Approaching Early Dyslexia Prediction with Ensemble Machine Learning: A Secondary Empirical Analysis of the Rello Et Al. Public Behavioural Dataset

Dyslexia is a specific learning disorder associated with persistent difficulties in reading accuracy, decoding, spelling and fluent word recognition. Early identification is important because timely educational support can reduce the educational consequences of undetected reading difficulties. The increasing availability of digital behavioural data provides an opportunity to complement conventional assessment with machine-learning-based screening approaches. This study presents a secondary empirical investigation of ensemble machine learning for early dyslexia prediction using the publicly available behavioural dataset originally reported by Rello et al. The study uses the Dyt-desktop dataset containing 3,644 observations and 196 predictive variables derived from demographic characteristics and performance across 32 linguistic tasks. The target distribution consists of 392 dyslexia cases and 3,252 non-dyslexia cases, producing a substantial class imbalance. Three tree-based ensemble learners—Random Forest (RF), Extreme Gradient Boosting (XGBoost) and Extra Trees (ET)—were evaluated. Random-Forest Recursive Feature Elimination (RF-RFE) was additionally investigated as a feature-selection strategy, while probability-level stacking combined predictions from RF, XGBoost and ET. Models were evaluated using accuracy, precision, recall, specificity, F1-score, ROC-AUC and Matthews Correlation Coefficient (MCC). XGBoost produced the strongest individual baseline performance, achieving 90.64% accuracy, 64.41% precision, 29.08% recall, 98.06% specificity, 40.07% F1-score, 0.8845 ROC-AUC and 0.3912 MCC. RF-RFE + XGBoost improved performance to 90.81% accuracy, 31.89% recall, 42.74% F1-score, 0.8858 ROC-AUC and 0.4122 MCC. The heterogeneous stacking model achieved substantially higher recall at 71.17%, together with an F1-score of 51.71% and MCC of 0.4644, although accuracy and specificity declined to 85.70% and 87.45%, respectively. The findings demonstrate that ensemble architecture materially affects the operating characteristics of dyslexia prediction. In particular, accuracy-oriented models and sensitivity-oriented screening models represent different deployment objectives. The study contributes a secondary empirical evaluation of the Rello dataset and demonstrates that heterogeneous ensemble learning can provide a potentially useful sensitivity-oriented screening strategy. Nevertheless, the findings should be interpreted as predictive screening evidence rather than clinical diagnosis, and independent external validation, threshold optimization, calibration and explainable artificial intelligence are required before practical deployment.