An objective approach could be objective parameters implemented in a supervised machine learning (SML). The goal of this study is to explore objective filling phase parameters in a large cohort of urodynamic studies (UDS) with DO, and to implement those parameters into a supervised machine learning model.

This retrospective study in a tertiary hospital included ICS standard UDS[2] performed on adult patients using the Ellipse urodynamics machine with a water-filled system and AUDACT software (Andromeda Medizinische Systeme GmbH, Taufkirchen, Germany). All UDS performed between 2003 and 2008 were anonymized and loaded into MATLAB R2022b (The MathWorks Inc., Natick, MA). A subset of 1001 UDS was selected. All measurements contained vesical pressures (Pves) and abdominal pressures (Pabd) and contained no major artifacts in the filling phase. 
One investigator analyzed the entire set of UDS and excluded all UDS without DO, with poor measurement quality, or with rectal contractions that would hamper parameter calculation. Each remaining DO pattern was visually scored based on the most recent definitions in literature[1], with some small additions by the investigators: 1: Phasic, cited: ‘characteristic waveform and may or may not lead to urinary incontinence’; 2: Terminal, cited: ‘detrusor contraction occurring near or at the maximum of cystometric capacity, which cannot be suppressed, and results in incontinence or even reflex bladder emptying’. We added that a terminal contraction may also result in the end of the filling phase, without incontinence.; 3: Compound, cited: ‘phasic detrusor contraction with a subsequent increase in detrusor and base pressure with each subsequent contraction’; 4: Sustained, cited: ‘continuous detrusor contraction without returning to the detrusor resting pressure’. There could be combined patterns of 1 & 2, 1 & 3, and 1 & 4. In all UDS with DO, the investigator used MATLAB to select each DO segment, which was defined as an elevation of detrusor pressure (Pdet) starting from the moment the pressure deviates from its baseline and ending the moment the pressure returns to its (original) baseline or with the end of the filling phase. 
In all UDS with DO, the following parameters were calculated for each DO segment: duration, amplitude, the area under the curve (AUC), the gradient between the beginning of a DO segment and its amplitude, volume, and time at (the beginning of) a DO segment as absolute values and as a percentage of the total filling phase volume and time (V%, t%). In phasic patterns with multiple DO segments, the volume and time between DO segments were calculated and the abovementioned parameter values were averaged.
The values of the calculated parameters were described for each visually scored DO pattern. Differences in parameter values between DO patterns were tested using a Kruskal-Wallis test in IBM SPSS Version 27 (IBM Corp., Armonk, NY) with p<0.05 being significant.
The parameters described above were included in a SML model, using the naïve Bayes classifier, using Python 3.9.13 with the SKlearn package. The dataset was split using the random split function included in the KSlearn package, in a test and a training dataset, with 25% of the measurements in the test dataset. The SML training was repeated 100 times with different random splits of the dataset to reduce the dependency on a certain split. After each training, the accuracy of the SML was calculated and the mean accuracy for the 100 repetitions was given. In addition, this was repeated  with only a selection of the parameters, to find the minimal amount of parameters needed.

In total 706 of 1001 visually analyzed UDS were excluded either due to the absence of DO, poor measurement quality, or hindering rectal activity. The distribution of the visually scored DO patterns in the 295 included UDS with DO was: 123 (42%) phasic, 76 (26%) terminal, 35 (12%) compound, 14 (5%) sustained and 47 (16%) had a combined pattern. 
Figure 1 shows the calculated parameters for the phasic, terminal, compound, and sustained patterns. All calculated parameters differed significantly between these four DO pattern groups. The boxplots in figure 2 highlight the value distribution in the duration, amplitude, and AUC of the DO segments as well as the percentage of the total filling phase volume at the first DO. 
108 of the 123 phasic patterns had multiple DO segments. The median time and volume between DO segments were 19 s (IQR, 9-37 s) and 26 cmH2O (IQR, 17-41 cmH2O) respectively.
The overall accuracy of the SML using the naïve Bayes classifier  resulted in an average accuracy for the 100 repetitions of 80% (IQR, 76%-84%). Training of the SML using only the duration, t%, and amplitude also resulted in an average accuracy of 80% (IQR, 77%-84%). Other selection of parameters resulted in a lower accuracy of the SML model.

Objective filling phase parameters that significantly differ between the patterns as currently defined can be calculated. Although figure 1 shows that overlap between groups exists, the value distribution seems to differ enough that relatively accurate objective definitions of DO pattern groups could be formulated based on such parameters. SML resulted in an overall accuracy of 80%, which is considered a good performance. Interestingly, using only three parameters, the accuracy of the SML did not change, indicating that the most discriminative power for distinguishing DO subtype is in these parameters. 
The value distribution of these parameters, however, might be user dependent. For patterns that occur at the end of the filling phase (terminal, compound, sustained), parameters like duration, amplitude, and AUC depend on when the urodynamicist decides to end the filling phase. Variability between hospitals thus might exist. 
Furthermore, this study only further analyzed DO patterns that fit a single definition while 16% of the UDS with DO showed two patterns. The parameters in these patterns will thus highly overlap with other groups. Although it might seem fitting in these cases to analyze the measurements based on the second pattern (i.e. terminal, compound, sustained), the phasic element should not be neglected.

Objective filling phase parameters were calculated that significantly differed between visually scored DO patterns based on the definitions discussed by the ICI-RS 2018. Thresholds based on the value distribution of these parameters could be used to objectively define DO patterns in future research, although some parameters might be user dependent and some measurements show a combination of two patterns. An SML model based on three of the parameters achieved an overall accuracy of 80%, which is considered good.

Gajewski JB, Gammie A, Speich J, Kirschner-Hermanns R, De Wachter S, Schurch B, Korstanje C, Valentini F, Rahnama’i MS (2019) Are there different patterns of detrusor overactivity which are clinically relevant? ICI-RS 2018. Neurourol Urodyn 38:S40–S45. https://doi.org/10.1002/nau.23964
Schäfer W, Abrams P, Liao L, Mattiasson A, Pesce F, Spangberg A, Sterling AM, Zinner NR, Van Kerrebroeck P (2002) Good Urodynamic Practices: Uroflowmetry, filling cystometry, and pressure-flow studies. Neurourol Urodyn 21:261–274. https://doi.org/10.1002/nau.10066

Continence 7S1 (2023) 100875
DOI: 10.1016/j.cont.2023.100875