Introduction
In recent years, breakthroughs in large language models (LLMs) have driven their widespread application in the medical field, such as medical question answering, electronic health record generation, and clinical decision support. At the same time, LLM-based AI agents are rapidly emerging. In clinical practice, healthcare professionals often face multimodal and highly heterogeneous data, heavy workloads, and time-critical clinical decision-making demands. AI agents can not only understand and generate human language, but also autonomously orchestrate multi-step tasks through tool invocation, demonstrating goal-oriented reasoning and decision-making capabilities, and are therefore regarded as a frontier in medical technology.
Existing studies have preliminarily explored the potential of AI agents in healthcare. For example, Qiu et al. explored their applications in diagnostic support and workflow optimization, while also pointing out challenges such as data privacy and over-reliance. Karunanayake analyzed the core functions of agentic AI in diagnosis, clinical operations, drug development, and robot-assisted intervention. Moritz et al. proposed a paradigm for coordinating multi-agent systems, emphasizing the potential of decentralized yet interoperable LLM-driven agents in optimizing clinical and operational workflows, and elaborated on implementation challenges such as secure communication, interoperability, and clinical validation. However, compared with the rapidly growing LLM medical literature, dedicated research on LLM-based AI agents remains limited, and existing reviews are insufficient in breadth, evaluation depth, and theoretical frameworks. This review aims to fill these gaps and provide a more comprehensive and structured perspective on the development and deployment of AI agents in healthcare.
Historical Evolution of AI Agents
The concept of the "agent" can be traced back to philosophical thought, spanning a journey from theoretical speculation to technological implementation. Ancient Greek philosophers had already begun to describe entities with desires, beliefs, intentions, and the capacity for action. Aristotle's "teleology" provided the philosophical foundation for the goal-oriented characteristics of later agents.
Entering the contemporary era, with the development of natural sciences and computer technology, artificial intelligence research shifted from philosophical speculation to practical application. In the 1950s, the Turing test became an important criterion for evaluating machine intelligence. In the 1970s, expert systems used human expert knowledge for reasoning and decision-making. The emergence of machine learning technology enabled agents to acquire knowledge and skills from data, significantly improving their intelligence. In the 21st century, deep learning achieved major breakthroughs in perception, decision-making, and execution capabilities, expanding application scenarios. Reinforcement learning, especially multi-agent reinforcement learning (MARL), made progress in sequential decision-making problems, enabling agents to make better decisions in complex environments.
After 2022, the proliferation of LLMs opened new avenues for the development of agents. Compared with reinforcement learning agents, LLM-based AI agents possess richer knowledge bases, more natural human-computer interaction capabilities, and better interpretability. For example, organizations such as OpenAI have continuously released more powerful models, accelerating the development of agents.
Core Application ScenariosThe application of LLM-based AI agents in healthcare covers multiple domains:
Assisted diagnosis: Agents analyze patient symptoms, medical history, and examination data to provide diagnostic suggestions, helping doctors reduce missed diagnoses.
Clinical decision support: In treatment selection, medication recommendations, and risk prediction, agents can provide personalized decision support based on the latest evidence and patient data.
Medical report generation: Automatically generate documents such as radiology reports and discharge summaries, reducing physicians' documentation burden and improving record accuracy.
Patient chatbots: Provide patients with 24/7 health consultations, preliminary symptom screening, and appointment guidance, improving patient experience.
Healthcare system management: Optimize operational processes such as hospital resource allocation, bed management, and staff scheduling to improve efficiency.
Medical education: Simulate clinical scenarios to assist medical students in diagnostic training and skill improvement, offering an interactive learning experience.
These applications demonstrate the great potential of AI agents in improving healthcare quality, efficiency, and accessibility.
Evaluation Framework
Establishing a multi-dimensional evaluation framework is essential to ensure the safety and effectiveness of AI agents. The evaluation should cover the following key dimensions:
Clinical performance: Including diagnostic accuracy, correctness of decision recommendations, accuracy of report generation, etc.
Safety and reliability: Evaluate agent behavior in edge cases to avoid harmful recommendations or hallucinated outputs.
User experience: Including satisfaction of doctors and patients, naturalness of interaction, trust, etc.
Ethics and fairness: Examine algorithmic bias, privacy protection, transparency, and accountability.
Efficiency and cost: Measure the time savings and resource utilization efficiency that agents bring to real-world workflows.
A comprehensive evaluation should combine quantitative metrics (such as precision, recall, and F1 score) with qualitative assessments (such as expert review and user surveys). In addition, validation in real clinical environments is needed to ensure generalization capability.
Future Development Directions
Looking to the future, we propose seven key development directions:1. Integration with Physical Systems: Combine AI agents with physical systems such as robots and wearable devices to achieve a closed loop from virtual decision-making to physical execution.
2. Mixture of Experts Models: Integrate general-purpose LLMs with domain-specific models to enhance performance and reliability in specialized medical tasks.
3. Expanded Evaluation Paradigms: Develop dynamic evaluation methods that are closer to clinical practice and cover long-tail scenarios, rather than relying solely on static test sets.
4. Safety and Controllability Assurance: Design explainable and auditable agent mechanisms that enable operators to understand and intervene in the agent's decision-making process.
5. Ethical Governance and User Trust: Establish clear ethical guidelines, transparency, and accountability frameworks to strengthen trust among patients and healthcare professionals.
6. Guidance on the Evolving Role of Healthcare Professionals: Clarify the division of labor between AI and humans, train healthcare professionals to collaborate effectively with agents, and adjust medical education content.
7. Interdisciplinary Collaboration: Promote cooperation across fields such as computer science, clinical medicine, ethics, and law to drive responsible innovation.
These directions will provide theoretical support and practical guidance for the sustainable development and implementation of AI agents in healthcare.
Conclusion
AI agents are profoundly transforming the landscape of healthcare, from assisted diagnosis to hospital operations management, demonstrating broad application prospects. However, their safety evaluation, ethical governance, and deep integration with medical workflows still face challenges. This review, by systematically examining historical evolution, application scenarios, evaluation frameworks, and future directions, provides a systematic reference for researchers, clinicians, and policymakers. We call for attention to safety and responsibility alongside technological innovation, to promote AI agents to serve human health in a trustworthy manner.