Introduction & Objective :
Patients with congenital ureteropelvic junction obstruction (UPJO) are at risk of renal function deterioration. While diuresis renography (DR) provides useful metrics by which pyeloplasty can be recommended, such as differential renal function and drainage half-time, we have previously reported improved accuracy for the prediction of the decision for pyeloplasty using machine learning. Here, we sought to validate this machine-learning model in a larger patient population.
Methods : A retrospective review of patients who underwent DR for suspected UPJO at our institution between 2009 and 2015 was performed. Patients who had reached the endpoint of pyeloplasty or discharged from urologic care were included. Eighteen patients were added to the previous cohort of 59 patients and the machine-learning model was retrained on the larger population of 77 patients. Six optimal features of the DR curves were selected to maximize the area under the receiver operator curve; these differed from the five optimal features used in the previously reported model. Both the six feature model and five feature model analyzed the 77 patient cohort to predict which patients would proceed to pyeloplasty. T1/2 of 20 and 30 minutes were used to analyze the cohort and were compared to the machine learning model.
The retrained six feature model in the current study population had an accuracy of 92% (91% sensitivity, 94% specificity) and the five feature model had an accuracy of 87% (100% sensitivity and 68% specificity). Both the five and new six feature machine learning models were better predictors of requiring pyeloplasty than the conventionally used measure of T1/2 of 20 minutes (76% accuracy) and T1/2 of 30 minutes (75% accuracy; Table 1).
The machine learning model improved the diagnostic accuracy of DR and outperformed T1/2. Furthermore, machine learning models may be better suited to analyze the dynamics of DR rather than the fixed variable of T1/2. While further study is warranted, machine learning could lead to earlier detection of severe UPJO and may reduce the number of DR performed prior to surgical management.