Improving Early Detection of Learning Disabilities Using Graph Convolutional Networks (GCN) and Multi-Modal Data

Abstract

Learning disabilities such as dyslexia, dyscalculia, ADHD, and dysgraphia remain among the most underdiagnosed neurodevelopmental disorders in early childhood, often leading to long-term academic and psychological challenges when not identified at an early stage. The purpose of this study is to design an accurate and interpretable artificial intelligence framework that will facilitate the earlier detection of learning disabilities using an attention-based graph convolutional network with multi-modal data. By utilizing multiple sources of heterogeneous neurocognitive data (e.g., eye-tracking metrics, fMRI, EEG, neuropsychological assessment scores, demographic data) in a unified graph representation, the authors propose to capture complex relational dependencies between the different modalities. The proposed methodology includes multi-stage preprocessing, constructing graphs using connections within the functional and cross-modal data, performing attention-based GCN learning, and performing interpretability analysis using GNNExplainer. The framework was evaluated using a dataset containing 1240 children aged 4-10 years and five diagnostic groups through the use of stratified 5-fold cross-validation. The results of the experiments showed that the overall performance of this method (i.e., MM-AGCN) was superior to that of conventional ML and DL baselines. MM-AGCN achieved an accuracy of 91.7%, sensitivity of 89.4%, specificity of 92.8%, F1 score of 90.6%, and AUC-ROC of 0.947%. Moreover, through ablation studies, it was noted that the performance of the model improved by 8.1% because of the use of an attention mechanism and decreased by 6.5% because of no cross-modal edges. Among all modalities, fMRI and EEG contributed the highest diagnostic value. The findings confirm that graph-based multi-modal learning provides a highly effective and clinically interpretable approach for early learning disability diagnosis. The proposed framework offers strong potential for supporting early intervention planning, personalized education strategies, and AI-assisted neurodevelopmental screening in clinical and educational environments.