Classification of working memory performance in primary school children using EEG time-frequency image features and convolutional neural networks

The capacity to keep and use information for short amounts of time is referred to as Working Memory. It is presumed that capacity of working memory is crucial for a variety of cognitive and noncognitive abilities. Because these capabilities have a substantial effect on children’s learning outcome, working memory assessment during early stage can benefit them. The purpose of the study is to propose an EEG-based working memory assessment system which can overcome the limitation of the current method that is carried out using scores which are subject to response bias during each testing, failing to directly quantify underlying brain activity. Artificial intelligence-based solutions can assist decrease downtime and enhance system performance by detecting tiny changes in memory performance that individuals might overlook. This study focuses on classifying the working memory performance of typically developing children by using visual stimuli assessments adapted from the Automated Working Memory Assessment and correlating the results with electroencephalogram data. The study consists of two stages. The first stage involves recording resting electroencephalogram data. The second stage includes two assessments: the Dot Game, which evaluates visual-spatial short-term memory, and Match the Shape Game, which assesses visual-spatial working memory. Children’s resting EEG recordings and performance scores from these working memory activities have been collected and categorised into Low, Medium, and High working memory group, defined based on the standardized score distribution. Spectral analysis revealed that the Low working memory group exhibited significantly higher theta power (4-8 Hz) while the high working memory demonstrated higher alpha power (8-12 Hz), supporting neural efficiency hypothesis. The Fz channel in the prefrontal cortex exhibiting the highest Power Spectral Density and Energy Spectral Density across participants. This channel was subsequently used to extract four time-frequency representations: spectrogram via Short Time Fourier Transform, scalogram via Continuous Wavelet Transform, Hilbert spectrum via Hilbert-Huang Transform, and fused image via Image Fusion as an input to Convolutional Neural Networks, specifically AlexNet and VGG16, trained with mini-batch sizes of 128, 256 and 512. All four feature representations demonstrated consistent trends aligned with the Neural Efficiency Hypothesis, with spectrogram and scalogram showing the most visually distinct pattern between WM groups. The classification models were developed to categorise working memory into three levels: Low, Medium and High. AlexNet achieved optimal performance with spectrogram features at mini-batch size 512 (90% accuracy for Dot Game) and with scalogram features at mini-batch size 128 (80% accuracy for Shape Game), while VGG16 excelled with scalogram features at mini-batch size 256 (82.33% for Dot Game, 80.67% for Shape Game). The recommended optimal configuration comprises AlexNet with spectrogram features and mini-batch size 512 for achieving highest overall performance among other tasks. To ensure the robustness of this result, a secondary evaluation was performed, which confirmed the 90% that the configuration captures fundamental neural oscillation in an entirely unbiased manner. These findings highlight the potential of EEG-based Artificial intelligence models for working memory assessment and provide insight for identifying children with early assessment and aid school administrators in making diagnoses and strategizing teaching methods to enhance academic achievement and mitigate learning challenge.