Stress Detection and Prediction Using CNNs from Electrocardiogram Signals

Stress prediction is a crucial aspect of mental health monitoring, with consequences for both psychological well- being and productivity. This work presents a unique way for stress prediction that uses binary and multiclass classification models. Through extensive experimentations with different durations and frequencies of Electrocardiogram Signal (ECG) signals, we identified a 5-second dataset sampled at 200Hz as the optimal configuration for our model. Moreover, we introduced an innovative feature i.e., the prediction of stress scores ranging from 0 to 100, providing nuanced insights into stress levels, where 0 represents no stress and 100 indicates high stress levels. The model obtains 95.04% accuracy, 95.27% precision, 94.95% F1 score, 86.69% sensitivity, and 99.44% specificity for the binary classification. With "Fun" added to the list of stress categories in addition to "Base" and "TSST," the model continues to perform well in the multi-class classification scenario, with accuracy of 88.10%, precision of 87.60%, F1 score of 87.35%, sensitivity of 95.97%, and specificity of 79.23%. These findings highlight how well this applied strategy predicts stress levels, providing important information for mental health and stress management strategies.