Title: The Impact of Scoring Protocols on AI Rater Bias in Complex Clinical Decisions

Abstract

As the evaluation of clinical AI increasingly relies on large language models (LLMs) functioning as AI raters, there remains a lack of quantitative characterization regarding their scoring behaviors across different evaluation conditions. To fill this void, we conducted a factorial study examining AI rater performance in the context of adult type 2 diabetes (T2D) pharmacotherapy during 12-month outpatient follow-ups—a clinical task demanding complex decision-making, operationalized through seven distinct evaluation questions. In this study, four open-source LLMs acted concurrently as both clinical decision support system (CDSS) models and AI raters.

Each CDSS output was assessed using two distinct scoring protocols: the Gold Rubric (GR) protocol, which utilized a patient-specific rubric, and the Non-Gold Rubric (Non-GR) protocol, which was rubric-free. We employed linear mixed effects models to analyze the interaction between the scoring protocol and five design factors: the CDSS model, CDSS prompt configuration (comparing document-referenced generation [DRG] against Baseline), the rater model, prompt character, and prompt type. The analysis estimated both main effects and their interactions with the protocol.

The results indicated that under the Non-GR protocol, AI raters produced consistently higher scores within a narrow band (averaging 74–78 points). In contrast, scores under the GR protocol were significantly lower, ranging from 7.69 to 49.64 points less on average, with interquartile ranges that were 1.68 to 3.67 times wider. Furthermore, within individual questions, the GR protocol enhanced the AI rater’s ability to discriminate between DRG and Baseline CDSS outputs by factors of 1.76 to 5.10. This protocol also exposed significant behavioral variations across different rater models, a nuance that was masked by the Non-GR approach.

These findings suggest that rubric anchoring is essential for maintaining discriminative power in clinical AI evaluation. Rubric-free scoring is inadequate for questions that necessitate patient-specific or jurisdiction-specific criteria, as rater models cannot deduce these requirements from parametric knowledge alone.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC