Multimodal Neurophysiological Representations of High School Students’ Situational Interest: A Machine Learning Approach

Abstract

Interest plays a vital role in students’ learning performance. Accurately measuring situational interest in the classroom environment is important for understanding the learning mechanism and improving teaching. However, self-report measurements frequently encounter issues of subjectivity and ambiguity, and it is hard to collect dynamic self-report scales without disturbance in the naturalistic environment. Thanks to the development of neuroscience and portable biosensors, it has become possible to represent psychological states with neurophysiological features in the classroom environment. In this study, multimodal neurophysiological signals, including electroencephalograph (EEG), electrodermal activity (EDA), and photoplethysmography (PPG), were applied to represent situational interest under both laboratory (Study 1) and naturalistic (Study 2) paradigms. A total of 33 features were extracted, and 7 different statistical indicators were calculated for each of them across all the epochs. Among these features, 47 in Study 1 and 49 in Study 2 demonstrated significant correlation with self-report situational interest. Employing a machine learning model, the analysis yielded a mean absolute error (MAE) of 0.772 and mean squared error (MSE) of 0.883 for the dataset in Study 1. However, the model was not robust on data from Study 2. These findings offer empirical support for the conceptual framework of situational interest, demonstrate the potential of neurophysiological data in educational assessments, and also highlight the challenges in naturalistic paradigm.