Beyond Accuracy: An Efficiency- and Safety-Aware Framework for Evaluating Clinical AI with Large Language Models
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Large language models (LLMs) demonstrate strong performance on medical reasoning tasks, but current evaluation approaches focus primarily on accuracy, neglecting the efficiency–safety trade-offs critical for real-world clinical utility.
Methods
We developed and validated the Clinical Value Density (CVD) framework, a novel metric quantifying clinical utility per unit of cognitive resource consumed. Six state-of-the-art LLMs (GPT-4o, Gemini-2.5, Claude-Sonnet-4, Grok-3, DeepSeek-R1, and Kimi-K2) were evaluated across 60 authentic clinical pharmacology scenarios derived from UAE healthcare practice and benchmarked against board-certified clinical pharmacist responses. Performance was assessed across efficiency, semantic similarity, safety, relevance, consistency, and conciseness, with triangulated validation against clinician preference and task efficiency.
Results
Traditional metrics (BLEU: 0.003–0.024; ROUGE-L: 0.079–0.166) failed to reflect clinical utility, while the CVD framework exposed critical efficiency–safety trade-offs. GPT-4o achieved the highest normalized CVD (0.475), delivering fourfold efficiency gains over pharmacists (41 vs 178 tokens) but with moderate safety (0.437), requiring supervised deployment. Grok-3 and Gemini-2.5 led in safety (0.605 and 0.542) but at the cost of efficiency. DeepSeek-R1 and Kimi-K2 produced unsafe brevity–accuracy trade-offs, generating concise but clinically inaccurate responses. Comparative validation revealed strong alignment between CVD and task efficiency (r = 0.924), but divergence from clinician preference (r = –0.845), reflecting a bias toward verbose outputs.
Conclusions
Current LLMs cannot reliably perform clinical functions autonomously. Instead, they are best positioned for AI-assisted, supervised integration, where efficiency and safety are balanced under professional oversight. The CVD framework provides a potentially useful for regulatory evaluation for quantifying deployment readiness, aligning AI evaluation with the real-world constraints of cognitive load, time, and patient safety. Future research should extend CVD across specialties, scale validation datasets, and conduct real-time workflow trials to establish specialty-specific safety thresholds for eventual autonomy.
Author Summary
Large language models (LLMs) are impressively accurate on medical reasoning benchmarks, yet clinical work is not a quiz, it is a race against time under strict safety constraints. We introduce Clinical Value Density (CVD), a simple yet powerful metric that captures clinical utility per unit of cognitive cost (i.e., how much useful care an answer delivers for the effort it imposes on clinicians). In a head-to-head evaluation of six state-of-the-art LLMs across 60 authentic clinical pharmacology scenarios benchmarked against board-certified pharmacists, traditional metrics (BLEU/ROUGE) barely moved, while CVD exposed the true efficiency–safety trade-offs that determine bedside usefulness. For example, GPT-4o delivered the highest normalized CVD (0.475) with ∼4× fewer tokens than pharmacists (≈41 vs 178), but only moderate safety (0.437) appropriate for supervised use, not autonomy. Conversely, Grok-3 and Gemini-2.5 were safest but less efficient; DeepSeek-R1 and Kimi-K2 risked unsafe brevity– accuracy trade-offs. Crucially, CVD strongly tracked task efficiency (r=0.924) yet diverged from clinician preference (r=–0.845), revealing a bias toward overly verbose answers that feel reassuring but slow care.
