Ensemble Transformers for Arabic Medical Text Classification

The development of social telehealth frameworks has integrated new dimensions of interaction into healthcare, enabling patients to discuss symptoms and obtain remote consultations. Such transformation has paved numerous streams of biomedical data, particularly in Arabic due to its rich morphology and scarcity of annotated resources. In this study, we address the problem of Arabic biomedical question classification (ABQC) using the MAQA dataset, with a focus on a carefully curated subset of ten critical medical categories. We implemented a single training pipeline for five transformer models: AraBERT, BioBERT, DistilBERT, mBERT, and XLM-RoBERTa to ensure they were evaluated consistently. To improve the reliability of the predictions to be both accurate and stable, we utilized an ensemble learning method based on majority voting of individual model outputs. Our ensemble approach achieved the highest classification performance with 92.62% Accuracy and 92.55% F1-score. These results are a step forward in the application of terminology and language resources towards Arabic biomedical text processing and validation of the assumption that ensemble techniques enhance the dependability of responses in telehealth question answering systems.