Abstract

In healthcare, fine-tuned LLMs offer significant potential to support clinical decision-making, personalised healthcare using patient medical history, medical documentation, patient engagement, and research translation, paving the way for a more efficient way of practising medicine. However, integrating these models into specialised medical domains, such as radiology, oncology, cardiology, or mental health, introduces complex technical, ethical, and regulatory challenges. With careful design, fine-tuned LLMs can become valuable tools that complement clinical expertise and advance modern healthcare, provided that various crucial concerns are mitigated, such as reliability, transparency, and accountability, where incorrect reasoning or interpretation can have serious consequences for patient safety.

Fine-tuned LLMs for specialised medical domains can assist clinicians in synthesising patient records, laboratory results, and unstructured clinical notes into actionable insights efficiently, leading to supporting diagnosis and treatment planning. Maintaining doctor-in-the-loop is essential until the tuning is ensured by the authorities in related medical domains. This work examines the opportunities enabled by domain-specific fine-tuning of LLMs. It also critically analyses the challenges related to hallucination, data quality, safety, explainability, bias, regulatory compliance, and real-world deployment. The work outlines a framework that helps deploy responsible fine-tuned LLMs in specialised medical settings.