HealthFlow: A Self-Evolving AI Agent for Autonomous Healthcare Research via Meta-Level Strategic Learning

Abstract

The rapid growth of biomedical literature and electronic health record (EHR) data presents a critical challenge: transforming vast, unstructured information into actionable scientific insight. Existing AI agent frameworks rely on static, human-designed strategies, limiting their ability to adapt to complex and evolving healthcare research workflows. This paper introduces HealthFlow, a self-evolving AI agent that learns to refine its high-level problem-solving strategies through experience. HealthFlow employs a multi-agent architecture with meta-planning, execution, evaluation, and reflection components, enabling the agent to distill successes and failures into structured procedural knowledge that continuously improves future planning. To rigorously evaluate autonomous healthcare research capabilities, we also introduce EHRFlowBench, a new benchmark consisting of complex, evidence-grounded research tasks derived from peer-reviewed scientific literature. Extensive experiments across multiple healthcare agent benchmarks demonstrate that HealthFlow significantly outperforms state-of-the-art agent frameworks in task success, robustness, and efficiency. These results highlight the importance of meta-level strategic learning and establish HealthFlow as a step toward more autonomous and effective AI systems for healthcare research.

Pegah Khorasani

Graduate Research Associate

As a doctoral candidate at Arizona State University (ASU), I am currently conducting research under the guidance of Professor Hassan Ghasemzadeh at the Embedded Machine Intelligence Lab (EMIL). My research interests encompass a range of topics, including machine learning, health monitoring system development, and mobile health.