Arabic Symptom Classification and Diagnosis Using Transformer Models and LLM-Based Augmentation

Arabic-speaking people have significant barriers to accessing intelligent medical devices, especially for symptom assessment and specialty routing. While transformer-based language models improve healthcare natural language processing, these gains have yet to be translated into Arabic, owing to a lack of annotated data and domain-specific resources. The research proposes an Arabic end-to-end medical triaging system that use a dual-task-fine-tuned Arabic large language model ( LLM) to generate a preliminary diagnosis in tandem with a dual-task transformer-based classifier. Users can interact with the system using native Arabic symptom descriptions and a chatbot to discuss their problems. The classifier will predict both the severity of the case and the right medical specialization, while the LLM will provide contextual diagnostic input. To address this shortage of training data, we collected and categorized 1,500 real-world Arabic symptom descriptions from social media, increasing the sample count to 4,000 by using Gemini Flash 2.0 for semantic rephrasing. This resulted in improved language diversity while maintaining label integrity. The experimental findings demonstrate the success of this method, with AraBERTv0.2 achieving the highest post-augmentation F1 score of 93.88%, followed by CAMeLBERT at 90.97%. The system demonstrates the ability to generate accurate low-resource Arabic clinical NLP models through guided data augmentation and transformer fine-tuning.