Abstract: FR-PO939
Stratifying Patient Risk for eGFR Decline Over a Year
Session Information
- CKD: Epidemiology, Risk Factors, Prevention - II
November 04, 2022 | Location: Exhibit Hall, Orange County Convention Center‚ West Building
Abstract Time: 10:00 AM - 12:00 PM
Category: CKD (Non-Dialysis)
- 2201 CKD (Non-Dialysis): Epidemiology‚ Risk Factors‚ and Prevention
Authors
- Mahmoodzadeh, Zahra, University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States
- Norris, Keith C., University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States
- Tuttle, Katherine R., Providence St Joseph Health, Spokane, Washington, United States
- Bui, Alex, University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States
- Nicholas, Susanne B., University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States
Group or Team Name
- CURE-CKD Registry Team
Background
Electronic health records (EHR) data enables assessment of patient-level risk by advanced data-driven artificial intelligence. This study used curated EHR data to stratify patients’ risk of eGFR decline over a 1-year prediction period. We tested a 2-stage model that first predicted the risk of unstable eGFR (decline >5 mL/min/1.73 m2) in the next year. For an unstable patient, a second model estimated the patient's expected eGFR in the next year.
Methods
We designed a 2-stage modeling technique with a binary classifier sequentially connected to a regressor, using the CURE-CKD data of 857931 patients. The binary classifier predicts the chance of having a stable eGFR (decline <5 mL/min/1.73 m2) over the next year, given a set of static (sex; race; rural-urban commuting area codes; hypertension, type 2 diabetes (DM), pre-DM, and CKD based on ICD coding) and temporal observations (past 2 years of eGFR, medications (ACEI/ARB, SGLT2 inhibitor, GLP1 RA, MRA, NSAID, PPI) and age. If the likelihood of remaining stable is <50%, the patient’s information is passed to a second model trained to estimate the eGFR in the next year. The first model is a deep neural network consisting of a convolutional neural network (CNN-processes longitudinal patient data; trained on the whole training set) concatenated to a feedforward neural network (FNN-processes static patient data). The second stage model is an extreme gradient boosting regressor, trained on a cohort of the training set with eGFR decline >5 mL/min/1.73 m2 in a 1-year prediction window. Both models were trained, validated, and tested on a 60/20/20 split.
Results
The first model’s area under the receiver operating curve was 0.814. When the classification threshold was set to 50%, the classifier accuracy reached 76%, and the precision reached 70%. The second model achieved the mean absolute error (MAE) of 3.47 and root mean squared error of 4.59, while the best stand-alone regressor trained on the whole training set had MAE of 5.73. Therefore, the 2-stage modeling technique increased the eGFR prediction accuracy for more critical patients.
Conclusion
Our 2-step model stratified patient risk for eGFR decline over a 1-year period. The model can support clinical decision-making on the risk of >5 mL/min/1.73 m2 eGFR decline in the next year. If the chance is high, the second model can estimate the eGFR value for the next year.
Funding
- Other NIH Support