Early Detection and Severity Assessment of Dysgraphia in Sinhala-Speaking Children Using a Multi-Modal Machine Learning Approach

Dysgraphia is a neurological learning disability that impairs handwriting, spelling, and the ability to express thoughts in written form. Despite its significant impact on children's academic performance and psychological well-being, early detection remains a challenge particularly in linguistically and culturally unique settings like Sri Lanka. Existing diagnostic tools are largely developed for English-speaking populations and fail to address Sinhala language-specific characteristics. To bridge this gap, this research introduces the first multi-modal sequential machine learning framework in Sri Lanka for the early detection and severity assessment of Dysgraphia among Sinhala-speaking primary school children. The proposed approach integrates two key components: a Convolutional Neural Network (CNN) utilizing VGG16 for handwriting image analysis, and a Gradient Boosting classifier for interpreting cognitive, behavioral, and personal data. The dataset comprises 373 digitized handwriting samples from 84 children, collected from local schools and a pediatric hospital, using psychological assessments from Indian contexts adapted to Sinhala language materials. Handwriting samples were preprocessed through binarization and color inversion for model input. The handwriting-based Dysgraphia detection model achieved an accuracy of 96%, while the severity assessment model reached 87%. This pioneering multi-modal sequential approach significantly enhances diagnostic precision and reliability, enabling earlier interventions and individualized academic support. The study represents a vital step toward inclusive education in Sri Lanka, leveraging advanced AI techniques to meet a critical need in local learning disability diagnostics.