Deep Learning Algorithms to Predict Differential Renal Function <40% in Unilateral Hydronephrosis Based on Key Parameters of Urinary Tract Ultrasound

Abstract

OBJECTIVE

To identify the correlation between ultrasound findings and the incidence of differential renal function (DRF) <40%, we conducted an analysis of the key parameters of urinary tract ultrasound in children with unilateral hydronephrosis. For children with unilateral hydronephrosis, DRF <40% serves as a compelling indication for surgical intervention, and it can be assessed through diuretic renogram. However, a significant number of patients do not have convenient access to high-quality renograms. So we conducted this analysis aiming to identifying value of urinary tract ultrasound in surgical intervention decision.

METHODS

We retrospectively reviewed the dynamic renogram and urinary tract ultrasound data of 802 children with hydronephrosis presented to our department. Independent risk factors related to DRF <40% were screened out. Several machine learning models were employed. The area under receiving operating curves (AUROC) was calculated for each model to compare their efficiency.

RESULTS

The renal pelvis anterior-posterior diameter, upper calyx dilation, and renal length ratio were found to be independent risk factors for DRF <40%. Among these factors, the renal length ratio had the highest AUROC of 0.820. These 3 factors, alone with the patients’ age, were then introduced into 3 machine learning models: random forest, logistic regression, and support vector machines (SVM), among which, the SVM exhibited the highest AUROC of 0.941, with a sensitivity of 90.32% and a specificity of 81.03%.

CONCLUSION

The length ratio exhibited the strongest correlation with DRF <40% among multiple ultrasound indices. Moreover, the SVM model exhibited significant improvement compared to individual factors. Therefore, it can be employed to identify high-risk children with impaired renal function in the assessment of congenital hydronephrosis.