Meta-learning Based Cardiopathy Detection from PPG Signals Using GAN and 1D CNN

Diagnosis and identification of cardiac disease efficiently in a short time with minimal computational complexity is still a challenge in the medical field. Critical patients needing immediate treatment are delayed in identifying the cardiopathy (for confirming the exact clinical procedure required). In this clinical study, we attempted to conduct a pilot cohort study to configure a method to identify cardiopathy from the photoplethysmography (PPG) signals of the 360 cardiac patients in the hospital. Initially, the classification network was constructed using ECG signals from the Physio-Net database. The proposed generative adversarial network (GAN) comprising BiLSTM was used to augment the ECG signals. Then they were classified by the proposed 1D convolution neural network (1DCNN) designed with self-attention and cross-attention properties. After obtaining excellent results (accuracy=98.30%), this computational technique was implemented to classify cardiopathy from hospital-based PPG signals using Meta-learning. Performance evaluation of PPG signals resulted in an accuracy of 98.60%, Micro Average of F1-score 97.80%, and Cohen's Kappa of 96.90%. The state-of-the-art studies showed the superiority of the proposed work over prior works. Establishing this Meta-learning-based network on future clinical data might evolve a robust technique to categorize cardiopathy from non-invasive PPG signals. Implementing such methodologies in the clinical field for cardiac disease diagnosis could fasten emergency patient treatment.