Summary
A medRxiv preprint reports that three adapted machine-learning models were evaluated on videourodynamic data from an independent cohort of patients with spina bifida. A model combining pressure-volume measurements and fluoroscopic images had the strongest reported results for predicting hydronephrosis and classifying bladder dysfunction.
A medRxiv preprint reports that machine-learning models adapted to an independent cohort of patients with spina bifida could classify bladder dysfunction and estimate the risk of developing hydronephrosis using videourodynamic data. The model combining pressure-volume measurements with fluoroscopic images had the strongest reported results for both tasks.
How the models were evaluated
Videourodynamic studies (VUDS) combine measurements of bladder pressure and volume with fluoroscopic imaging. Clinicians use them to assess bladder function and the risk of deterioration in the upper urinary tract. The preprint notes that interpretation can vary between reviewers.
The researchers adapted three previously developed machine-learning approaches for an independent dataset from one institution, with VUDS data collected between 2016 and 2025. One deep-learning model used pressure-volume data, another analysed fluoroscopic images, and an ensemble combined risk estimates from both. The models were evaluated for prediction of incident hydronephrosis and for classification of bladder dysfunction severity.
The hydronephrosis analysis included 70 patients; 13 (18%) developed the condition. For this outcome, the ensemble model had a concordance index of 0.77, compared with 0.70 for the pressure-volume model and 0.72 for the imaging model. Its overall area under the receiver operating characteristic curve (AUROC) was 0.75, compared with 0.67 and 0.74, respectively. These measures describe how well model scores distinguish or rank outcomes; they are not percentages of patients correctly classified.
For bladder dysfunction classification, the analysis covered 95 VUDS studies. The ensemble model achieved 71% accuracy, compared with 60% for pressure-volume data alone and 66% for imaging alone. The authors also report no substantial disagreements between the ensemble model and expert pediatric urologist reviewers.
What the results indicate
The combined model performed better than either individual data modality on the reported measures in this cohort. That is relevant because VUDS interpretation informs assessment of bladder dysfunction and potential upper urinary tract risk. Combining measurements and images may offer a way to support more consistent interpretation, although the study evaluated model performance rather than a change in clinical care.
The findings come from a single-institution cohort, with 13 incident hydronephrosis cases in the risk analysis. The report is a medRxiv preprint, so its findings have not yet undergone peer review. The results provide an external-cohort evaluation of the adapted models, but the reported sample is modest for assessing how performance may carry across clinical settings.