Abstract 15414: Diagnostic Performance of Deep Learning on 12-lead Electrocardiography to Distinguish Takotsubo Syndrome and Acute Anterior Myocardial Infarction

Introduction: Takotsubo syndrome (TTS) and acute anterior myocardial infarction (ant-AMI) show very similar 12-lead electrocardiography (ECG) featured at onset, and it is often difficult to distinguish them without cardiac catheterization. The difference of ECG between them was studied, but the diagnostic performance of machine learning (deep learning) for them had not been investigated. Hypothesis: Deep learning on 12-leads ECG has high diagnostic performance to diagnose TTS and ant-AMI at onset. Methods: Consecutive 50 patients of TTS were one-to-one matched to ant-AMI randomly by their age and gender, and total 100 patients were enrolled. No sinus rhythm patients were excluded. All ECGs were divided into each 12-lead, and 5 heart beats from one lead were extracted. For each lead, 250 ECG waves of TTS/AMI were sampled as 24bit bitmap image, and prediction model construction by convolutional neural network (CNN: transfer learning, using VGG16 architecture) underwent to distinguish the two diseases in each lead. Next, gradient weighted class activation color mapping (GradCam) was performed to detect the degree and position of convolutional importance in the leads. Results: Lead aVR (mean accuracy 0.748), I (0.733), and V1 (0.678) were the top 3 leads with high accuracy. In aVR lead, GradCam showed strong convolution of negative T wave in TTS, and sharp R wave in ant-AMI. In I lead, it spotlighted several parts of ECG wave in ant-AMI. However in TTS, whole shape of the wave, P wave onset, and negative T were invertedly convoluted in TTS. Conclusions: Deep learning was a powerful tool to distinguish TTS and ant-AMI at onset, and GradCam method gave us new insight of the difference on ECG between the two diseases.