Abstract
Abstract
English Language Anxiety (ELA) remains a significant barrier to academic success, particularly for students in regions where English is not the most spoken language. This study presents a lightweight and interpretable machine learning (ML)-based approach for the early detection of ELA among rural high school students in Bangladesh, a low-income, non-English-speaking country. Five supervised learning algorithms, Support Vector Machine (SVM),
k
-Nearest Neighbors (
k
-NN), Naïve Bayes, Decision Tree, and Random Forest, were evaluated on 2,600 student records. Foreign Language Classroom Anxiety Scale (FLCAS) scores and teacher assessments were used only to construct ground-truth ELA labels, whereas the predictive models were trained exclusively on academic and demographic features. The Decision Tree achieved the highest accuracy (99.36%), with similarly strong performance across precision, recall, and
$$\:{F}_{1}$$
-score, using the holdout evaluation method. Stratified 10-fold cross-validation produced a mean accuracy of 99.15 ± 0.28%, indicating stable internal performance at the student-record level. SHAP was used to quantify the contribution of the actual model inputs, including English performance, other-subject performance, class level, school-wide English performance, and gender. An independent qualitative evaluation by 15 experienced English teachers provided contextual support for the model’s findings. These results suggest that interpretable ML can provide useful decision support for ELA screening in resource-constrained rural educational settings, while external validation across unseen schools and districts remains necessary.