Optimization of Convolutional Neural Network for Accurate Stress Detection Using Multimodal Physiological Signals

Abstract

Stress is an integral part of daily life that most individuals must manage regularly. However, long-term stress or high levels of stress can compromise safety and disrupt normal lifestyles. Early detection of mental stress can help prevent numerous stress-related health problems. When an individual experiences stress, noticeable changes occur in several physiological signals, including impedance, thermal, electrical, and optical signals. By analyzing these signals, stress levels can be effectively determined. Even with the use of advanced technology, existing research on stress detection has failed to produce satisfactory accuracy. To address this gap, the research proposed two ideas. First, instead of using a single physiological signal, the study used multimodal signals. Electrocardiogram (ECG) and Electroencephalogram (EEG) signals under no-stress, low-stress, and high-stress conditions were acquired from the Kaggle site. Second, the Convolutional Neural Network (CNN) hyperparameters were tuned using a bio-inspired optimization technique called the Firefly Algorithm (FA). The drawbacks of the FA were identified and further improved, leading to the development of an Improved Firefly Algorithm (IFA) to fine-tune the CNN hyperparameters. The multimodal data from Kaggle was processed to remove noise and then fed into the proposed IFA-CNN, FA-CNN, and a baseline CNN model to predict the stress levels of individuals. Additionally, the three models were also tested with ECG and EEG data separately. The outcomes of all three models, using ECG, EEG, and multimodal data, were compared using positive metrics (accuracy, recall, precision, F1-score) and negative metrics (False Negative Rate (FNR), and False Positive Rate (FPR). The experimental results showed that the proposed IFA-CNN using multimodal data achieved the highest correct stress-level prediction, with 47 out of 48 samples correctly identified, yielding an accuracy of 97.92%. The comparison results also highlighted the advantage of using multimodal data over single-signal data. The proposed approach is highly beneficial for reliable stress-level prediction in individuals.