A S.C.O.R.E. framework for evaluating open-ended responses from large language models in healthcare

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Natural Language Processing

Background:

• Evaluating Large Language Models (LLMs) in healthcare requires domain-specific clinical validation.
• Existing quantitative metrics often fail to accurately assess the clinical appropriateness of LLM-generated responses.
• A structured, expert-driven framework is needed for reliable LLM evaluation in medical contexts.

Purpose of the Study:

• To introduce S.C.O.R.E. (Safety, Consensus & Context, Objectivity, Reproducibility, Explainability), a novel five-dimensional framework for expert evaluation of LLM healthcare outputs.
• To validate S.C.O.R.E. against quantitative metrics and assess its reliability and utility in clinical settings.
• To demonstrate the necessity of domain-specific tuning for LLMs in healthcare and its detectability via expert evaluation.

Main Methods:

• Developed the S.C.O.R.E. framework encompassing Safety, Consensus & Context, Objectivity, Reproducibility, and Explainability.
• Validated S.C.O.R.E. against quantitative metrics (BLEU, ROUGE, BERTScore) using three LLMs (GPT-4o, Claude 4 Sonnet, DeepSeek) across ophthalmology, medication, and anesthesia domains.
• Assessed S.C.O.R.E.'s internal consistency (Cronbach's α) and effect sizes (Cliff's δ) in optimized and non-optimized domains.

Main Results:

• S.C.O.R.E. demonstrated acceptable internal consistency (Cronbach's α = 0.745) and detected significant differences (Cliff's δ = 0.68–0.92) related to model optimization status.
• Quantitative metrics frequently misclassified clinically appropriate LLM responses, highlighting their limitations.
• LLM performance varied significantly across specialties, with GPT-4o excelling in ophthalmology (optimized) while others performed better in non-optimized domains, underscoring the need for domain-specific tuning.

Conclusions:

• S.C.O.R.E. provides a reliable and practical framework for the clinical validation of LLM-generated healthcare responses.
• Expert evaluation using S.C.O.R.E. is superior to quantitative metrics for assessing clinical appropriateness and guides iterative model refinement.
• Domain-specific optimization is crucial for LLM performance in healthcare, and S.C.O.R.E. effectively detects this, supporting regulatory compliance and safe deployment.