Amazing Power of DINOv2 for Automatic Diagnosis of 12-Lead ECG

The Electrocardiogram (ECG) is a reliable and fast diagnostic test for detecting several heart anomalies to enable timely intervention. However, accurate interpretation is crucial and requires skilled professionals. Automatic diagnosis of heart disease from ECG can solve the problem. The proposed method in this paper uses a Visual transformer, DINOv2, for ECG diagnosis. DINOv2 is trained on many real-world object images. The model has been finetuned with ECG images. This finetuned model is used for extracting relevant features required by the classification method for anomaly detection. The main aim of this paper is to classify the widely known CODE-15 dataset with increased accuracy. This dataset is 15% of data from a larger CODE dataset. This 12-lead ECG dataset has samples for six heart conditions: 1dAVb, RBBB, LBBB, SB, AF, and ST. The model was finetuned for 10 epochs and achieved a test accuracy of 96.30%. This accuracy has been achieved with a short duration (2 seconds) of ECG data only. Ribeiro et al. (2020) have tried to classify the entire CODE data and have been able to achieve an average precision of 92.37%. In this paper, we have achieved an increase in accuracy to the tune of 3.3% with a smaller subset dataset.