Graph Based ECG Classification: A Continual Learning Framework for Evolving Cardiac Diagnosis

Abstract

Recognition of ECGs is essential to early detection and tracking of heart conditions. Standard machine learning algorithms are not well suited to adapting to new data distributions and they do not generalize well to the real world clinical world where patient information continues to arrive. To eliminate this, we suggest a robust model that can be implemented as a combination of Graph Neural Networks (GNNs) and continual and lifelong learning to identify severe cardiac disorders: Normal (NORM), Myocardial Infarction (MI), ST/T Change (STTC), Conduction Disturbance (CD), and Hypertrophy (HYP). Based on the PTB-XL dataset, we clean ECG data and obtain significant timedomain features, entropy features, and complexity characteristics. These characteristics provide the global patient similarity graph where a patient is a node and the edges are built based on K-Nearest Neighbor (KNN) similarity. This graph is subsequently fed to GNN models that each learn a representation of patients and are then continuously adapted to new ECG classes. The system adapts itself on the arrival of new information related to the simulation of the deployment conditions to provide scalable and specific diagnosis of ECG. In our experiments, GraphSAGE has the highest accuracy, 97.29%, in comparison with GCN and GAT. It means that GraphSAGE is more confident with assigning various cardiac conditions. To increase flexibility, GraphSAGE model was used with continual learning. The model was trained step-by-step instead of training on all classes at once, with every step the model was presented with new classes and was to learn these classes with maintaining knowledge of previous classes. This has been effective in blocking catastrophic forgetting which is typical of machine learning where earlier learning is overridden when training new tasks. The conclusions were quite encouraging There was a slight increase in training accuracy overall, with the accuracy increasing by one percentage point between the first and the last task 92.57 to 96.94. This indicates the fact that the model maintained an accuracy and even increased with the addition of each class. Therefore, the approach of graph-based computation, innovative GNNs, and continual learning offer an ECG classification system that is effective, yet flexible. The proposed framework has a great potential of being applied in the real world clinical practice where medical data is variably changing with time and new disease patterns are created. It allows a variety of existing devices to be scaled, adapted, and be long-term datadrift-resistant, providing a very strong basis of intelligent, dynamic clinical decision support systems.