Benchmarking Reliability Metrics for Large Language Models in Computational Health Applications

Abstract

The deployment of large language models (LLMs) in computational health applications demands rigorous evaluation frameworks capable of quantifying not only predictive accuracy but also output reliability, calibration, and uncertainty. Despite rapid advances in model capability, the absence of standardized benchmarking protocols for reliability metrics represents a critical gap that impedes safe clinical translation. This work addresses that gap through a systematic comparative study of reliability evaluation methodologies applied to LLMs across diverse health-related natural language processing tasks.

We introduce a comprehensive benchmarking suite encompassing six reliability dimensions: calibration error, selective prediction performance, distributional shift robustness, semantic consistency, factual grounding fidelity, and conformal coverage guarantees. Formally, for a model producing probability estimates $\hat{p}$ over label space $\mathcal{Y}$, we define reliability as a composite functional $\mathcal{R}(\hat{p}, y) = \sum_{k=1}^{K} \lambda_k \, \phi_k(\hat{p}, y)$, where $\phi_k$ denotes each constituent metric and $\lambda_k$ its associated weight determined through task-specific sensitivity analysis.

Experiments are conducted across clinical note summarization, medical question answering, diagnostic coding, and adverse event extraction benchmarks, evaluating seven state-of-the-art LLMs under both zero-shot and fine-tuned regimes. Our results reveal substantial inter-metric disagreement, demonstrating that accuracy-centric evaluation systematically overestimates model trustworthiness in high-stakes health contexts. Calibration and conformal prediction metrics expose failure modes invisible to standard performance measures.

This study provides actionable recommendations for practitioners and regulators, advocating for multi-dimensional reliability reporting as a prerequisite for responsible LLM deployment in healthcare. All benchmarking code and evaluation protocols are released publicly to facilitate reproducibility and community adoption of standardized reliability assessment practices.