Subject-Independent Stress Classification using Multimodal Wearable Biosignals

Abstract

Our goal is to develop a stress detection framework that works across different individuals utilizing a variety of physiological signals from
wearable devices, considering that stress affects both mental and physical health. As we have access to more sensing technology around us, we
can monitor physiological signals continuously, which can then be utilized for the automated detection of stress. We have proposed an approach
for classification of stress, which is independent of the subject. For this purpose, multimodal wearable biosignals have been utilized along with
machine learning. Physiological signals, such as Heart Rate(HR), Inter-Beat Interval (IBI), Electrodermal Activity (EDA), Blood volume pulse
(BVP), Skin Temperature (TEMP), and Accelerometer (ACC) signals, have been utilized from various subjects under different activity
conditions. Simple statistical features such as the mean and standard deviation, have been extracted after the noise filtering.
IN order to reduce the Inter-variability between individuals, we normalized the data on a per-subject basis before model training. We have used
ten different types of classifiers, such as Logistic Regression, support vector machine, K-Nearest Neighbours, Decision Trees, Random Forest,
Gradient Boosting, Extra Trees, AdaBoost, Naive Bayes, and XGBoost. In our initial phase of experiments, we have observed around 84%
accuracy with a random train-test split. However, this could lead to a scenario where the model is exposed to same individuals in the train and
test set, leading to biased accuracy estimates. To assess the model in a more realistic scenario, we employed a subject-independent validation
strategy using the Group K-Fold cross-validation method. In this case, the Extra Trees ensemble model was able to achieve an average accuracy
of around 72%.For the feature separability, we employed the PCA method and observed a partial clustering effect with some overlap between
the stress levels, this was scientifically expected. This could indicate the non- linear nature of the physiological stress response. From the
feature importance results, it was observed that HR variability features obtained using IBI and EDA signals are among the most valuable
information in determining the levels of stress.Our results indicate the efficacy of using ensemble methods in the development of a reliable
model for stress detection in a wearable healthcare system.