Predicting posterior urethral obstruction in boys with lower urinary tract symptoms using deep artificial neural network.

PURPOSE: To assess the prediction model for late-presenting posterior urethral valve (PUV) in boys with lower urinary tract symptoms (LUTS) using artificial neural network (ANN).

MATERIALS AND METHODS: 408 boys aged 3-17 years (median 7.2 years) with LUTS were examined and had bladder diary, ultrasound, uroflowmetry, urine, and urine culture. Cystoscopy was recommended when peak flow rate (Qmax ) was persistently ≤ 5th percentile in patients who were unresponsive to urotherapy and pharmacological treatment (oxybutynin). With four-layered backpropagating deep ANN, the probability of finding PUV was estimated using noninvasive, quantitative parameters (age, Qmax , time to Qmax , voided volume, flow time, voiding time, average flow rate).

RESULTS: There were 97 patients with low Qmax and 74 were unresponsive. In 41, cystoscopy was performed and PUV was diagnosed in 37 (9.1%). In multivariate analysis, significant variables in favor of PUV were urgency (OR = 3.96, 95% CI = 1.30-12.03, p = 0.015), increased voiding frequency (OR = 3.81, 95% CI = 1.03-14.11, p = 0.045), and weak stream/intermittency (OR = 8.30, 95% CI = 2.49-27.63, p = 0.001). The ANN dataset included 87 uroflows of children with PUV and 114 uroflows classified as normal. The best performance was with two hidden layers with four neurons each. The best test accuracy was 92.7% and AUROC was 98.0%. With cutoff value of 0.8, sensitivity was 100.0%, specificity 89.7%, positive predictive value 80.0%, and negative predictive value 100.0%.

CONCLUSIONS: With ANN, we accurately predicted 92.7% of late-presenting PUV using uroflow. Considering the high frequency of PUV in boys with LUTS, especially in cases of urgency, increased voiding frequency, and weak stream or intermittency, accurate prediction could lead to timely treatment.