Multimodal Wearable Sensors-based Stress and Affective States Prediction Model

In the modern world, stress is considered one of the critical health problems. Moreover, its long-term retention develops hypertension, heart disease, depression, sleep apnea, and other related diseases. Hence, it is necessary to detect the stress condition, and its cause and mitigate that to avoid its prolonged effects. It is required to develop portable technology to keep an eye on stress conditions. The present work proposed a multimodal wearable sensors-based stress and affective states prediction model. It utilizes wearable sensor signals of different modalities acquired from the wrist and chest. The performance of six classifiers with individual modality has been estimated for the multifactor physiological dataset Wearable Stress and Affect Detection (WESAD). Temperature signal has been found the most informative from the pool of wrist-worn sensor. Whereas ECG and GSR have been found most discriminative from chest-worn sensors. The random forest (RF) classifier emerges as the best-performing classifier. In general, the chest-worn sensors performed better as compared to wrist-worn sensors. For the wrist-worn and chest-worn sensors, the best performance has been achieved as 95.54±2.14% and 97.15±2.67%, respectively (pvalue<0.05) while utilizing all sensor information. The developed model is based on conventional machine-learning models and is suitable for embedded applications.