Early dyslexia prediction is difficult because of its varied behavioral signals and the need for timely, reliable screening in educational settings. This study proposes a soft-voting ensemble framework that integrates multiple boosting classifiers (AdaBoost, Gradient Boosting, XGBoost, and CatBoost) to improve predictive performance while maintaining interpretability. The model is evaluated using stratified 10-fold cross-validation on a publicly available dataset derived from gamified cognitive assessments, with minimum Redundancy Maximum Relevance feature selection and SHapley Additive exPlanations analysis used to enhance transparency. Experimental results reveal that the ensemble outperforms individual models and recent methods, achieving 91.6% accuracy, 90.7% precision, 91.6% recall, and a 90.9% F1-score using 165 selected features with class weighting, thereby supporting its practicality for explainable early dyslexia screening.