A deep learning-based acute coronary syndrome-related disease classification method: a cohort study for network interpretability and transfer learning

Accurate and efficient diagnosis of acute coronary syndrome (ACS)-related cardiac diseases is crucial for optimizing medical resources and enhancing clinical treatment efficiency. In this paper, we construct two distinct ECG datasets: a clinical 12-lead original ECG dataset and a clinical 12-lead standard ECG dataset. Our focus is on the automatic diagnosis of diseases clinically associated with acute coronary syndrome, specifically T wave changes, ST-T segment changes, and pathological Q waves. We propose a fine-tuned EfficientNet-based network for model construction and conduct comparative experiments using various state-of-the-art networks. Our experiments demonstrate satisfactory results in the classification of four types of ECGs: normal ECGs and three types of abnormal ECGs, with an overall classification accuracy of 73.333%. Notably, the model achieves a classification precision of 0.963 for pathological Q wave types and a classification specificity of 0.877, 0.859, 0.911, and 0.993 for the four types of models. To evaluate the model's performance, we randomly select a subset of ECGs and invite two groups of doctors, each with 4–5 years of experience in ECG interpretation, to determine the gold standard. A comparison of the proposed model's classification performance with the doctors' annotations reveals that for T wave changes and ST-T segment changes, the individual doctors' accuracy is 70.115%, while the proposed fine-tuning model achieves an accuracy of 71.111%. Thus, the performance of our model matches that of experienced doctors. Furthermore, we apply transfer learning to generalize the model trained on the original 12-lead ECG dataset to the new standard 12-lead ECG dataset. The experimental results demonstrate that the transferred model achieves a classification accuracy of 71.066% on the standard dataset, equivalent to the accuracy obtained on the original ECG dataset. Notably, for the normal sinus rhythm type, the precision and sensitivity of the transferred model improved by 7.4% and 1.8%, respectively. Additionally, the ROC AUC values for normal sinus rhythm, T wave changes, and ST-T segment changes after transfer are 0.857, 0.714, and 0.770, respectively. Moreover, we utilize Grad-CAM technology to provide network interpretability, offering theoretical support for clinical electrocardiogram diagnosis based on ECG records and classification results. Regarding network model efficiency, the average classification time for the transferred learning network on each ECG is approximately 0.0035 s, indicating the model's suitability for clinical practice and its potential to enhance clinical diagnosis efficiency.