SNNLCnet: A shallow neural network integrated model for electroencephalogram-based classification of learning comprehension
Shin-ichi Ito, Ryota Miyake, Momoyo Ito, Minoru FukumiJOURNAL/atin/04.03/02274269-990000000-00035/figure1/v/2026-10-01T120209Z/r/image-tiff
Objectives:
To address the inherent difficulties in assessing student comprehension using educational technology, this study develops and validates an integrated model that combines shallow and convolutional neural networks. The model aims to achieve refined quantification and classification of student learning outcomes.
Methods:
A shallow neural network-driven electroencephalogram-based classification method for learning comprehension (SNNLCnet) is designed. The proposed workflow comprises electroencephalogram measurement, signal normalization, input-data image generation, feature extraction, and classification. Electroencephalography signals are recorded using a commercially available, user-friendly electroencephalography device. To adapt the convolutional neural network models for electroencephalography feature extraction, input data images are generated from normalized electroencephalography signals. For feature extraction and classification, we propose a shallow neural network comprising one convolutional layer and two fully connected layers. The proposed method is validated using macro-averaged accuracy, macro-averaged precision, macro-averaged recall, and macro-averaged F1 score.
Results:
SNNLCnet achieves macro-averaged accuracy, macro-averaged precision, macro-averaged recall, and macro-averaged F1 scores of 0.938, 0.926, 0.930, and 0.924, respectively. These values are higher than those of benchmark models, namely EEGNet, a shallow convolutional neural network, and other convolutional neural network models, demonstrating the effectiveness of the proposed approach.
Conclusion:
SNNLCnet can accurately classify the presence or absence of learning comprehension in human-common responses by connecting one convolutional layer to two fully connected layers. Although these results did not identify a specific electroencephalography signature (e.g., an event-related potential) indicative of comprehension, a feature-extraction method and a shallow convolutional neural network are proposed, which enable straightforward and highly accurate classification of comprehension and non-comprehension.