This study evaluates seven machine learning algorithms for predicting the academic performance of high school students by training models on a Ghanaian cohort and externally validating them on a U.S. cohort. Following TRIPOD-ML guidelines, the study employed train-only preprocessing steps, including median/mode imputation, one-hot encoding, and scaling for LR/SVM/KNN, along with stratified 5-fold cross-validation and a held-out test set. Decision thresholds were tuned to emphasize recall while maintaining acceptable precision. In Ghana, RF and XGB achieved ROC-AUC and PR-AUC scores of 99%, with accuracy ranging from 96% to 97%, and an F1-score of approximately 97%. The held-out results closely matched the cross-validation findings, with XGB and RF achieving accuracies of 98% and 97%, respectively, both attaining ROC-AUC and PR-AUC scores of 99%. In the external validation cohort, performance remained strong, with RF and XGB achieving approximately 95% accuracy, an F1-score of nearly 97%, and ROC-AUC and PR-AUC scores of 99%, surpassing recent U.S. benchmarks. SHAP and permutation importance consistently identified prior achievement measures, such as BECE results, mock scores, and study hours, as the most influential predictors. Subgroup analyses by gender and region revealed minor, statistically uncertain differences. Overall, tree-based ensemble models offer accurate, well-calibrated, and interpretable risk stratification, making them suitable for early-warning systems that enable timely and targeted support.