Cognitive load detection through wearable devices

Abstract

This study addresses the challenges of accurately classifying Cognitive Load (CL) between "rest" and "non resting" state. This classification has gained significant attention in recent years, due to the ability to enhance various industries such as Human-Computer Interaction (HCI), that also plays a vital role in the evolving industry of Artificial Intelligence (AI). Accurate real-time classification also enables better performance in various applications such as education, medical solutions and smart automotive. In this study the CogLoad dataset is used, created and used at the CogLoad@UbiComp 2020 Competition, which contains physiological features such as Respiratory Rate intervals (RR), Heart Rate (HR), Skin Temperature (TEMP), and Galvanic Skin Response (GSR). Our method proposes a robust Machine Learning approach to accurately classify the CL, by applying transformation on the "raw" physiological data into statistical calculated features based on a rolling window of 11 seconds, and also participant-domain statistics, to address the limitation of inter-subject variability. SMOTE technique to balance the dataset was utilised, to create a nearly balanced dataset. The best performing model achieved the results of; 95.79% Accuracy, 99.27% ROC-AUC, 96.38% Precision, 96.38% Recall and an F1-Score of 95.77%, which was the XGB boosting ensemble technique. This work provides a baseline for real time classification, and with potential for further improvements through expanded datasets that contains a larger variety of participants and task variety. It also highlights the potential of utilising wearable technology with advanced ML to predict CL, facilitating more effective and adaptive HCI systems in real-world scenarios.