ECG based human activity-specific cardiac pattern detection using machine-learning and deep-learning models

Monitoring cardiac patterns under relaxed, cognitive, and physical stressors is crucial for identifying early signs of cardiac stress or abnormalities. This study analyzes ECG signals recorded during diverse activities such as sitting, math-reasoning, walking, jogging, and hand-biking, simulating these stressors. A deep-learning image-based convolutional neural network (CNN) model utilizing bispectrum-based contours was proposed to classify cardiac patterns by capturing the non-linear dynamics of cardiac behavior. Two approaches were employed: a feature-based random forest (RF) machine-learning model using time-domain, frequency-domain, and statistical features, and an image-based CNN model utilizing Continuous wavelet transform (CWT) based scalograms and bispectrum-based contours. Feature selection techniques, including Pearson correlation and least absolute shrinkage and selection operator (LASSO) regularization, were used to identify significant features for RF model input. RF model achieved 96.80 % accuracy and an F1-score of 92.22 %. CNN model outperformed it, achieving 98.44 % accuracy and a 96.11 % F1-score with CWT scalograms, and 99.16 % accuracy and a 97.89 % F1-score with bispectrum-based contours. Key features such as stress index and SNS-to-PNS ratio increased with cognitive and physical stressors, highlighting autonomic responses. Based on the results of analysis, the proposed CNN model with bispectrum-based contours demonstrated superior accuracy and reliability, showcasing significant potential for monitoring cardiac functions across diverse activities.