Generative machine learning applied to the electrocardiogram (ECG) to uncover relevant ECG features in distinguishing takotsubo syndrome from myocardial infarction

Background Takotsubo syndrome (TTS) almost always presents with a pathological electrocardiogram (ECG) that initially is indistinguishable from myocardial infarction (MI). Studies have tried to develop an ECG-criteria to diagnose TTS but have been hampered by heterogeneity and small sample size. Previously machine learning (ML) classification of TTS from MI, using neural networks, have been limited by the black-box nature of the method and by low precision-recall, and low positive predictive value. Purpose Our aim was to apply generative machine-learning to the dynamical clustering of heartbeat morphologies to distinguish between TTS and MI, improving previous classification results, and emphasising visualisation of potentially new ECG features for takotsubo syndrome. Methods 507 TTS cases and 14,978 controls, from a database of takotsubo syndrome and myocardial infarction subjects, admitted to hospital between 2010 -2019 with a suspicion of MI and an indication for coronary angiography, were identified from the Swedish SCAAR register. The main outcome of the study was the ability of ML to identify (1) TTS from suspected MI, (2) TTS from ST-elevation MI (STEMI), (3) TTS from non-ST-elevation MI (NSTEMI), (4) TTS from verified MI, and (5) apical TTS from anteriorSTEMI. Bayesian non-parametric clustering was applied to the sequence of heartbeats, identifying relevant features, and classified recordings by logistic regression. 10-fold cross-validation was performed to emphasize the robustness of our results. Results ML differentiated TTS from MI: 1) TTS from suspected MI with area under the curve (AUC) of 0.89 (95% CI 0.88-0.91); 2) TTS from verified MI with 0.92 (0.90-0.93); 3) TTS from ST-elevation MI (STEMI) with 0.94 (0.93-0.95); 4) TTS from non-ST-elevation MI with 0.91 (0.89-0.92); and 5) apical TTS from anterior STEMI with 0.90 (0.88-0.92). In each scenario the model characterized electrocardiographic patterns for classification. Pointing towards a combination of heterogenic features, including ST-elevations/depression, inverted T-waves, QT-prolongation associated with takotsubo syndrome. As a novel finding atrial activation disturbances was found to be associated with TTS and important for classification. Conclusion In the present approach ML could discriminate TTS against suspected and verified MI with good to excellent performance. The study shed light on new ECG features for takotsubo syndrome. Providing evidence for a heterogenic pattern of the ECG features and supporting criteria previously recognised in the literature. For the first time atrial activation disturbances were found to be associated with TTS and important for classification.