Heart Failure Prediction using Longitudinal analysis of ECG Biomarkers

Zuzana Koscova1, Radovan Smisek2, Ivo Viscor3, Pavel Jurak3, Filip Plesinger3
1Department of Biomedical Informatics, Emory University; Institute of Scientific Instruments of the Czech Academy of Sciences, 2Institute of Scientific Instruments of the CAS, v. v. i., 3Institute of Scientific Instruments of the CAS


Abstract

Although the electrocardiogram (ECG) is not a primary diagnostic tool for heart failure (HF), it can provide early markers that are useful for HF risk stratification. Most ECG-based HF prediction studies rely on a single-time-point recording; however, we hypothesize that longitudinal changes across multiple ECGs contain additional predictive information.

In this study, we used data from the MIMIC-IV database and selected patients with at least two ECG recordings. Incident HF was defined based on ICD-9/10 codes with up to 10 years of follow-up. The final cohort included 65,427 patients (HF-prevalence 10%). Patients were split into training and test sets. Using 5-fold-cross-validation on the training set, we developed two random forest models for 3-year HF prediction: (1) a model based on interpretable ECG biomarkers from the most recent ECG (RR interval, P-wave duration, QRS duration, QT interval, QTcB, P-, QRS-, T-wave-axes, ventricular electrical dyssynchrony (VED), and ventricular duration (VD)); and (2) a model that augmented these features with longitudinal changes, defined as the range between minimum and maximum feature values across all available ECGs.

In 5-fold-cross-validation, the combined model (single-time-point + longitudinal ECG features) outperformed the single-time-point ECG model, achieving an AUROC of 0.84 vs. 0.79 and an AUPRC of 0.70 vs. 0.63 for the 3-year prediction task.

On an independent test set with up to 10-year follow-up (13,086 subjects, 10% HF-prevalence), the longitudinal model again improved risk stratification compared to the single-time-point ECG model. In Cox regression analysis, the longitudinal model demonstrated a higher hazard ratio (2.16 vs. 1.39) and improved Harrell's concordance index (0.77 vs. 0.69) after adjusting for age and sex.

SHAP analysis identified the ranges of T-wave axis, VD, P-wave duration, and VED as the most important predictors. Overall, these results demonstrate that incorporating longitudinal ECG information enhances HF risk stratification.