Adaptive Multimodal Dyslexia Detection Using Incremental XGBoost with Cognitive Feature Embeddings

The timely intervention of dyslexia through early diagnosis is important to facilitate the implementation of individualized learning techniques in children. The paper introduces a new adaptive model of real-time dyslexia detection with a combination of multimodal data sources, which can be seen as reading exercises, writing tasks, and eye-tracking interactions. Out of these sources, the primary and derived cognitive characteristics, such as reading rate, omitted words, duration of fixation, spelling errors, error density ratings, fixations-error associations, cognitive load ratings, and writingreading discrepancy ratings are derived. The proposed Incremental XGBoost model is based on the difference with the traditional methods of batch-learning where it is constantly updated with new student records, which enables the system to adapt dynamically to the changing learning trends without undergoing total retraining. This is an online learning ability that enhances prediction accuracy and strength and minimizes the computational burden. Educators can use the explainable allowances of AI, e.g. SHAP values, to offer interpretable data on the significance of features, which can help in comprehending specific learning challenges and designing customized responses. This is supported by experimental testing on a dataset of 150 students which indicates that the inclusion of cognitive features with incremental learning is much better than the normal XGBoost and only single-modality models. The system suggested is scalable, effective and can be implemented to work in real time in the classroom, providing a viable tool to improve early intervention strategies on dyslexia