Diagnostic Accuracy of GPT-4o and Claude for HEART Score Calculation in Chest Pain (LLM-HEART)

AI SYSTEM SPECIFICATIONS AND PROMPT PROTOCOL Two distinct large language models (LLMs) will be evaluated as index tests: OpenAI GPT-4o (model string: gpt-4o-2024-11-20) and Anthropic Claude (model string: claude-sonnet-4-6). To ensure reproducibility and eliminate stochastic variation, both models will be accessed via standardized API calls using deterministic parameters (temperature = 0, max_tokens = 500, and strict JSON response format). The exact system prompt layout will be locked prior to initialization, and its integrity will be verified using a SHA-256 cryptographic hash. The models will evaluate each patient record independently in zero-shot isolation, with no cross-contamination or conversational history retention between runs.

REFERENCE STANDARD CONSENSUS PROTOCOL The reference standard consists of a structured consensus HEART score established by three independent emergency medicine physicians (each possessing >=3 years of clinical experience and specific training on HEART score criteria). The physicians will review the anonymized clinical charts while remaining strictly blinded to the LLM outputs and the final 30-day MACE outcomes. For each of the 5 HEART components (scored 0, 1, or 2), a majority vote (2/3 agreement) will determine the final component score. In the event of complete disagreement across all three reviewers on a specific component, a fourth independent adjudicator will resolve the tie.

INDETERMINATE RESULTS MANAGEMENT

In strict compliance with STARD-AI 2025 guidelines, cases with missing or uninterpretable parameters within the free-text clinical notes will be classified into predefined indeterminate tiers:

1. Complete Cases: 0 indeterminate components (eligible for primary diagnostic accuracy analysis).
2. Partial Indeterminate: Exactly 1 missing component preventing definitive automatic calculation.
3. Full Indeterminate: >=2 missing components. The proportion of indeterminate classifications will be quantified for both LLMs and evaluated alongside the routine documentation quality of the charts.

STATISTICAL ANALYSIS AND AGREEMENT WEIGHTING Statistical power and sample size calculation are based on the Hanley-McNeil methodology for the Area Under the ROC Curve (AUC). To achieve an expected AUC of 0.85 with a non-inferiority margin of 0.05, a power of 80%, and a two-sided alpha of 0.05, the primary complete-case analysis requires 600 evaluable patients. Accounting for an anticipated 15% indeterminate rate, a total enrollment target of 690 patients is set. Inter-rater agreement between each LLM and the expert consensus will be computed using quadratic weighted Cohen's Kappa for the ordinal total HEART score (0-10) and linear weighted Kappa for individual components (0-2). Diagnostic performance metrics (sensitivity, specificity, PPV, NPV) will be calculated at prespecified binary (>=4) and trimodal thresholds with 95% Wilson confidence intervals. Pairwise comparison of AUC values between GPT-4o and Claude will be executed using the DeLong test.

DATA ANONYMIZATION AND PRIVACY To ensure full compliance with local personal data protection legislation (KVKK), all free-text emergency department notes will undergo strict de-identification. Patient names, institutional ID numbers, precise dates, and specific demographic identifiers will be stripped entirely before formatting the data payload for API transmission.

PATIENT AND PUBLIC INVOLVEMENT BEYANI Patient and public involvement was not applicable to this study as it involves the analysis of routinely collected clinical data.