Agentic AI Explained: Peter Lee on the Next Revolution in Healthcare. Part 3

Summary of YouTube Video: Agentic AI with Dr. Peter Lee

Key Concepts:

• Agentic AI: AI systems capable of memory, entitlements, action, and reasoning.
• Memory (Episodic): AI's ability to remember past interactions and context.
• Entitlements: Permissions granted to an AI to access tools and data.
• Action: AI's ability to change the state of the world (e.g., moving a mouse, sending emails).
• Reasoning: AI's ability to think and make decisions.
• Meeting Facilitator Agent: An AI agent in Microsoft Teams that helps manage meetings.
• Healthcare Orchestrator Agent: An AI agent that assists in healthcare settings, such as tumor board meetings.
• MAI DXO: A multi-agent AI system for medical diagnosis.
• Chain of Debate: A process where multiple AI agents with different roles debate to arrive at a conclusion.
• Differential Diagnosis: The process of differentiating between diseases with similar symptoms.

1. Introduction and Framing of Agentic AI

• Don introduces Dr. Peter Lee to discuss Agentic AI, framing it as the next step beyond GenAI's writing capabilities.
• Agentic AI involves AI systems that can perform actions, make decisions, and utilize tools.
• Dr. Lee agrees that Agentic AI is a significant development, comparing the current state to the "middle innings" of a tech revolution.

2. Components of Agentic AI

• Dr. Lee outlines four key components of Agentic AI:
  • Memory: AI systems are evolving to have episodic memory, remembering user context and past interactions.
  • Entitlements: Defining the permissions and access rights of the AI, such as accessing tools, sending emails, or querying databases.
  • Action: The ability of the AI to change the state of the world, including physical actions (robots) and digital actions (moving a mouse, writing files).
  • Reasoning: The overarching ability of the AI to think, analyze, and make decisions.
• These components are being integrated by researchers and industry to develop agent-based systems.

3. Meeting Facilitator Agent Example

• Dr. Lee mentions the meeting facilitator agent in Microsoft Teams as an example of Agentic AI in use.
• This agent lives in the meeting chat, knows the agenda, keeps the meeting on schedule, and can prompt participants.

4. Healthcare Orchestrator Agent Case Study (Stanford Medicine)

• In collaboration with Stanford Medicine, a healthcare orchestrator agent was developed for tumor board meetings.
• The agent participates in the meeting, retrieves patient records, invokes AI models for analysis (radiology, pathology), and accesses relevant research literature.
• This agent exhibits memory (context of the meeting), entitlements (access to tools), action (recording notes, creating presentations), and reasoning (contributing to treatment decisions).
• The tumor board example highlights how AI can act like a person, participating in problem-solving.

5. MAI DXO: Multi-Agent System for Medical Diagnosis

• Microsoft AI division developed MAI DXO, a multi-agent system for medical diagnosis.
• The system includes multiple AI agents with different roles:
  • One agent interacts with the patient to arrive at a diagnosis.
  • Another agent acts as a contrarian, challenging the primary agent's thought process.
  • A third agent focuses on cost considerations, questioning the need for lab tests.
• These agents engage in a "chain of debate" to arrive at a correct diagnosis.
• Testing with the New England Journal of Medicine on rare diagnostic cases showed that MAI DXO achieved over 80% accuracy, significantly better than seasoned human clinicians.
• This demonstrates how agentic computing can reinforce AI intelligence and reduce hallucination rates.

6. AI as a Second Set of Eyes

• Dr. Lee suggests that doctors and nurses can use AI as a "second set of eyes" to review their work.
• Presenting a differential diagnosis to an AI can help identify overlooked factors or potential errors.
• This use case is valuable even if the AI is slightly wrong, as it can expand the decision-maker's perspective.

7. Conclusion

• Dr. Lee emphasizes the potential of Agentic AI to facilitate human processes, reinforce AI intelligence, and reduce errors.
• The examples of the meeting facilitator agent, healthcare orchestrator agent, and MAI DXO demonstrate the diverse applications of Agentic AI.
• The key takeaway is that Agentic AI is a promising area of development with the potential to significantly impact various industries.