Machine Learning Analysis of Equivocal Renal Scans to Predict Renal Complications
John Weaver, MD1, Joey Logan, MS2, Maria Antony, BS2, Reiley Broms, BS2, Lauren Erdman, PhD3, Mandy Rickard, BSN3, Bernarda Viteri, MD2, Neeta D'Souza, BS2, Armando Lorenzo, MD4, Yong Fan, PhD5, Gregory Tasian, MD2.
1Rainbow Babies and Children's Hospital/Case Western Reserve University, Cleveland, OH, USA, 2Children's Hospital of Philadelphia, Philadelphia, PA, USA, 3Hospital for Sick Children, Toronto, ON, Canada, 4Hospital for Sick Children, Toronto, ON, USA, 5University of Pennsylvania Perelman School of Medicine, Philadelphia, PA, USA.
Introduction:Patients with antenatal hydronephrosis (ANH) are often examined with mercaptuacetyltriglycerine(MAG3) Lasix renal scans when there is suspicion for a ureteropelvic junction obstruction (UPJO). However, human interpretation of MAG3 renal scans is subject to variation and uncertainty. We applied machine learning methods to discover imaging features in equivocal MAG-3 renal scans associated with future renal complications. We hypothesize that quantitative machine learning tools will discriminate children at risk of future renal complications from those who are not.
Methods:We performed a case-control study of children with ANH and concern for a UPJO based on ultrasound findings of Society of Fetal Urology grade 3 or greater, or Urinary Tract Dilation (UTD) Classification grade 2, or greater hydronephrosis without ureteral dilation. Exclusion criteria included history of ipsilateral vesicoureteral reflux or renal anatomic anomalies (e.g., renal malrotation, horseshoe kidneys). The inclusion criterion was an equivocal MAG3 renal scan, defined as scans that did not lead to surgical intervention for a UPJO within 6 months after the scan was performed. Cases were selected based on the occurrence of renal complications that occurred ≥ 6 months after the equivocal MAG3. Complications included: decline of greater than 5% in renal function of the affected kidney on subsequent MAG3 renal scan, increased parenchymal thinning or worsening hydronephrosis on renal ultrasound, new onset flank pain, or occurrence of pyelonephritis. Patients were eligible to be controls if no renal complications occurred for a minimum of 3 years after the renal scan. MAG3 output plots were converted to numerical features using parametric MRI software developed at our institution. We created a classification model using random forests to distinguish cases from controls based on their MAG3 features. The cohort was divided into training (80%) and test (20%) sets; results were internally validated using 3-fold cross validation. The performance of the model was assessed using measures of diagnostic accuracy.
Results:Of the 117 patients eligible for inclusion with available images, 51 patients were cases and 66 patients were controls (Table 1). The model predicted future complications with a sensitivity of 81.8%, a specificity of 61.2%, a positive predictive value of 74%, a negative predictive value of 71.4%. A 2x2 table illustrating the model's performance is shown in Figure 1. Positive and negative likelihood ratios were 2.1 and .3, respectively. Assuming a 43% probability of complications before extracting the MAG3 features, the post-test probabilities of renal complications increased to 62% and decreased to 18.7% when the model did and did not predict a complication, respectively.
Conclusion: Our machine learning model was able to predict patients at high risk of developing renal complications following an equivocal renal scan and also discriminate those who are at low risk.
| Table 1. Comparison of clinical characteristics of the complication and non-complication groups. | ||
| Cases | Controls | |
| Patients (n) | 51 | 66 |
| Side of interest | Right 14, Left 37 | Right 26, Left 40 |
| Male sex, n(%) | Male 38, Female 13 | Male 51, Female 15 |
| Age at time of renal scan in days, median (IQR) | 64 (47,154) | 91.5 (61,200) |
| Time from equivocal MAG3 to pyeloplasty (cases) and to most recent follow up (controls), median (IQR) days | 464 (321,890) | 1682 (1394.75, 2180.25) |
| Median T1/2 (IQR) | 20 (13.75,31) | 11 (5,25) |
| Median split function of affected kidney (IQR) | 48.5 (44.25,51) | 49 (47, 53) |
| Hydronephrosis grade on ultrasound that prompted renal scan (UTD), median (IQR) | 2 (2,3) | 2 (2,3) |
| Hydronephrosis grade on ultrasound that prompted renal scan (UTD), average (STDEV) | 2.43 (0.5) | 2.45 (0.56) |
| Hydronephrosis grade prior to pyeloplasty (cases) and on most recent follow up (controls) (UTD), median (IQR) | 3(3,3) | 2 (1,2) |
| Hydronephrosis grade prior to pyeloplasty (cases) and on most recent follow up (controls) (UTD), average (STDEV) | 2.76 (0.48) | 1.56 (0.93) |
| Random Forest Model Prediction | |||
| Complication | No Complication | ||
| True Outcome | Complication | 54 | 12 |
| No Complication | 19 | 30 | |
| Figure 1. 2x2 confusion matrix depicting model performance | |||
Back to 2022 Abstracts