Abstract: TH-PO009
Artificial Intelligence (AI)-Driven Screening for Undiscovered CKD
Session Information
- Augmented Intelligence for Prediction and Image Analysis
October 24, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Salazer, Thomas L., Hackensack Meridian Health, Edison, New Jersey, United States
- Sheth, Naitik, Hackensack Meridian Health, Edison, New Jersey, United States
- Masud, Avais, Hackensack Meridian Health, Edison, New Jersey, United States
- Serur, David, Hackensack Meridian Health, Edison, New Jersey, United States
- Hidalgo, Guillermo, Hackensack Meridian Health, Edison, New Jersey, United States
- Aqeel, Iram, Hackensack Meridian Health, Edison, New Jersey, United States
- Adilova, Linara, Carenostics, Philadelphia, Pennsylvania, United States
- Kamp, Michael, Carenostics, Philadelphia, Pennsylvania, United States
- Fitzpatrick, Tim, Carenostics, Philadelphia, Pennsylvania, United States
- Krishnan, Sriram, Carenostics, Philadelphia, Pennsylvania, United States
- Rao, Kanishka, Carenostics, Philadelphia, Pennsylvania, United States
- Rao, Bharat, Carenostics, Philadelphia, Pennsylvania, United States
Background
We developed and validated a predictive model to identify patients at risk of undiagnosed CKD. Given the progressive nature of CKD, our model incorporates a crucial temporal dimension, allowing it to assess whether undiagnosed patients have Stage 3 CKD based on prior EHR data. We developed the model to enable proactive evaluation and intervention on patients.
Methods
Patients aged 18 to 85 with no previous diagnosis of CKD as of January 1, 2018, were considered to identify which patients are at risk of undiagnosed CKD. The data set included patients with no prior eGFR, or at most one abnormal eGFR. We extracted 237 variables from the EHR, including demographics, labs, medications, and prior diagnoses. Multiple machine learning models were trained on a training set, and performance was measured on the test set using ROC curves, sensitivity, specificity and accuracy, as well as temporal stability.
Results
A Random Forest model trained on 46,775 patients achieved an AUC of .95 to predict whether next eGFR measured would be abnormal (eGFR < 60) on an independent test set of 23,039 patients whose eGFR was measured in the subsequent year. A selected set point achieved a sensitivity of 0.51 and PPV of 0.84. (ML models trained to predict the first abnormal eGFR in subsequent 2/3/4 years, achieved AUCs of 0.94/0.93/0.92 demonstrating the temporal stability of our approach.)
Conclusion
The model developed using a random forest machine learning algorithm was able to identify undiagnosed patients at risk of Stage 3 CKD with high accuracy and discriminatory power. This predictive model demonstrates the potential of an AI-driven approach to identify patients earlier in the disease process. In the next phase of this work, we intend to apply this model to identify high risk patients, and proactively recommend them for evaluation and targeted interventions.
Random Forest model ROC Curve