ECG-based deep learning for chronic kidney disease detection and cardiovascular risk prediction

Comprehensive Summary

While chronic kidney disease (CKD) is routinely assessed using glomerular filtration rates (eGFR) and albuminuria, abnormal heart rhythms are also a common symptom found in patients with or at risk for CKD. In this study, researchers developed a Deep Learning Model (DLM) based on data from a standard 12-lead ECG to predict CKD. ECG and eGFR data from a 66,578 patient cohort was used to train, tune, and internally validate the DLM, while 10,476 patients from a separate cohort were used for external validation. Both an eXtreme Gradient Boosting (XGB) and an elastic net model were trained to estimate CDK stage, and their performance was compared to the DLM. The DLM displayed a strong ability to detect CDK with an AUC of 0.885 for the internal validation and 0.861 for the external validation. While the internal validation was outperformed by XGB and the elastic net models when age and sex were included, the external validation outperformed both models even with those covariates. Additionally, the DLM outperformed both models in predicting eGFR based on ECG, and it flagged patients who initially presented with normal eGFR (eGFR ≥ 60 mL/min/1.73 m²) during CDK screening but had elevated risk of developing it.

Outcomes and Implications

Cardiovascular disease is a common complication and cause for mortality for patients with CDK. However, prior research has shown only weak associations between CDK and ECG, likely due to complex relationships and patterns that go undetected by modern statistical and assessment methods. The ability of a DLM to recognize these relationships may be a vital tool for providers to recognize early signs of or increased risk for CDK. Elevated-risk patients with negative eGFR tests were flagged by the DLM, suggesting that heart complications may present earlier than impaired renal function. However, further study in diverse populations and DLM interpretability is needed before implementation of this technology into clinical practice.