Deployment-Centered Evaluation: Predicting Query-Level Rejection Risk in a Clinical LLM System

2026-06-10

AI Summary

Existing static evaluation methods struggle to predict the risk that a single query from a clinical large language model will be rejected by users in real-world deployment and often rely on densely annotated data. This work proposes a deployment-oriented, prospective evaluation framework that, for the first time, integrates deployment context—such as physician type, department, and model version—with query content to train a pre-response classifier capable of fine-grained prediction of user rejection behavior. Leveraging 4.5 months of real-world electronic health record system logs and sparse user feedback, the model achieves an AUROC of 0.719 and demonstrates promising utility in dynamic intervention tasks, including triggering safety mechanisms and model abstention.

Abstract

Large language models (LLMs) are increasingly integrated into clinical systems, making it essential to evaluate the real-world utility of these systems. However, static benchmarks tend to measure correctness rather than user acceptance, aggregate performance across queries, and require densely annotated datasets -- leading to major blind spots for evaluating clinical systems. In this work, we perform a deployment-centered evaluation of an LLM system embedded within electronic health records at an academic medical center, where user feedback is sparse but closely reflects the deployment conditions. Specifically, we train a pre-response classifier that estimates the risk that a future interaction will result in the user rejecting the LLM response, based on query content and deployment-specific context available before generation. We conduct a prospective analysis of our model over 4.5 months of user feedback, finding that our prediction model achieves an AUROC of 0.719. Further, we estimate the benefit of such predictions in two downstream use cases (guardrail triggering and abstention). Our key conceptual insight is that making use of deployment-specific context (i.e., the provider type, department name, language model used for response), as opposed to only query content, improves the ability to predict whether the user will reject the system output. Altogether, our empirical case study demonstrates the feasibility of predicting user rejection using deployment-specific context, opening the door to targeted guardrails.

Problem

• clinical LLM
• user rejection
• deployment-centered evaluation
• query-level risk prediction
• real-world utility

Innovation

• deployment-centered evaluation
• query-level rejection prediction
• clinical LLM systems
• pre-response classifier
• deployment-specific context