DOMAIN-SPECIFIC LARGE LANGUAGE MODELS IN CLINICAL DECISION-MAKING: AN EMPIRICAL ANALYSIS OF MED-PALM AND THE FUTURE OF MEDICAL AI

Abstract

The rapid evolution of large language models (LLMs) has precipitated unprecedented opportunities for their application in the healthcare domain. This paper presents a comprehensive empirical and analytical review of domain-specific LLMs, with particular emphasis on Medical Pathways Language Model (Med-PaLM) and its successor, Med-PaLM 2, Google’s clinically-oriented generative AI systems. Using the United States Medical Licensing Examination (USMLE) benchmark as a primary performance metric, we examine the progressive advancements from general-purpose LLMs to medically fine-tuned models. Med-PaLM 2 achieved an accuracy of 86.5% on USMLE-style questions, representing a 19% improvement over Med-PaLM v1 and approximating expert physician-level performance. We further evaluate training strategies including instruction fine-tuning, chain-of-thought (CoT) prompting, self-consistency, and ensemble refinement (ER), alongside qualitative clinician-led evaluation frameworks. The paper addresses challenges in clinical integration, including multimodal data requirements, regulatory constraints, and the necessity of prospective validation. Our findings suggest that while domain-specific LLMs demonstrate transformative potential in medical question-answering, clinical decision support, and patient engagement, their safe and effective deployment requires rigorous multi-step evaluation, cross-functional collaboration, and a human-centric design philosophy. This work contributes novel insights for researchers, clinicians, and health technology developers pursuing responsible AI innovation in medicine.