JoWUA

Abstract

Accurate identification of Learning Disabilities (LD) at an early age is vital for timely intervention, individualized learning plans, and optimal learning results. Conventional evaluation techniques like clinical evaluations and teacher observations can be time-consuming, subjective, and may not be able to identify subtle trends in academic and behavioral performance. This work suggests an explainable AI framework that combines real-time context-aware deployment in edge and mobile platforms with a Random Forest (RF) classifier to predict if a student is at risk for identification as an LD. A total of 348 student records containing academic, cognitive, behavioral, and demographic characteristics. To build trustworthy models, categorical encoding, feature standardization, and class balancing (RandomOverSampler) were performed. Multiple machine learning models, such as Random Forest, Decision Tree, K-Nearest Neighbor, Extra Trees, and Support Vector Machine, were trained and tested using 3-fold cross-validation. The overall best predictive performance was given by the Random Forest model with 96.36% accuracy, 96.00% F1 Score, 98.30% ROC-AUC, and 92.71% Cohen's Kappa, showing the reliability and stability of the model in classification tasks. In order to be interpretable, Shapley Additive exPlanations (SHAP) were computed, and academic performance, writing ability, and impulsivity were found to be the strongest predictors. Local SHAP explanations helped teachers get knowledge of the factors that are putting each student at risk, providing teachers with transparency and informing intervention. It is feasible also for ubiquitous and robust deployment in distributed education with the monitoring and decisions being done in real time in the edge or mobile device. This framework described here is capable to support the future deployment on mobile and edge computing environments by the computational efficiency of the under lying machine learning algorithms.