SPU Main Site  |  Past and Future Meetings
Society For Pediatric Urology

Back to 2019 Abstracts


Machine-learning Algorithm To Identify Urodynamic Detrusor Overactivities
Hsin-Hsiao Scott Wang, MD, MPH, MBAn1, Dylan Cahill, BA1, Caleb P. Nelson, MD, MPH1, Hau-Tieng Wu, MD, PhD2, Carlos Estrada, MD, MBA1.
1Boston Children's Hospital, Boston, MA, USA, 2Department of Mathematics, Duke University, Durham, NC, USA.

Backgrounds: Overactive contractions (OC) of the bladder is a finding on urodynamic studies (UDS) that often correlates with lower urinary tract symptoms and drives management. However, UDS interpretation remains non-standardized and prone to inter-observer variability. We sought to develop a statistical model and apply machine learning algorithms to reliably identify OC in UDS.
Methods: We utilized UDS archive files for studies performed at our institution between 2013 to 2019. Raw tracings of vesical pressure (Pves), abdominal pressure (Pabd), detrusor pressure (Pves), infused volume (Vinf), and all annotations during UDS were obtained. Patients less than 1 year old, studies with calibration issues, or those with significant artifact were excluded. All signals were normalized to 10 minutes and sampled at the 10Hz sampling rate and detrended by the smoothed median filter of length 60 seconds. In the training set, 5 representative OC patterns were clustered using K-means (Figure 1). Dynamic Time Warping algorithm was used to match candidate Pdet signal segments to representative OC patterns considering temporal deviations. A probability of OC was then assigned to the candidate Pdet segments. A 5-fold cross validation (CV) was used to evaluate the model performance.
Results: 800 UDS were included. Median age was 9 years (range 1-33). There were 1764 OC events that did not overlap with annotated artifacts (cough, cry, valsalva, movements). The AUC of the training sets from the 5-fold CV was 0.840.01. The 5-fold CV leads to an overall accuracy 81.27%, and sensitivity and specificity of detecting OC events are 77.77% and 81.31% respectively in the testing set.
Conclusions: Our predictive model using machine learning algorithms provides promising performance to facilitate automated identification of OC in UDS. This would allow for standardization and potentially more reliable UDS interpretation. Signal processing and machine learning interpretation of the other components of UDS are forthcoming.



Back to 2019 Abstracts