A 1D CNN using all 12-lead 10-sec ECGs predicted VT in post-MI patients with 0.79 accuracy, 0.84 recall, and 55% patient-level VT detection years before onset.
Can artificial intelligence applied to sinus rhythm ECGs predict the long-term risk of ventricular tachycardia in post-MI patients?
An AI model applied to standard sinus rhythm ECGs can identify post-MI patients at risk for ventricular tachycardia months to years before onset, potentially enabling earlier personalized interventions.
Absolute Event Rate: 0% vs 0%
Abstract Background Post-myocardial infarction (MI) patients are at risk of developing scar-based ventricular tachycardia (VT) years later, potentially leading to sudden cardiac death (SCD). Long-term risk prediction will allow for personalized care pathways. Artificial Intelligence (AI) may improve VT risk stratification using electrocardiography (ECG) data. However, current models predict VT only hours in advance—too short for clinical interventions, such as ICD implantation. Purpose To classify sinus rhythm ECGs of post-MI patients, recorded months to years before VT onset, into VT and non-VT using AI. Methods In this retrospective study, patients with primary coronary intervention for acute ST-elevation myocardial infarction and at least one cardiology appointment within a 5 year follow-up period were included. Patients were classified as VT (monomorphic sustained VT 48 hours post-MI) or non-VT. ECGs recorded between two month post-MI and VT or last clinical contact, with normal sinus rhythm and a QRS duration ≤120ms were selected for further analysis. This resulted in a unique longitudinal dataset of 436 ECGs from 90 VT patients and 10,381 ECGs from 1,933 non-VT patients, recorded between 2002 and 2023. The data were denoised and split into training (70%), validation (15%), and test (15%) sets, ensuring ECGs from each patient appeared in only one set while maintaining similar patient distributions. Three AI models were trained: a 1D convolutional neural network (CNN), a residual Network, and a pretrained CNN, across 40 configurations with a special focus on addressing the class imbalance between VT and non-VT patients:• Two ECG formats (full 10-second ECG or single median beat)• Four lead selections (lead I, lead aVR, lead II+aVR+V1+V4 or all 12 leads)• Five balancing approaches (none, random upsampling, random downsampling, Synthetic Minority Oversampling Technique or class weighting) Resutls The best-performing model was the 1D CNN trained on full 10-second ECGs with all 12 leads, balanced using class weights. This model achieved• Accuracy: 0.79• Precision: 0.10• Recall: 0.84• Area under the Receiver Operating Characteristic Curve: 0.81 Figure 1 displays the confusion matrix for ECG predictions, while Figure 2 illustrates the prediction results for a typical VT patient. Each bar represents an ECG, with its height indicating the predicted probability of belonging to the VT class. Averaging these probabilities across all ECGs per patient enabled prediction at patient-level instead of ECG-level, correctly classifying 55% of VT patients while misclassifying only 7% of non-VT patients. Conclusion AI models can identify VT risk from sinus rhythm ECGs of post-MI patients recorded months to years before VT onset. This may enable earlier high-risk patient identification and timely intervention strategies offering a promising approach for long-term, personalized care in patients at risk for scar-related VT.ECG Classification Confusion Matrix ECG predictions of a VT patient
Buck et al. (Sat,) reported a other. A 1D CNN using all 12-lead 10-sec ECGs predicted VT in post-MI patients with 0.79 accuracy, 0.84 recall, and 55% patient-level VT detection years before onset.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: