JAMA Clinical Reviews How Artificial Intelligence Has Evolved and the Implications for Health Care
22 snips
Jan 16, 2024 Dr. Michael Howell discusses the evolution of AI in healthcare, from symbolic AI to AI 3.0. The limitations of AI 1.0 and the impact of biases in AI 2.0 are explored. Advancements in AI 3.0, such as foundation models and generative AI, are discussed. The importance of clinicians understanding different AI models and gaining first-hand experience is emphasized.
AI Snips
Chapters
Transcript
Episode notes
Three Distinct Epochs Explain AI Differences
- AI evolved in three distinct epochs with different capabilities and risks.
- AI 1.0 (symbolic/probabilistic), AI 2.0 (deep learning for single tasks), and AI 3.0 (foundation/large language models that can perform many tasks and generate content).
Clinical Pathways Are Classic AI 1.0
- AI 1.0 encodes human knowledge as explicit rules used in clinical pathways.
- Howell describes them as a million if-then rules used in EHR flow tools for sepsis, ARDS, and other clinical pathways clinicians write.
Rule Based Systems Break Down In Real World
- AI 1.0 is fundamentally limited and brittle in messy real-world settings.
- Howell contrasts rule-based spam detectors and clinical rules with deep learning, noting they fail empirically and for complex tasks.