Multimodal Physiological Signals and Machine Learning for Stress Detection by Wearable Devices

Wearable technology is growing in popularity, and wearable devices, such as smartwatches, are used in many applications, from fitness tracking and activity recognition to health monitoring. As the affordability and popularity of such devices increase, so does the amount of personal and unique data that they provide. At the same time, advantages in microprocessor and memory technology enable multiple physiological signal sensors integrated into wearable devices to collect personal and unique data. After the data is extracted, machine learning classification algorithms can help investigate the insights of the data. In this work, we examine the performance of a real-time stress detection system based on physiological signals collected from wearable devices. Specifically, three physiological signals, electrodermal activity (EDA), electrocardiogram (ECG), and photoplethysmo-graph (PPG) that can be collected through smartwatches, are examined for stress classification. Six machine learning methods are used for the classification in a post-acquisition phase, at a computer, including Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest, Naive Bayes, Logistic Regression, and Stacking Ensemble Learning (SEL). Data from two publicly available datasets are used for training and testing. We examine the accuracy of each modality and the combination of all modalities. According to evaluation results, EDA has the best accuracy when SEL is used for classification. Also, the accuracy of EDA outperforms the other signals and combinations, in comparison with any of the other machine learning approaches, for both datasets. EDA collected from the wearable device has a great potential to be used for a real-time stress detection system.