MedMI-Bench: Pre-Registered Analysis Plan for Supplementary Experiments

Le résumé fourni par la source

Pre-registered analysis plan for three supplementary experiments extending the MedMI-Bench physician audit of LLM-generated medication decision benchmarks derived from MIMIC-IV. This document commits to hypotheses, sample sizes, statistical tests, and decision rules before data collection to protect the analysis from post-hoc selection effects. The three experiments are: 1. A position-shuffling causal test on 2,000 consensus-accepted probes. The original 34,201-probe pool analysis showed an observational position bias (13.8% consensus acceptance for position A vs 16.0% for position D, P=.002). This experiment shuffles the correct-answer position while holding clinical content constant and re-runs all three validators (GPT-4o, Claude Sonnet 4.6, Gemini 2.5 Flash) to determine whether the effect is causally driven by position or confounded with content. 2. A 100-probe physician audit of the consensus-rejected pool (n=16,386), matching the 100-probe accepted-pool audit from the main paper, stratified by task type and department. This enables bidirectional characterization of validator error (false-acceptance rate vs false-rejection rate) using matched sample sizes and equivalent confidence interval widths. 3. A cross-validator disagreement analysis on the full 34,201-probe pool, computing pairwise Cohen's kappa between validator pairs, sole-dissenter identity distributions, and a correlated-error test comparing observed joint-error counts against the product of individual error rates under the statistical independence null. Decision rules for paper integration, deviation policies, and null-result reporting commitments are specified in the attached document. This pre-registration accompanies the main MedMI-Bench paper targeting JMIR AI and will be cited in the paper's Methods section with its Zenodo DOI. Author: Aman Sharma (ORCID: 0009-0005-5107-4485)Contact: Aman_sharma007@yahoo.comMain paper target venue: JMIR AIPre-registration timestamp: April 2026 (before experiment execution)

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.