PSEBench: A Controllable and Verifiable Benchmark for Evaluating LLMs in Patient Safety Event Triage

Researchers introduced PSEBench, a 5,074-case benchmark dataset designed to evaluate large language models on patient safety event triage—the critical task of determining whether clinical incidents require reporting under regulatory policy. The methodology uses policy-grounded clause cards and verification mechanisms to ensure reliable evaluation of LLM reasoning, information-seeking behavior, and appropriate abstention in ambiguous cases.

PSEBench addresses a significant gap in AI evaluation infrastructure for high-stakes healthcare applications. Patient safety event triage remains largely manual despite its routine nature, creating bottlenecks in clinical workflows while introducing human error risks. The benchmark's innovative clause card methodology—which decomposes regulatory language into auditable decision specifications—enables systematic testing of whether LLMs can reason about complex policy requirements with the precision demanded in healthcare settings.

The development reflects broader recognition that general-purpose LLM benchmarks fail to capture domain-specific reasoning patterns. Healthcare regulation requires not just pattern matching but genuine policy comprehension, along with the critical ability to abstain when evidence is insufficient. Traditional benchmarks lack mechanisms to evaluate this safety-critical behavior. PSEBench's closed-loop verification pipeline and agentic evaluation environment represent methodological advances that could inform benchmark design across regulated industries.

Evaluation across 15 representative LLMs establishes capability baselines while exposing consistent gaps. These findings matter for healthcare organizations considering LLM deployment. The benchmark provides actionable evidence about which models meet minimum reliability standards and where additional safeguards are necessary. Healthcare IT vendors and patient safety software companies now have concrete data to guide model selection and integration strategies.

Longer term, PSEBench establishes a replicable framework for policy-grounded evaluation applicable to other compliance domains—insurance claims processing, regulatory reporting, financial fraud detection. The methodology demonstrates that rigorous AI evaluation in regulated contexts requires domain-specific infrastructure beyond generic leaderboards. Healthcare systems implementing AI governance frameworks can reference this benchmark as evidence of due diligence in model assessment.

• →PSEBench provides the first large-scale benchmark specifically designed for evaluating LLM performance on patient safety event triage with 5,074 policy-grounded test cases.
• →The clause card methodology breaks regulatory text into auditable decision specifications, enabling systematic evaluation of policy reasoning and appropriate model abstention.
• →Evaluation of 15 LLMs reveals consistent capability gaps, providing healthcare organizations with data-driven guidance for safe model deployment decisions.
• →The benchmark's closed-loop verification ensures ground truth by construction, addressing a fundamental limitation in evaluating high-stakes healthcare AI tasks.
• →The policy-grounded evaluation framework is replicable across other regulated industries requiring compliance reasoning and principled decision-making.