Graph Attention Networks Based Cardiac Arrest Diagnosis

Abnormal waveforms of electrocardiogram (ECG) signals, which also known as shockable rhythms including ventricular fibrillation and ventricular tachycardia, are the primary source of sudden cardiac arrest (SCA). In this work, an effective paradigm for SCA diagnosis is proposed as the shock advice algorithm for the application in automated external defibrillators using ECG signals. The proposed algorithm is carefully constructed with graph attention networks (GAT) using a set of various features extracted from 8s- Ecgsegments. The GAT structure is optimized with exhaustive search method in which parameter selection is based on a gird search. 5-fold cross validation procedure is also implemented to obtain reliable performance results for both model optimization and validation stages. The proposed algorithm shows an accuracy of 99.21 %, a specificity of 95.59%, a sensitivity of 99.62% and a F1-score of 99.56%.