Deep learning-based differentiation of Takotsubo syndrome and anterior acute myocardial infarction using electrocardiograms

Background Takotsubo syndrome often mimics acute myocardial infarction (AMI) on electrocardiograms (ECGs), making differentiation challenging, especially in patients involving recanalized AMI. Purpose We assessed the hypothesis that artificial intelligence (AI) models could distinguish Takotsubo syndrome from anterior AMI with or without coronary recanalization using ECGs. Methods Patients who were diagnosed with Takotsubo syndrome between 1996 and 2024 were included. Among patients diagnosed with anterior ST-segment elevation myocardial infarction (STEMI) due to occlusion of the mid-left anterior descending artery between 2010 and 2024, 321 were randomly included. Healthy volunteers were also included. ECGs recorded within 2 weeks of the index event were labeled as Takotsubo or AMI groups. ECGs between 2 weeks to 1 year after the index event were excluded, and the remaining were managed as non-Takotsubo/AMI. ECGs from heathy volunteers were also treated as non-Takotsubo/AMI. Three deep neural network-based models were pretrained and fine-tuned using the study datasets with both 3-class and 2-class labelling, and the final ensemble model for binary prediction (Takotsubo/AMI) was developed. The performance of the models was evaluated using an external test dataset from an independent institute. Results Among 308 patients with Takotsubo Syndrome and 321 with anterior STEMI we developed the models. The final model achieved an accuracy of 0.914, a receiver-operating-characteristics area-under-curve (AUC) was 0.970, precision-recall AUC was 0.984, the Expected Calibration Error was 0.0253, Brier Score 0.0654, and Brier Skill Score was 0.712 using external dataset (positive AMI, Figure 1, 2). Conclusions The AI model could differentiate Takotsubo syndrome from anterior STEMI on ECGs and has the potential to assist clinicians.