JU INSIGHT Deep Learning of Videourodynamics to Classify Bladder Dysfunction Severity in Patients With Spina Bifida

By: John K. Weaver, MD, MSTR, Rainbow Babies and Children’s Hospital/Case Western Reserve University School of Medicine, Cleveland, Ohio; Madalyne Martin-Olenski, BA, The Children’s Hospital of Philadelphia, Pennsylvania; Joseph Logan, BA, MS, Translational Research Informatics Group, The Children’s Hospital of Philadelphia, Pennsylvania; Reiley Broms, BA, The Children’s Hospital of Philadelphia, Pennsylvania; Maria Antony, BA, The Children’s Hospital of Philadelphia, Pennsylvania; Jason Van Batavia, MD, MSTR, The Children’s Hospital of Philadelphia, Pennsylvania; Dana A. Weiss, MD, The Children’s Hospital of Philadelphia, Pennsylvania; Christopher J. Long, MD, The Children’s Hospital of Philadelphia, Pennsylvania; Ariana L. Smith, MD, Perelman School of Medicine, University of Pennsylvania, Philadelphia; Stephen A. Zderic, MD, The Children’s Hospital of Philadelphia, Pennsylvania; Jing Huang, PhD, Center for Pediatric Clinical Effectiveness, The Children’s Hospital of Philadelphia, Pennsylvania; Perelman School of Medicine, University of Pennsylvania, Philadelphia; Yong Fan, PhD, Perelman School of Medicine, University of Pennsylvania, Philadelphia; Gregory E. Tasian, MD, MSc, MSCE, Center for Pediatric Clinical Effectiveness, The Children’s Hospital of Philadelphia, Pennsylvania Perelman School of Medicine, University of Pennsylvania, Philadelphia | Posted on: 18 May 2023

Weaver JK, Martin-Olenski M, Logan J, et al. Deep learning of videourodynamics to classify bladder dysfunction severity in patients with spina bifida. J Urol. 2023;209(5):994-1003.

Study Need and Importance

Videourodynamics (VUDS) is the gold standard for the evaluation of the lower urinary tract in patients with spina bifida to characterize bladder function. The interpretation of these data is important to determine whether a bladder has safe storage and emptying functions and whether bladder dysfunction could contribute to the loss of kidney function. VUDS data are rich in detail but laborious for an individual to fully understand, and their interpretation has high interobserver variability. These characteristics create challenges for the utility of VUDS in the longitudinal evaluation of children with spina bifida.

What We Found

In this study, deep learning models that automatically extracted features from pressure and volume tracings and/or fluoroscopic images from VUDS studies classified severity of bladder dysfunction with moderate accuracy (see Figure). The highest performance was observed in models that included longitudinal volume and pressure data and fluoroscopic images. The best performing model was the 75% estimated bladder capacity ensemble model, which had an overall accuracy of 70%. Model performance was directly related to the degree of bladder filling; the accuracy of the classification increased and the discordance between categories decreased as the percentage of estimated bladder capacity achieved increased.

Figure. Schematic of the pressure-volume model (1D convolutional neural network) and the imaging model (VGG-16 pretrained convolutional neural network). Averaging the risk probabilities of these respective models resulted in the ensemble model.

Limitations

We used the expert reviewers’ majority rating as the ground truth by which our models were assessed. Although there is known significant interrater reliability among pediatric urologists, a majority rating by expert reviewers was deemed to be the most scientifically rigorous way to establish ground truth for training and testing our models. Additionally, overfitting and poor generalizability are two known limitations of machine learning models built from small, single institution cohorts.

Interpretation for Patient Care

Our deep learning models were able to automatically classify bladder dysfunction severity. Retrospective and prospective studies performed at other institutions are needed to validate our models, but in the future, these models could influence clinical decision-making by extracting informative features from VUDS studies that would not otherwise be available to clinicians.