Machine Learning Analysis of Videourodynamics to Predict Incident Hydronephrosis in Patients With Spina Bifida.

Purpose: Variability in the interpretation of videourodynamics studies limits reliable classification of kidney injury risk for patients with spina bifida. We developed machine learning models to predict incident hydronephrosis in patients with spina bifida using videourodynamics data.

Materials and methods: We trained machine learning models using data from videourodynamics studies performed between 2016 and 2022 on patients with spina bifida aged 2 months to 42 years. We evaluated the performance of 4 models to predict incident hydronephrosis following an index videourodynamics study: (1) random survival forest model using data prospectively abstracted from videourodynamics studies by urologists, (2) random survival forest of bladder volume-pressure data, (3) random survival forest using deep learning features extracted from fluoroscopic images of the bladder, (4) ensemble model averaging the probabilities of the volume-pressure and fluoroscopic models.

Results: We included 354 and 200 patients in the training and validation cohorts, respectively. Among the training and validation cohorts, 89 (25.1%) and 71 (35.5%) patients developed incident hydronephrosis at a median time of 1.6 (IQR, 0.5-3) and 2.49 (IQR, 1.72-3.03) years after the index videourodynamics study, respectively. The ensemble model that included data from studies during which ≥ 75% expected bladder capacity was reached had the best discrimination (C statistic 0.73; 95% CI, 0.68-0.76). The specificity of high-risk scores (top 10% in the ensemble model) was 97%.

Conclusions: Automated extraction of features from pressure/volume recordings and fluoroscopic images of the bladder predicted incident hydronephrosis in patients with spina bifida.