Ultrasound-based imaging features as predictors of kidney function decline: a machine learning model in the Chronic Kidney Disease in Children (CKiD) study
Bernarda Viteri, MD, MSTR1, Joya Sims, MS2, Joey Logan, MSc1, Hannah Derwick, MPH1, Jennifer Roem, MS3, Susan Furth, MD, PhD1, Gregory Tasian, MD, MSc, MSCE1.
1Children's Hospital of Philadelphia, Philadelphia, PA, USA, 2Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA, USA, 3Johns Hopkins Bloomberg School of Public Health, Baltimore, MD, USA.
BACKGROUND: Prediction models of kidney function decline based on biological specimens are evolving, but few integrate imaging. We previously reported that features extracted from kidney ultrasounds using deep learning improved early prediction of chronic kidney disease progression among boys with posterior urethral valves. We are now applying and calibrating these machine learning-derived features in the prospective Chronic Kidney Disease in Children (CKiD) cohort study.METHODS: We determined the predictive performance of a previously developed machine learning algorithm among 119 participants with non-glomerular chronic kidney disease enrolled at 14 sites in CKiD. The primary outcome was a 35% decline of estimated glomerular filtration rate (eGFR). We assessed the predictive performance of three different models: (1) Clinical model, (2) Imaging model, and (3) Ensemble model. The clinical model included clinical features measured at start of follow-up in CKiD, including eGFR measured by CKiD “U25” formula, systolic blood pressure, urine creatinine, urine protein, age, serum creatinine. The imaging model used features extracted automatically from ultrasound images of the first kidney ultrasound available for each participant. The kidney images were mapped using optical character recognition from burned in labels and images with poor quality were discarded. We used a random survival forest model to estimate eGFR drop over time and assigned a high and low risk group based on predicted risk scores. We then extracted the average pooling layer to ingest it into a random survival forest model. We united the Clinical and Imaging model to create the Ensemble model.
Results: One-hundred nineteen and 91 participants were included in the Clinical model and Imaging model, respectively. The median age of the cohort at time at which ultrasound was obtained was 24.7 days old (IQR 5.4). Estimated GFR decline of 35% or more was reached by 25.2% of the cohort over a median follow-up of 9 years (IQR 6). The primary diagnosis of the cohort was aplastic/dysplastic/hypoplastic kidney disease (45/119, 38%). The models successfully identified “High” vs. “Low” risk populations. Of the three models presented, the Ensemble model had the highest C- index of 0.75, which outperformed the Clinical model (C-index 0.73) and the Imaging model (C-index 0.72). CONCLUSIONS:Our models accurately predicted kidney function decline in children with chronic kidney disease using deep learning imaging features. The Ensemble model had the best performance, highlighting the importance of including imaging in the model. These results in a prospective cohort suggest promise for translation to clinical practice.
Back to 2023 Abstracts