CardioMultiNet for cardiovascular disease diagnosis using multi lead electrocardiogram data

Since cardiovascular diseases (CVDs) are the leading cause of death in the world, it is high time to independently produce an automated diagnostic system of ECG signals that determines heart disease. Traditional deep learning models, such as CNNs, LSTMs, and other Transformer-based models, isolate local morphology (through convolution), dynamic behaviour (through recurrent units), and inter-lead contextual dependencies, creating a trade-off between performance, interpretability, and computational efficiency. This work introduces CardioDeepAI, a complete multi-lead ECG analysis framework comprising CardioMultiNet, a CNN-LM-Transformer hybrid architecture tailored to this task. The model uses convolutional layers to extract morphological features, stacked Bi-LSTM modules to model temporal dependencies, and Transformer-based attention mechanisms to capture global context and lead correlations. We also introduce a robust uncertainty-aware mechanism and explainability methods such as Grad-CAM, SHAP, and attention visualisation to improve robustness and provide insightful visualisations. On the PTB-XL, MIT-BIH, and CPSC datasets, our model achieves an accuracy of 98.2%, an F1-score of 98.1%, an AUC of 0.992, and a Matthews correlation coefficient (MCC) of 0.963, providing a strong basis for real-time clinical access and wearable health applications.