Cognitive Load Recognition Based on T-Test and SHAP from Wristband Sensors

The understanding of cognitive load plays a key role in increasing the potential of human-centric systems. Recently, cognitive load detection has attracted the attention of researchers. Competitions are being held and relevant data are being released in this regard. Managing cognitive load through wearable devices in daily life contributes, amongst others, to industrial safety. Wearable bands require high prediction results with less data because of their limited battery and processing power. Therefore, by detecting important features that characterize cognitive load, we aimed to obtain a high load detection classification accuracy using few features. In total, we detected 179 features such as heart rate variabilities, descriptive statistical, and frequency features. Important features were detected using the independent t-test and SHapley additive exPlanation (SHAP). Furthermore, an accuracy of 70.3% was obtained with only ten important features using the LightGBM classifier. Heart rate variability and galvanic skin response were used as the important features. Additionally, the discrete wavelet transform was used as a more important frequency-domain feature than the discrete cosine transform. The proposed cognitive load detection method achieved higher accuracy with fewer features using a lighter classifier than those reported by existing CogLoad data studies

Automated summary

Understanding cognitive load is crucial for enhancing the potential of human-centric systems. Recently, there has been a growing interest in cognitive load detection, with competitions and data releases in this field. Managing cognitive load through wearable devices contributes to industrial safety and requires high prediction results with limited data. By detecting important features characterizing cognitive load, this study achieved high load detection accuracy with few features and a lightweight classifier.