Towards seamless myocardial infarction and cardiac rhythm detection: a deep learning approach with automated...

In India, myocardial infarction (MI) is a significant cause of mortality related to cardiovascular diseases. Timely diagnosis is critical for addressing this issue. While prior studies have concentrated on digital electrocardiogram (ECG) data, the presence of noise and artifacts in paper electrocardiogram is not well addressed in the literature. So, it is important to incorporate a more diverse range of data to understand the intricacies involved in MI diagnosis comprehensively.

This paper proposes an approach for MI and non-MI classification and cardiac rhythm classification using ECG. Modified versions of GoogleNet-Gated Recurrent Unit (GRU) and ResNet50- Bidirectional Long Short-Term Memory (BiLSTM) models were utilized for MI/non-MI and cardiac rhythm classification, respectively. Three different datasets (PTB-XL (dataset 1), ECG images of cardiac patients' data samples (dataset 2), and hospital paper ECG (dataset 3) were used for training the datasets, and final testing was done on digitized hospital paper ECG.

The proposed model achieved 88.8% accuracy on the PTB-XL dataset for MI/non-MI classification but showed performance disparity on dataset 3 test data. After a training-retraining process on datasets 2 and 3 and testing on dataset 3 test data, the model showed an accuracy of 92.50% with a balanced performance. For five-class cardiac rhythm classification, the proposed model attained 93.50% accuracy.

Retraining models on diverse datasets enhances generalization to real-world scenarios. The study's findings emphasize the importance of incorporating various data sources to improve model robustness and reliability in clinical applications. The developed application facilitates the seamless integration of digitization and classification processes for enhanced diagnostic accuracy.