Abstract: SA-PO362
Ultrasound-Based Machine Learning Model and Renal Parenchymal Area as Predictors of Kidney Function Decline in Children with CKD
Session Information
- Pediatric Nephrology - III
November 04, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Pediatric Nephrology
- 1900 Pediatric Nephrology
Authors
- Viteri, Bernarda, The Children's Hospital of Philadelphia, Philadelphia, Pennsylvania, United States
- Furth, Susan L., The Children's Hospital of Philadelphia, Philadelphia, Pennsylvania, United States
- Sims, Joya M., The Children's Hospital of Philadelphia, Philadelphia, Pennsylvania, United States
- Derwick, Hannah C., The Children's Hospital of Philadelphia, Philadelphia, Pennsylvania, United States
- Fischer, Katherine M., The Children's Hospital of Philadelphia, Philadelphia, Pennsylvania, United States
- Roem, Jennifer, Johns Hopkins University, Baltimore, Maryland, United States
- Logan, Joseph R., The Children's Hospital of Philadelphia, Philadelphia, Pennsylvania, United States
- Tasian, Gregory Edward, The Children's Hospital of Philadelphia, Philadelphia, Pennsylvania, United States
Background
Few prediction models of kidney function decline include imaging features. Here, we apply machine learning-derived renal ultrasound features and compare the performance to renal parenchymal area (RPA) in the Chronic Kidney Disease in Children (CKiD) cohort study.
Methods
We determined the predictive performance of a previously developed machine learning algorithm among 119 subjects with non-glomerular chronic kidney disease (CKD) enrolled at 14 sites in the prospective CKiD study. The primary outcome was CKD progression defined as initiation of renal replacement therapy or 50% decline of estimated glomerular filtration rate (eGFR) assessed by the CKiD U25 equation. We assessed the predictive performance of three different models: (1) Clinical model (CM), (2) Deep Learning model (DLM), and (3) Deep Learning +Clinical model (DL+CM). The CM included baseline clinical features: age, eGFR, systolic blood pressure, urine creatinine, urine protein, and serum creatinine. The DLM used features extracted automatically from the first available kidney ultrasound. In a subset of 46 subjects, we compared the performance of ensembled models (3) DL+CM to (4) RPA +Clinical model (RPA +CM). The RPA model included age and the mean RPA values measured by an attending urologist and nephrologist blinded to outcomes. We used a random survival forest model to estimate CKD progression.
Results
Ninety-one subjects with a median age of 11.5 months, IQR [0.2, 78.2] at ultrasound were included in the DL+CM. CKD progression occurred for 20% over a median follow-up of 9 yrs, IQR [3.8,9.7]. For the subset of 46 subjects, median age at ultrasound was 9.7 months, IQR [1.4, 79.3] and CKD progression occurred in 28% over a median follow-up of 6.4 yrs IQR[4.1, 9.2]. Of the models presented, the DL+CM and RPA+CM had the best performance with C-index of 0.74 and 0.77, respectively.
Conclusion
Our ensemble models accurately predicted CKD progression using imaging features in children with CKD. The comparison of the performance of deep learning features and RPA was limited by the small sample size. These results in a prospectively enrolled cohort suggest that early ultrasound imaging features could identify children at greatest risk of CKD progression in clinical practice.
Funding
- NIDDK Support