Comparative Analysis of Deep Learning Models for Myocardial Ischemia Classification Using Ambulatory Electrocardiograms

Alarm fatigue is a significant issue in healthcare where clinicians become desensitized to an alert from a monitored patient, primarily due to the high frequency of false alarms. Importantly, electrocardiogram (ECG) changes indicative of myocardial ischemia (M-Isch) are not reliably detected by current algorithms. The clinical use of deep learning (DL) is becoming more prevalent, with a growing emphasis on detecting subtle changes in various biosignals that may be otherwise overlooked due to human error. Here, we investigate the robustness of various DL architectures for the detection of M-Isch. We hypothesized that DL models will be able to achieve a high level of specificity and sensitivity (> 90%) when classifying ECG recordings indicative of M-Isch and normal sinus rhythm. This study leverages the advanced capabilities of timeseriesAI (tsai), a library of DL frameworks specializing in time series decoding, applied for the first time to M-Isch classification. Additionally, the models were trained on the European ST-T database (available from PhysioNet), which consists of ECGs annotated by cardiologists for various ST-T changes associated with M-Isch. The dataset also utilizes ambulatory ECGs, which is not only clinically relevant, but increases the decoding difficulty, as the presence of noise and artifacts inherent in ambulatory recordings necessitates robust generalizability. M-Isch is often transient, resulting in a limited number of beats with detectable ST-T changes within the given dataset. To address class imbalance, the data was up sampled using various data augmentation techniques. The dataset was divided into a training set, validation set, and test set, with a split ratio of 0.70, 0.20 and 0.10 respectively, which was determined to provide sufficient data for training, while allowing for adequate validation and training. To mitigate data leakage, the data was stratified to ensure balanced class representation in each batch, with sufficient temporal separation between training, validation, and test sets. Additionally, each batch was standardized using z-score normalization to ensure consistent input distributions across batches during training. The performances of six deep learning models—Fully Convolutional Network (FCN), Residual Network (ResNet), Temporal Convolutional Network (TCN), OmniScale Convolutional Neural Network (CNN), Residual Convolutional Neural Network (ResCNN) and InceptionTime— were compared. Initial training was constrained to binary classification between either single-lead sinus rhythm or ST-elevation (>0.2 mV) M-Isch. Individual beats, defined by 276 milliseconds (ms) before and 500 ms after the R wave respectively, were used to train and test the models from a single subject within the database. The results show that the ensemble-based architectures outperformed all other architectures, with high sensitivity, specificity and accuracy, precision and F1 scores (InceptionTime: 91.75%, 92.20%, 93.01%, 90.86% and 0.9371, respectively; ResCNN: 89.64%, 91.84%, 90.25%, 87.52% and 0.9187, respectively). These findings suggest that DL has the potential to improve the sensitivity and specificity of ambulatory M-Isch detection from a general performance level of sensitivity ranging from 50% to 72% and specificity between 69% and 90%. This could conceivably lead to earlier identification of ischemia and facilitating important timely interventions in remote monitoring of ECGs that could prevent adverse cardiac events associated with M-Isch outside of traditional clinical settings. This study was supported by: NIH (R01 NS131493) This abstract was presented at the American Physiology Summit 2025 and is only available in HTML format. There is no downloadable file or PDF version. The Physiology editorial board was not involved in the peer review process.