Key points are not available for this paper at this time.
Abstract Early identification of learners at risk of learning difficulties (LDs) is essential for timely educational support, particularly in low-resource school settings where access to formal diagnostic services is limited. However, existing school-based identification practices often rely on subjective teacher judgment, leading to delayed or inconsistent intervention. This study proposes an explainable machine-learning (ML)–based risk assessment and decision-support framework for stratifying LD risk among upper primary school learners in Ghana’s basic education system. Data were collected from 2,115 learners across public schools over two academic years (2022–2024), incorporating psychometric, behavioral, and demographic indicators derived from validated instruments. Learners were categorized into three LD risk tiers-low, moderate, and high-aligned with Response to Intervention (RTI) principles using statistically informed composite thresholds. Several classifiers, including Random Forest, Support Vector Machine, XGBoost, and Logistic Regression, were evaluated alongside a stacking ensemble model. The ensemble achieved the strongest screening performance (accuracy = 95%, macro-F1 = 93%, ROC-AUC = 0.995), interpreted within the context of educational risk identification rather than clinical diagnosis. Explainability analyses identified working memory, rapid naming, math-related constructs, ADHD-related behaviors, and vocabulary as key contributors to LD risk stratification, consistent with established cognitive and learning theories. The proposed framework demonstrates the potential of explainable ML to support teacher-informed early screening and prioritization of learner support within the study context, while external validation and local tuning are required before application across other regions, languages, or school systems.
Odoom et al. (Tue,) studied this question.