Myocardial infarction (MI) is alethal heart condition that occurs due to the lack of blood flow to the heart tissues. Based on the time from symptoms on set, it is categorized into three severity stages: early MI (EMI), acute MI (AMI), and chronic MI (CMI). Electrocardiogram (ECG) signals are often used to diagnose MI with pathological changes in its characteristics. In clinical practice, accurate diagnosis and risk-stratification are essential to optimize various treatment strategies, hence clinical outcome. However, most automated methods focus only on identifying MI patients from healthy controls (HC). Therefore, in this project, we propose a novel multi-lead diagnostic attention-based recurrent neural network (MLDA-RNN) for automated diagnosis of the three MI severity stages from HC subjects. The method systematically processes the 12-lead ECGs to capture the multi-scale temporal dependencies from each ECG leads for improved classification. Specifically, we first employ the RNNs to encode the temporal variations in the 12-leadECG signals. These encoded vectors are fed to the intra-lead attention module to summarize the within-lead discriminative vectors to obtain lead-attentive representations. Then, the inter-lead attention module aggregates these representative vectors based on their clinical relevance to obtain a high-level feature representation for a reliable diagnosis. Using 12-lead ECGs from the PTBDB and STAFF III datasets, we achieved an overall accuracy of 97.79%without compromising on the class-wise detection rates. With improved performance, the MLDA-RNN also shows promising results for model interpretability as the learned attention weights often correlate with clinicians’ way of diagnosing MI severity stages. This project is implemented with MATLAB software.