Risk stratification using QT-based ECG biomarkers requires accurate heart-rate adjustment. Conventional correction formulas (Bazett and Fridericia) assume a memoryless QT–RR relationship and ignore temporal adaptation. This limitation is pronounced in high-risk groups (post-myocardial infarction (MI)), where abnormal repolarization responses to stressors are clinically informative.
We introduce a system-identification framework modeling QT as output of a nonlinear dynamical system driven by current and past RR intervals, enabling personalized QT adaptation dynamics for generalizable dynamical ECG biomarkers.
RR and QT intervals were extracted from Holter ECGs of 418 post-MI and control participants in the MIMS2 cohort during 699 mental and physical stress test recordings. We trained a recurrent neural network (LSTM) to predict beat-wise QT intervals based on both past and current RR intervals. The global model was initially trained on mental-stress data, then fine-tuned for each subject, and tested on each subject's unseen physical-stress data; the opposite setup was also evaluated. Accuracy was assessed using mean absolute error (MAE) and correlation coefficient (CC) over memory lengths of 5–120 seconds. QT dynamics exhibit finite memory, with an optimal 60s history window yielding an MAE of 12.8±3.4ms and a CC of 0.80±0.11; performance plateaus beyond this.
Then, trained models simulated QT responses to controlled RR sequences mimicking a stress test, producing normalized QT–RR hysteresis loops. Exercise- and recovery-phase QT values on a common RR grid (400:50:1200ms) defined subject-specific features for MI classification, compared to nine classical QT-correction features. XGBoost, SVM, and Logistic-Regression classifiers were trained on 309 participants and tested on 104 unseen participants. Dynamic QT–RR features improved MI classification over classical QTc, increasing training AUROC from 65.0±10.4% to 66.6±8.7% and testing AUROC from 63.1% to 65.7%.
These findings demonstrate that subject-specific finite-memory models capture QT–RR dynamics, providing dynamic biomarkers that outperform traditional MI classification and underscore the significance of repolarization dynamics.