Toward early dyslexia screening through handwriting: A trustworthy XAI-enhanced deep learning framework

Dyslexia, a prevalent learning disorder, significantly impairs children's ability to process written language, often leading to challenges such as letter reversals, inconsistent spacing, and scrambled writing sequences. Early detection is critical for timely intervention, yet traditional diagnostic methods such as behavioral assessments and neuroimaging-are either time-consuming, subjective, or prohibitively expensive. Recent advances in deep learning offer promising alternatives, particularly through handwriting analysis, which is non-invasive and scalable. However, the "black-box" nature of many deep learning models raises concerns about interpretability, a key requirement for clinical and educational adoption. This study introduces a novel deep learning framework combining a custom Convolutional Neural Network (CNN), Gated Recurrent Units (GRUs), and an Attention mechanism to detect dyslexia from handwriting samples. The study uses handwriting samples from adolescents aged 12-15 with an established dyslexia diagnosis and a control group. It therefore provides a proof-of-concept for handwriting-based classification and a foundation for early screening, rather than a validated early screening tool. The CNN extracts spatial features, while GRUs model temporal patterns in writing sequences. The Attention mechanism highlights the most discriminative regions for screening. To enhance transparency, Explainable AI (XAI) techniques such as Grad-CAM, LIME, and feature scatter plots are added to provide visual explanations of the model's decisions. Evaluated on a dataset of 3170 handwriting images from dyslexic and non-dyslexic students, the model achieved 99.5% accuracy using leave-one-out cross validation, outperforming pre-trained models like ResNet-50. Ablation studies confirmed the CNN's pivotal role while GRUs and Attention provided complementary improvements. These findings demonstrate technical feasibility; however, the model has not yet been evaluated in younger, undiagnosed children, and conclusions about early screening utility should therefore be interpreted with caution. XAI visualizations, such as Grad-CAM and LIME, offer interpretable saliency maps, and the results show that the predictions made by the model depend on the regions where known dyslexic characteristics, such as irregularities in letter curvature, spacing, and alignment, often occur.