Artificial intelligence-based analysis of uroflowmetry patterns in children: a machine learning perspective

Abstract

Aim Uroflowmetry (UF) is one of the most commonly used noninvasive tests in the evaluation of children with lower urinary tract symptoms (LUTS). However, studies have highlighted a weak agreement among experts interpreting voiding patterns. This study aims to assess the impact of Machine Learning (ML) models, which have become increasingly prevalent in medicine, on the interpretation of voiding patterns.Materials and Methods The study included UF tests of children aged 4-17 years who were referred to our clinic with LUTS. Voiding patterns were independently interpreted by three experts in pediatric urology. Discrepancies in interpretations were jointly re-evaluated by these three observers, and a consensus was reached. Voiding volume (VV), voiding duration (VD), and urine flow rates at 0.5-s intervals were converted into numerical data for analysis. Eighty percent of the data set was used as training data for ML, while the remaining 20% was reserved for testing. A total of five different ML models were employed for classification: Decision Tree, Random Forest, CatBoost, XGBoost, and LightGBM. The models that most accurately identified each voiding pattern were determined.Results We included a total of 500 UF tests in our study, comprising 221 boys (44.2%) and 279 girls (55.8%). The mean age of the children was 9.17 +/- 3.41 years. In the initial assessment, 311 tests (62.2%) were interpreted identically by the observers, while 189 tests (37.8%) were interpreted differently by at least one observer (Fleiss' kappa = 0.608). Of the samples used for ML training, 253 (50.6%) exhibited a bell-shaped pattern, 52 (10.4%) a tower pattern, 103 (20.6%) a staccato pattern, 40 (8%) an interrupted pattern, and 52 (10.4%) a plateau voiding pattern. Among the models tested, the highest accuracy was achieved with XGBoost (85.00% +/- 2.90), while the lowest accuracy was observed with the Decision Tree model (81.80% +/- 1.47). When evaluating voiding patterns individually, the interrupted voiding pattern demonstrated the highest accuracy rates (95%-100%), where as the tower (63.46%-73.08%) and plateau (61.54%-71.15%) patterns had the lowest.Conclusion The current trial demonstrated, for the first time, that ML models achieved an acceptable accuracy rate in interpreting UF patterns in children. Consequently, artificial intelligence (AI) models have the potential to help standardize the analysis of UF voiding patterns in the future.Trial Registration ClinicalTrials.gov (Ref: NCT06814847).