Jeff Zhang

BACKGROUND AND OBJECTIVES: Traditional medical board examinations present clinical information in static vignettes with multiple-choices (MC), fundamentally different from how physicians gather and integrate data in practice. Recent advances in large language models (LLMs) offer promising approaches to creating more realistic clinical interactive conversations. However, these approaches are limited in neurosurgery, where patient communication capacity varies significantly and diagnosis heavily relies on objective data such as imaging and neurological examinations. We aimed to develop and evaluate a multi–artificial intelligence (AI) agent conversation framework for neurosurgical case assessment that enables realistic clinical interactions through simulated patients and structured access to objective clinical data. METHODS: We developed a framework to convert 608 Self-Assessment in Neurological Surgery first-order diagnosis questions into conversation sessions using 3 specialized AI agents: patient AI for subjective information, system AI for objective data, and clinical AI for diagnostic reasoning. We evaluated generative pretrained transformer 4o's (GPT-4o's) diagnostic accuracy across traditional vignettes, patient-only conversations, and patient + system AI interactions, with human benchmark testing from 10 neurosurgery residents. RESULTS: GPT-4o showed significant performance drops from traditional vignettes to conversational formats in both MC (89.0%-60.9%, P < .0001) and free-response scenarios (78.4%-30.3%, P < .0001). Adding access to objective data through system AI improved performance (to 67.4%, P = .0015; and 61.8%, P < .0001, respectively). Questions requiring image interpretation showed similar patterns but lower accuracy. Residents outperformed GPT-4o in free-response conversations (70.0% vs 28.3%, P = .0030) using fewer interactions and reported high educational value of the interactive format. CONCLUSION: This multi-AI agent framework provides both a more challenging evaluation method for LLMs and an engaging educational tool for neurosurgical training. The significant performance drops in conversational formats suggest that traditional MC testing may overestimate LLMs' clinical reasoning capabilities, while the framework's interactive nature offers promising applications for enhancing medical education.

Hospitals and healthcare systems rely on operational decisions that determine patient flow, cost, and quality of care. Despite strong performance on medical knowledge and conversational benchmarks, foundation models trained on general text may lack the specialized knowledge required for these operational decisions. We introduce Lang1, a family of models (100M-7B parameters) pretrained on a specialized corpus blending 80B clinical tokens from NYU Langone Health's EHRs and 627B tokens from the internet. To rigorously evaluate Lang1 in real-world settings, we developed the REalistic Medical Evaluation (ReMedE), a benchmark derived from 668,331 EHR notes that evaluates five critical tasks: 30-day readmission prediction, 30-day mortality prediction, length of stay, comorbidity coding, and predicting insurance claims denial. In zero-shot settings, both general-purpose and specialized models underperform on four of five tasks (36.6%-71.7% AUROC), with mortality prediction being an exception. After finetuning, Lang1-1B outperforms finetuned generalist models up to 70x larger and zero-shot models up to 671x larger, improving AUROC by 3.64%-6.75% and 1.66%-23.66% respectively. We also observed cross-task scaling with joint finetuning on multiple tasks leading to improvement on other tasks. Lang1-1B effectively transfers to out-of-distribution settings, including other clinical tasks and an external health system. Our findings suggest that predictive capabilities for hospital operations require explicit supervised finetuning, and that this finetuning process is made more efficient by in-domain pretraining on EHR. Our findings support the emerging view that specialized LLMs can compete with generalist models in specialized tasks, and show that effective healthcare systems AI requires the combination of in-domain pretraining, supervised finetuning, and real-world evaluation beyond proxy benchmarks.