Adaptive Multidimensional Dual Attentive DCNN for Detecting Cardiac Morbidities Using Fused ECG-PPG Signals
Poulomi Pal, Manjunatha Mahadevappa
Abstract
Clinicians refer to standard physiological signals (ECG and PPG) to diagnose conduction disorders and coronary arterial malfunctions. The exact disease is identified from the results of confirmatory medical tests. This procedure creates a time lag in treating patients and causes difficulty in attending emergency cases. In this study, a novel method is proposed to detect and classify diseases related to electrical impulses and coronary arteries from other cardiac abnormalities efficiently. The ECG and PPG signals are collected simultaneously from 300 cardiovascular diseased patients, following a predecided inclusion and exclusion criteria. A fused signal is generated using an algorithm from the PPG and ECG signals. The deep neural network is constructed involving self and cross attention properties, multidimensional convolution, and skip connections. Reinforcement learning is utilized to induce an adaptive property in the model. The model’s ability in distinguishing the targeted diseases is achieved with an accuracy of 0.80, recall of 0.75, precision of 0.73, micro avg F-score of 0.78, and specificity of 0.85. Ablation studies are performed to get an efficient network. Comparison with the state-of-art methods and other established networks is made to denote the superiority of the proposed network. The fused signal is more efficient in comparison to the individual signals. Analysis of the ROC and PR curves produced an AUC and AP of 0.79 and 0.76, respectively. This technique could be used to identify the type of CVD at a preliminary stage of clinical diagnosis effectively involving less time, effort, and resources.