Key result
Recurrent neural networks achieve a ~93% F1 score for cross-subject stress classification.
Why the study?
Early detection of mental stress via bio-signals can prevent related health issues, prompting the use of machine and deep learning on multimodal wearable sensor data.
Can machine learning and deep learning models accurately detect stress states using multimodal physiological data from wearable sensors?
Can machine learning and deep learning models accurately detect stress states using multimodal physiological data from wearable sensors?
Machine learning and deep learning models, particularly RNNs and ensemble tree classifiers, can accurately classify stress states using multimodal physiological data from wearable sensors.
This is the authors' abstract. We don't add key points for this paper.
Stress can disrupt daily activities and harm health if prolonged or severe. Early detection of mental stress, indicated by changes in bio-signals like thermal, electrical, and acoustic signals, can prevent related health issues. This study employs machine learning and deep learning techniques on multimodal dataset from wearable sensors, focusing on processed metrics for the three-axis acceleration (ACC), electrocardiogram (ECG), blood volume pulse (BVP), body temperature (TEMP), respiration (RESP), electromyogram (EMG), and electrodermal activity (EDA) from the 15 subjects in the WESAD dataset to effectively classify four different states – baseline, stress, amusement, and meditation. Seven traditional machine learning algorithms – Logistic Regression, Gaussian Naïve Bayes Classifier, AdaBoost Classifier, XGB Classifier, Decision Trees Classifier, Extra Trees Classifier, and Random Forest Classifier and three widely used deep learning algorithms – Deep Neural Network, Convolutional Neural Network and Recurrent Neural Network were trained and tested on the dataset on two phases to predict the state of different subject at any given time. Our findings indicate that Recurrent Neural Networks achieved an F 1 score of 93% when trained on a selected set of subjects and tested on the data from different subjects. Traditional machine learning algorithms, including Random Forest, Extra Trees, and XGB Classifiers, on the other hand, each achieved an F 1 score of 99% when trained and tested on the data for the same set of subjects. Additionally, models performed better on chest data when trained and tested on the same subjects, while they perform better on wrist data when trained on a random group of subjects and tested on the remaining ones.
Answers are grounded in this paper and cite their sources.
Cautions against clinical adoption of wearable stress classifiers; leaves open cross-subject generalizability and prospective validation.
No takes yet. Share an insight, caveat, or question.
Abdelfattah et al. (2025) studied Mental stress (n=15). Machine learning and deep learning algorithms was evaluated on Classification of four states: baseline, stress, amusement, and meditation (F1 score 93% (RNN) and 99% (Traditional ML)). Recurrent Neural Networks achieved an F1 score of 93% for cross-subject stress classification, while traditional machine learning models achieved 99% when trained and tested on the same subjects.
Synapse has enriched 10 related papers. Consider them for comparative context: