Explainable Machine Learning Framework for Predicting the Academic Performance of Deaf and Hard-of-Hearing (DHH) Students

Abstract

Deaf and Hard-of-Hearing (DHH) students face communication and language barriers that most academicperformance prediction systems overlook, since such systems are designed for hearing populations and ignore deafness-related factors such as communication mode, hearing-loss degree, family history of deafness, and age of speech-therapy intervention. This paper proposes an explainable machine learning framework, developed as a Computer Engineering academic project, to classify DHH students into High, Middle, and Low-performance categories using socio-economic, prior-academic, and deafness-related features. Since real-world DHH datasets are small and imbalanced, the pipeline uses synthetic data generation and SMOTE-based balancing before training classical and ensemble classifiers such as SVM, Random Forest, XGBoost, ExtraTrees, and a stacked ensemble. SHAP and LIME
generate transparent, actionable explanations for teachers, parents, and policymakers. As the system is at the design stage, this paper reports the proposed architecture, methodology, and evaluation protocol rather than empirical results, contributing a reproducible template for explainable, imbalance-aware learning in an underserved educational domain.