Prognostic evaluation of boys with posterior urethral valves using a recursive feature elimination machine learning...

To improve prognostic prediction and identify key predictors of renal and bladder outcomes in boys with posterior urethral valves (PUV) using machine learning (ML).

Single center study of boys with PUV. A bespoke ML framework incorporating recursive feature elimination, supervised learning models, and 5-fold cross validation was developed for prognostic prediction and identification of key features.

Patients diagnosed after one year of age (n = 22, late diagnosis group) had normal renal function at last follow up. Renal outcome analysis was therefore focused on the 83 patients diagnosed before one year of age (early diagnosis group), of whom 22% (18/83) developed chronic kidney disease (CKD) stage 3-5. Eighty of 105 patients were explored in the bladder outcome analysis. Nadir creatinine was the strongest prognostic feature for renal function. Among features known shortly after diagnosis, the most important were baseline creatinine followed by renal dysplasia. Among tested ML models, random forest classification performed the strongest, both at 1 year of age (0.96 PR-AUC) and shortly after diagnosis (0.73 PR-AUC). Across all features combinations, unregularized logistic regression consistently demonstrated the lowest predictive performance. Bladder dysfunction was observed in 69% (55/80) of boys. However, using our ML framework, the available features did not allow reliable prediction of this outcome (around 0.82 PR-AUC against the 0.69 baseline).

This study demonstrates that a recursive feature elimination ML framework can reliably identify the most important predictors and predict long-term renal function in boys with PUV. Baseline creatinine and renal dysplasia were identified as strong predictors and can inform parental guidance shortly after diagnosis. As expected, nadir creatinine was identified as the most powerful predictor. This study functions as a proof of concept demonstrating how ML methods can provide robust analysis of small datasets, an approach that may be valuable to other rare disease-research settings. Importantly, all models were developed using data from a single-center cohort and should undergo external validation to demonstrate generalizability.