The impact of a domain-specific harness on large language model evaluations of medical necessity in oncology.

Abstract

740 Background: While large language models (LLMs) have been effective at medical domain question and answer, their effectiveness and sufficiency in determining the medical necessity (MN) of complex therapeutics in oncology for a given patient’s clinical context is unclear. Recent advances in artificial intelligence have been driven by a transition from LLMs acting independently through chat prompts to agents that can support complex automated processes. Given the recent success of agent harnesses (software infrastructure surrounding an LLM) in coding applications, we hypothesize that constructing an agent system consisting of an LLM with a domain-specific harness will lead to an improved ability to perform evaluations of MN. Methods: We tested our hypothesis by using an LLM to evaluate the MN of initiating pertuzumab for treatment of HER2+ breast cancer in the neoadjuvant, adjuvant, and advanced/metastatic settings in accordance with a payer policy that is grounded in NCCN guidelines. An evaluation dataset consisting of synthetic case notes for 95 breast cancer patients was constructed using Tuple Health’s medical necessity agent environment, which draws on research literature. Each case was manually reviewed to ensure clinical fidelity and to establish ground truth of MN in accordance with the policy. For each patient case, we evaluated MN using the chosen LLM (DeepSeek V3.2) under three scenarios: 1) using the capabilities of the LLM alone, 2) providing the policy text at inference using retrieval augmented generation (RAG), 3) as an agent system consisting of the LLM and a Tuple Health harness for MN evaluation in oncology. Performance was measured by MN evaluation accuracy (pass@1). Results: The baseline accuracy of the LLM was 67.4%. Providing the policy with RAG improved accuracy to 85.3%, a 17.9% absolute improvement over baseline performance. Operating as an agent system increased overall accuracy to 92.6%, a 25.2% absolute improvement over baseline performance. Conclusions: While the performance of LLMs on clinical benchmarks suggests significant capability in the domain, we found that the baseline performance of our chosen LLM to evaluate the MN of initiation of pertuzumab for HER2+ breast cancer was poor (67.4% accuracy). We found that providing the policy as context to the LLM through RAG improved accuracy (85.3%) and that using an agent system composed of a Tuple Health MN harness and the LLM further improved accuracy (92.6%). The poor baseline performance of the LLM and the improvements in performance observed with the agent system (w.r.t. both the baseline and RAG evaluation scenarios) motivate the use of domain-specific harnesses to improve the reliability of LLMs in evaluations of medical necessity in complex domains such as oncology. Evaluation Scenario MN Evaluation Accuracy(Pass@1) LLM only 67.4% (64/95) LLM + RAG 85.3% (81/95) Agent System (MN Harness + LLM) 92.6% (88/95)