Optimizing Deep Neural Models for Early Dyslexia Detection Using Novel Bilingual Handwritten Dataset

ABSTRACT

Dyslexia is one of the most common learning disabilities and is difficult to detect in the early stages of learning. This is especially complicated in multicultural education systems. An AI‐based framework for the detection of dyslexia using handwritten samples of children between the ages of 6 and 10 years is presented in this paper. Due to the scarcity of multilingual dyslexia datasets, we propose a dataset using balanced English‐Urdu handwriting containing 852 samples in total. The dataset consists of 426 samples of dyslexic children and 426 samples of non‐dyslexic children. Several deep learning architectures using Convolutional Neural Network (CNN), VGG16, InceptionV3, MobileNet1, MobileNet2, and MobileNet3 were used with three optimizers, Adam, SGD, and RMSProp. Evaluation of the optimization of the deep learning methods was suggested to assess the reliability of the results. It was found that the MobileNet1 optimizer with RMSProp managed to achieve results that were overall the best with a high mean accuracy of and resulted in the highest mean ‐score of 0.7823. These results were especially notable for MobileNet2 with RMSProp and SGD optimizers and a slight decrease in the mean accuracy of the CNN was observed with the SGD optimizer as baseline. This indicates that lightweight deep learning architectures with apt optimizers are effective in capturing handwriting patterns of learners that have dyslexia in a bilingual English–Urdu context. The results go on to show that for small handwriting datasets, repeated‐run statistical evaluation offers a more consistent assessment than single‐run accuracy reporting. While the framework shows promise to some extent as an early screening support tool, its practical educational or clinical use is contingent upon its large datasets, participant‐wise validation, and real‐time deployment evaluation. This study advances multilingual dyslexia research by providing a bilingual handwriting dataset and a statistically solid evaluation of lightweight deep learning models for early dyslexia detection.