AI in Healthcare Series: From Decision Support to Drug Prescriptions, Dr. Graham Walker, Kaiser

Key Concepts

• GPT-4.5: A newer iteration of the GPT-4 language model with incremental improvements.
• Hallucination Rate: The tendency of AI models to generate incorrect or nonsensical information.
• GPT Slop: Poorly edited or unrefined content generated by AI models.
• Model Proliferation: The increasing number and variety of AI models available.
• Augmented Intelligence: The use of AI to enhance human capabilities, rather than replace them.
• AI Prescribing: The concept of allowing AI models to legally prescribe medications.

GPT-4.5 and Model Improvements

• Incremental Improvements: GPT-4.5 shows some improvements over GPT-4, but not as dramatic as the leap from GPT-3 to GPT-4. This is attributed to the diminishing returns of pre-training with scaling laws.
• Hallucination Rate Reduction: The hallucination rate is improving, making the models more reliable, but still remains at a surprising 37% in some benchmarks.
• Trust but Verify: Users are becoming more comfortable with AI responses, potentially leading to less verification.
• GPT Slop: The issue of poorly edited AI-generated content remains a concern.
• Example: Sebastian Buck's updated "Tixie Unicorn" image visually represents the model's progress.

Model Proliferation and Choice Paralysis

• Many Models Available: There's a growing number of AI models, creating a "paralysis by choice" for users.
• Competition: Companies are embedding AI capabilities into their platforms to compete and retain users.
• Model Selection: Choosing the right model for a specific task is challenging due to the rapid pace of development.
• Multi-Model Approach: The future likely involves using multiple models for different purposes.
• Cost: Subscriptions to various AI tools can amount to around $300 per month.
• Examples: Chat-GPT is used for unique answers, Claude for editing, and DeepSeek for more "devious" behavior.

Physician Use of AI

• Increased Adoption: A recent survey indicates that two-thirds of clinicians are now using AI in their practice, a significant jump from previous estimates.
• Unauthorized Use: Much of this use is happening outside of officially sanctioned tools, with clinicians using personal devices and public models.
• Motivations: The adoption is driven by the need to keep up with clinical, administrative, and prior authorization tasks.
• Benefits: Clinicians are reporting time savings and improved efficiency.
• Scribing: AI-powered scribing is gaining momentum, with mixed opinions on its helpfulness.
• Definite Advantage Group: A subset of clinicians who are seeing clear advantages from using AI, indicating a potential future trend.
• Examples: Clinicians are using AI for discharge summaries, planning tough conversations, and looking up current evidence.

AI in Medical Education

• Education Potential: AI can be used for empathetic coaching, simulation, and teaching.
• Concept Explanation: AI can explain complex concepts in different ways, tailored to the learner's level of understanding.
• Personalized Learning: AI can act as a personalized tutor, leading to significant learning gains.
• Curriculum Reform: AI may force a re-evaluation of the medical education curriculum, focusing on relevant knowledge and skills.
• Examples: A school in Nigeria used GPT forges as a tutor, resulting in two grade levels of advancement in six weeks. Students are using AI to practice Japanese conversation.

Risk and Comparison

• AI Misses: AI will inevitably make mistakes, but it's important to compare this risk to the risks of current practices.
• Comparison Standard: AI tools should be compared to the risks of human error, delays in care, and other existing problems in healthcare.
• Total Risk: It's crucial to consider the total risk to patients, including the risks of not using AI.
• Ethical Considerations: The loss of control associated with AI tools requires a higher standard of performance.
• Example: The analogy of Tesla's full driving capability, where the system needs to be significantly better than the average human driver due to the loss of control.

AI Prescribing

• Proposed Bill: A bill has been proposed that would allow AI models to legally prescribe medications.
• Concerns: There are concerns about the safety and regulation of AI prescribing.
• Telehealth Comparison: The current practice of telehealth providers quickly prescribing medications raises questions about the level of human oversight.
• Hospital Ownership: One suggestion is to allow hospitals to own the risk and use AI prescribing under a hospital license.
• Likelihood of Passage: The bill is unlikely to pass in its current form.

Notable Quotes

• "This is the worst it'll ever be" - referring to the current state of AI technology, implying continuous improvement.
• "People are going around and using phones using other things" - highlighting the unauthorized use of AI tools by clinicians.

Technical Terms

• Pre-training: The initial training of a language model on a large dataset.
• Scaling Laws: The relationship between the size of a model, the amount of training data, and its performance.
• Scaffolding: The additional software and processes built around AI models to constrain their behavior and improve their reliability.
• Phi: Protected Health Information, patient data that must be kept confidential.

Synthesis/Conclusion

The discussion highlights the rapid advancements and increasing adoption of AI in healthcare. While GPT-4.5 shows incremental improvements, the proliferation of models presents challenges in choosing the right tool. Clinicians are increasingly using AI, often without official authorization, driven by the need to improve efficiency and manage workload. AI holds significant potential for medical education, personalized learning, and curriculum reform. However, it's crucial to carefully consider the risks and compare them to existing practices. The idea of AI prescribing raises concerns and is unlikely to be implemented soon. Overall, the panelists express optimism about the future of AI in healthcare, emphasizing the need for safe and responsible adoption.