A multi-level classification model for patients’ medical questions using pretrained transformers, graph attention networks and bidirectional LSTMs

The rapid advancement of medical services on various online platforms highlights the rising demand for intelligent models that can support users of these services. One such model is the classification of patients’ needs based on their questions, which can significantly improve the quality of service provided. Several models have been developed to categorize Arabic medical questions, based on various deep learning (DL) architectures and techniques to address the challenges posed by the complexity of the Arabic language. Despite significant progress, existing studies in the broad literature have focused on classifying Arabic medical questions at a single level. To address this gap, this paper presents a new model for the multi-level classification of patients’ questions in Arabic. The developed model uses various DL architectures and mechanisms, including graph attention networks (GATs), pretrained transformers, Bidirectional LSTMs (BiLSTMs), and multi-head attention, to overcome the challenges related to the complexity of the Arabic language and the hierarchical structures of medical specialties and disciplines. The model performance is evaluated using both standard and hierarchical metrics, along with two calibration measures and confidence intervals via stratified bootstrapping to strengthen statistical validity. Moreover, the model is trained and validated on two datasets annotated according to a predefined taxonomy. The findings demonstrate the key role of each model component across an ablation study, as well as its superiority over baseline models developed in the literature. In particular, the best configuration combines the CAMeLBERT transformer, GATs, and BiLSTM with multi-head attention, achieving performance exceeding 88% and 76% across all metrics at the first and second levels, respectively. Additionally, formal complexity and error cluster analysis demonstrate the substantial practicality of the developed model for clinical platforms.