Neurocognitive prediction of dyslexic handwriting pattern using an explainable AI-driven custom LiteBinaryNet-CNN

Abstract

Artificial intelligence (AI) based on deep learning, particularly convolutional neural networks (CNNs), shows strong potential in handwriting recognition. However, their limited transparency constrains use in sensitive domains such as dyslexia prediction. From a neurocognitive standpoint, dyslexia stems from atypical neural processing reflected in handwriting irregularities, making handwriting prediction a neurocognitive inference task. This study introduces a neurocognitively informed framework, Custom LiteBinaryNet, a lightweight CNN integrated with Explainable AI (XAI) for transparent dyslexia prediction from handwriting. Custom LiteBinaryNet-CNN was evaluated in baseline and tuned configurations, the latter optimized through aggressive augmentation and hyperparameter tuning. Compared to LeNet-5 (60.49% accuracy, AUC 0.56), the baseline achieved 78.73% accuracy and AUC 0.87, while the tuned model reached 83.36% accuracy and AUC 0.91. Loss analysis confirmed improved stability and generalization. XAI methods, including Grad-CAM and Occlusion Sensitivity, revealed neurocognitive interpretability by highlighting handwriting traits, such as letter reversals and spatial inconsistencies, which linked to dyslexic motor patterns. These results align computational predictions with cognitive evidence, enhancing transparency and diagnostic value. The proposed model offers a practical and explainable approach for early neurocognitive prediction of dyslexia through handwriting analysis.