Benchmarking large language model-based agent systems for clinical decision tasks

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Computational Medicine

Background:

• Agentic AI systems, capable of autonomous reasoning and tool use, show potential in healthcare applications.
• Systematic real-world performance evaluation of these advanced AI systems in medicine is currently limited.
• Existing benchmarks do not fully capture the complexities of clinical decision-making and tool integration.

Purpose of the Study:

• To systematically benchmark the real-world performance of two agentic AI systems in healthcare settings.
• To evaluate the efficacy of agentic AI across diverse medical tasks, including diagnostics, QA, and complex examinations.
• To assess the trade-offs between performance gains, resource utilization, and hallucination rates in medical AI agents.

Main Methods:

• Evaluated OpenManus (Llama-4 based) and Manus (proprietary multistep architecture) on AgentClinic, MedAgentsBench, and Humanity's Last Exam (HLE) benchmarks.
• Assessed performance on text-based and multimodal medical question-answering and diagnostic simulations.
• Quantified accuracy, token usage, latency, and hallucination rates, with in-agent safeguards.

Main Results:

• Agentic AI systems provided modest accuracy improvements over baseline LLMs, with significant increases in token usage and latency.
• Accuracy on AgentClinic MedQA reached 60.3%, MedAgentsBench 30.3%, and HLE text 8.6%.
• Multimodal accuracy was low (15.5% on HLE, 29.2% on AgentClinic NEJM), and hallucinations persisted despite safeguards.

Conclusions:

• Current agentic AI designs offer limited performance benefits in healthcare relative to their substantial computational and workflow costs.
• There is a critical need for the development of more accurate, efficient, and clinically viable agent systems for medical applications.
• Further research is required to optimize agentic AI architectures for practical healthcare deployment.