Purpose: Uroflowmetry is a key diagnostic tool for assessing bladder bowel dysfunction in children, with voiding curve shape being the most critical parameter. However, subjective interpretation leads to low inter-observer agreement. This study evaluates the potential of artificial intelligence (AI) and machine learning (ML) to objectively classify uroflowmetry curves, aiming to reduce variability and enhance diagnostic accuracy.
Methods: This cross-sectional study analyzed 586 uroflowmetry curves from children aged 5–17 years, excluding tests with voided volumes below 50% of expected bladder capacity. Curves were standardized per ICS recommendations (1 mm = 1 s on x-axis, 1 ml/s on y-axis) and classified by three pediatric urology specialists into bell, tower, plateau, staccato, or interrupted patterns per ICCS definitions. The YOLOv5x6 algorithm was trained on 85% of the dataset, with 15% for validation, using a high-performance system. Performance was assessed via accuracy, precision, recall, F1-score, and mean Average Precision (mAP).
Results: Inter-rater agreement was high (Fleiss’ kappa: 0.948 ± 0.007). The AI model achieved 85.8% accuracy, with 96% success in identifying bell-shaped curves. Plateau curves showed the highest precision (1.00), while staccato had the lowest (0.64). mAP@0.5 reached ~90%, stabilizing after 50 epochs.
Conclusions: AI-driven classification of uroflowmetry curves offers high accuracy and reduces observer variability. Future work should focus on multi-center datasets and standardized reporting to enhance clinical utility and integration into uroflowmetry devices for real-time analysis.
Keywords: artificial intelligence, uroflowmetry, voiding dysfunction
Anahtar Kelimeler: