AI in Healthcare Series: Navigating Medical Innovation, Dr. Amy Abernathy, Highlander Health

Key Concepts

• LLMs (Large Language Models): AI models capable of understanding and generating human-like text.
• AGI (Artificial General Intelligence): A hypothetical level of AI that can perform any intellectual task that a human being can.
• Implementation Science: The study of methods to promote the adoption and integration of evidence-based practices and research into regular use.
• Evidence-Based Optimization: Fine-tuning and improving processes based on empirical data and analysis.
• Software as a Medical Device (SaMD): Software intended to be used for one or more medical purposes that perform these purposes without being part of a hardware medical device.
• Clinical Decision Support (CDS): Providing clinicians and patients with knowledge and person-specific information, intelligently filtered or presented at appropriate times, to enhance health and health care.
• Overhang of Capability: The situation where the capabilities of AI models far exceed their current practical applications and understanding by users.

Advancements in AI and LLMs

• Rapid Progress: The capabilities of AI models are advancing rapidly, saturating benchmarks and exceeding expectations.
• Underutilization: The potential of these models is not being fully realized in clinical practice. Physicians may be dabbling but not fully leveraging their capabilities.
• Adoption Curve: A significant challenge is integrating AI solutions into daily workflows and communicating their benefits effectively.
• Prompting Skills: Mastering the art of prompting LLMs is crucial for effective use. Personal experience and learning from others (e.g., family members) can be more effective than formal training.
• Legitimate Use Cases: Determining appropriate and ethical applications of LLMs in various professional contexts (e.g., editorial work, patient care) is essential.
• Uncertainty and Liability: Concerns exist regarding the reliability and potential liability associated with using LLMs in high-stakes situations like patient care.
• Consumer Adoption: Widespread consumer adoption of AI tools is occurring, even as healthcare systems grapple with how to integrate them.

The Analogy to the Early Internet

• Echoes of the Past: The current situation with AI adoption in healthcare is reminiscent of the early days of the internet, where patients brought in stacks of information and clinicians struggled to make sense of it.
• Guiding Patients: Clinicians need to guide patients thoughtfully on how to use AI tools, acknowledging their potential while also addressing their limitations.
• Learning from History: Lessons from the rapid adoption of the internet in the late 1990s can inform the current approach to AI integration.

The Evolution of Technology and Cognition

• Sequential Offloading: Technology has progressively offloaded computation (computers), memory (internet), and attention (social media). The current frontier involves offloading cognition to LLMs.
• Threat to Cognition: There's a concern that AI could potentially threaten human cognition.
• Daily Use: The goal is for LLMs to become as commonplace as the internet in daily clinical practice, assisting with tasks like treatment planning and data analysis.
• Patient Expectations: Patients are increasingly asking about the specific AI models used and their capabilities, requiring clinicians to have a deeper understanding of these tools.

AI vs. AI + Physician Performance

• Surprising Results: Studies have shown that AI alone can outperform physicians, even when physicians are using AI tools.
• Google's AMIE System: Research on Google's AI AMIE system demonstrates that clinicians using the tool sometimes underperform compared to the AI alone.
• Best Practices Needed: The lack of established best practices for interacting with AI models may contribute to this underperformance.

The Human Element and Emotional Intelligence

• Offloading Cognition, Not Social Interaction: While AI can offload cognitive tasks, it cannot replace human social interaction, empathy, and leadership.
• Empathy and Emotion: LLMs can help offload some of the demands of emoting, potentially freeing up humans to focus on more distinctly human aspects of care.
• Humanity in the Loop: There will be an increased need for humans to interpret AI outputs and provide personalized care and guidance to patients.

Training and Integrated Thinking

• Internship Training: Medical training, particularly during internship, focuses on developing a "gut instinct" for patient assessment, which involves integrating facts and applying them to individual cases.
• Bias and Instinct: This instinctive plateau is embedded with biases that are not yet built into LLMs.
• Structured vs. Fuzzy Thinking: Clinicians may initially pull LLMs down because their structured thinking clashes with the AI's capabilities. The goal is to integrate fuzzy thinking with AI to enhance decision-making.

Evidence Generation and Implementation

• Divorce Between Technology and Practice: A disconnect exists between the capabilities of AI and its actual implementation in healthcare systems.
• Implementation Science: There is a need for more evidence and research on how to effectively implement AI in clinical settings.
• Evidence-Based Optimization: Implementation should be based on evidence-based optimization, similar to how algorithms are embedded in consumer applications.
• Dynamic Settings: Clinical settings change as AI is implemented, requiring ongoing study of the impact on care delivery and data generation.
• Regulatory Challenges: The regulatory framework for AI in healthcare is still evolving, particularly for LLMs used for clinical decision support.
• Flexible Governance: There is a need for more flexible governance systems, including self-governance from industry and adaptive regulation from the FDA.

Regulatory Considerations

• Clear Regulation: Software as a Medical Device (SaMD) with AI algorithms (e.g., mammogram reading) has a clear regulatory path.
• Less Clear Regulation: LLMs supporting the practice of medicine and clinical decision support have less defined regulations.
• Fixed World Presumption: Current regulations often presume a fixed world with predetermined change control plans, which may not be suitable for rapidly evolving AI technologies.

Conclusion

The integration of AI, particularly LLMs, into healthcare presents both opportunities and challenges. While AI models are rapidly advancing and demonstrating impressive capabilities, their full potential is not yet being realized in clinical practice. Effective implementation requires a focus on implementation science, evidence-based optimization, and flexible governance systems. Furthermore, the human element remains crucial, as AI cannot replace the empathy, social interaction, and leadership skills that clinicians provide. The future of healthcare will likely involve a new balance of human and AI capabilities, with clinicians playing a vital role in interpreting AI outputs and providing personalized care to patients.