Ambulatory Holter ECG monitoring provides parameters that may support risk stratification for coronary heart disease (CHD). Several ECG-derived biomarkers have been proposed for CHD risk assessment; however, they are influenced by physical activity, which is often not explicitly considered. Accelerometer (ACC) data, commonly collected alongside Holter ECG recordings, can provide information on whether an individual is active or at rest, while the electrophysiological response to physical activity may reflect underlying cardiovascular function. In this study, we perform CHD classification from long-term ECG and ACC recordings to identify activity-aware ECG biomarkers and assess their utility for informing future approaches to CHD risk prediction.
We analyzed one week of three-channel Holter ECG and simultaneous triaxial ACC from 443 veteran twins (CHD class prevalence 22%). CHD included myocardial infarction, coronary artery bypass grafting, and percutaneous coronary angioplasty. Features extracted in 5-minute segments included heart rate variability, cardiac repolarization, and activity counts. An activity-aware convolutional neural network was used to classify active and resting states from ACC data. We then evaluated four Random Forest models using different feature aggregation strategies: global mean, activity-stratified features, and combined active–resting representations.
Using 5-fold cross-validation, activity-stratified feature aggregation improved performance over global aggregation. The activity-stratified model achieved a mean (SD) AUROC of 0.80 (0.07) and AUPRC of 0.56 (0.13), outperforming the overall mean aggregation model (AUROC 0.72 (0.06), AUPRC 0.43 (0.10)). SHAP analysis identified cardiac repolarization features — including QT interval (resting only), ST-segment level, and T-wave alternans (resting and active) estimated using modified moving average and a Markov-model-based state-transition approach — as the strongest CHD classifiers.
These results suggest that ECG biomarkers combined with activity-aware modeling could potentially improve future approaches for CHD risk prediction in long-term wearable data.