Democratizing cost-effective, agentic artificial intelligence to multilingual medical summarization through knowledge...

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Healthcare Technology

Background:

• The demand for multilingual AI in healthcare is growing, particularly for underrepresented languages like Arabic.
• Linguistic complexities in Arabic pose challenges for existing large language models (LLMs) in specialized tasks such as medical summarization.

Purpose of the Study:

• To develop a domain-specific AI agent, AraSum, for summarizing Arabic clinical documentation.
• To address the limitations of foundational LLMs in handling Arabic's linguistic nuances for medical tasks.

Main Methods:

• Utilized a novel knowledge distillation framework to create a small language model (SLM) from a large multilingual LLM.
• Trained AraSum on a synthetic dataset of Arabic medical dialogues.
• Employed BLEU, ROUGE scores, and Arabic-speaking evaluator assessments for performance evaluation.

Main Results:

• AraSum demonstrated superior performance compared to the foundational Arabic LLM, JAIS-30B, on key metrics.
• Evaluators rated AraSum higher in accuracy, comprehensiveness, and clinical utility.
• AraSum achieved these results with significantly reduced computational and environmental costs.

Conclusions:

• AraSum offers an efficient and effective solution for Arabic medical text summarization.
• The study highlights the potential of SLM-based agents for advancing multilingual healthcare and sustainable AI.
• This approach promotes equity in healthcare access by supporting resource-efficient AI in low-resource settings.