Machine Learning for Urodynamic Detection of Detrusor Overactivity

Abstract

Objective: To develop a machine learning algorithm that identifies detrusor overactivity (DO) in Urodynamic Studies (UDS) in the spina bifida population. UDS plays a key role in assessment of neurogenic bladder in patients with spina bifida. Due to significant variability in individual interpretations of UDS data, there is a need to standardize UDS interpretation.

Materials and methods: Patients who underwent UDS at a single pediatric urology clinic between May 2012 and September 2020 were included. UDS files were analyzed in both time and frequency domains, varying inclusion of vesical, abdominal, and detrusor pressure channels. A machine learning pipeline was constructed using data windowing, dimensionality reduction, and support vector machines. Models were designed to detect clinician identified detrusor overactivity.

Results: Data were extracted from 805 UDS testing files from 546 unique patients. The generated models achieved good performance metrics in detecting DO agreement with the clinician, in both time- and frequency-based approaches. Incorporation of multiple channels and data windowing improved performance. The time-based model with all 3 channels had the highest area under the curve (AUC) (91.9 ± 1.3%; sensitivity: 84.2 ± 3.8%; specificity: 86.4 ± 1.3%). The 3-channel frequency-based model had the highest specificity (AUC: 90.5 ± 1.9%; sensitivity: 68.3 ± 5.3%; specificity: 92.9 ± 1.1%).

Conclusion: We developed a promising proof-of-concept machine learning pipeline that identifies DO in UDS. Machine-learning-based predictive modeling algorithms may be employed to standardize UDS interpretation and could potentially augment shared decision-making and improve patient care.

Figure 1.

Example data. A) Example data from one test/file showing p_ves, p_abd, and p_det signals where p_det is calculated by subtracting p_abd from p_ves. The file also contains numerous events, most of which were placed manually during data collection. Two DO events are evident on the right side as evidenced by two, operator placed, red lines after t=2000s. Note that in some time periods there is poor signal quality (e.g, p_ves between 1000 and 1500s). These artifacts potentially complicate automated analyses. Example p_ves and p_abd signals are shown in panels B-E. The goal of this work is to create an algorithm that, using these signals (and p_det), can reliably identify DO. In B) the p_ves contraction is clearly a DO event. In C), the sharp contractions and simultaneous p_abd activity indicates that the increase in vesical pressure is likely from a cough. In D) there is an isolated increase in p_ves, but it is relatively small, making it less clear whether the increase represents a DO event. In E) there is an increase in p_ves, but also an increase in p_abd, although of lower size, complicating the interpretation. Whether E) qualifies as DO may depend, in part, on the quality of the p_abd signal in matching p_ves in response to external (to the bladder) pressures. The diversity of signal types makes it a challenging problem to reliably and consistently identify DO events.

Figure 2.

Feature extraction. A) Example p_ves trace. B) Data from (A) as a histogram. Five features are extracted from this data including the minimum value, the maximum value, the mean, median, and standard deviation. C) In other models the Fourier transform of the data was computed. The resulting frequency domain data served as an input to the prediction model. D) Another example trace with sparse DO activity. E) Subsets of data from (D). Subsets of 60s in duration were extracted, advancing 20s before the next extraction. F) Normalized histograms from the entire dataset, and a subset that has DO. Reducing the time-period over which the features are computed increases the representation of the DO in the signal. G) Similar to (F), but for the frequency domain.

Figure 3.

Model Performance. A) Receiver operating characteristic (ROC) curve for all 8 models tested. B-I) Confusion matrices for all 8 models with AUC, specificity, sensitivity values, as well as the average ± standard deviation of each group size for TN (true negatives), FP (false positives), FN (false negative) and TP (true positives). All results shown are the average of 20 splits of the data into 85% for training and 15% for testing.

Figure 4.

Examples from the model with windowed data, time features, and 3 channels. Each panel shows model predictions for each subset shown in the panel (bars that span the duration of the subset). The center subsets also list the probability of DO that the model assigned for that subset, an indicator of the model’s confidence in the outcome (closer to 0.5 indicates less confidence). A) Center examples show TPs with high confidence. B) Center examples show TPs with low confidence, presumably due to the relatively low amplitude of the contraction. C) FP error due to coughing with a potential contribution from 1–2 small humps following the cough. D) FP from pressure applied to the abdomen. E) FN presumably due to the high amount of abdominal activity. F) FN where DO occurred just prior to an increase in abdominal pressure (rather than during).