#309 ‒ AI in medicine: its potential to revolutionize disease prediction, diagnosis, and outcomes, causes for concern in medicine and beyond, and more | Isaac Kohane, M.D., Ph.D.
Isaac "Zak" Kohane, physician-scientist and chair of Biomedical Informatics at Harvard Medical School, discusses AI's evolution and its transformative impact on medicine. He shares insights from early GPT-4 interactions, highlighting AI's role in image-based specialties, early disease diagnosis, and future possibilities like autonomous surgery, while also addressing ethical and regulatory concerns.
Deep Dive Analysis
15 Topic Outline
Zak Kohane's Unconventional Journey and Early GPT-4 Experience
Evolution of AI: Turing Test and First Generation
Second Generation AI: Rule-Based Systems and Their Limitations
Three Revolutions Driving Modern AI: Data, Neural Networks, and GPUs
AI Breakthroughs in Image-Based Medical Specialties
Third Generation AI: Transformer Architecture and Large Language Models
Concerns and Debates Surrounding AI Regulation
AI's Role in Augmenting Clinicians and Addressing Physician Shortages
AI's Potential for Early Disease Diagnosis and Prediction
Challenges of Data Accessibility and Integration in Healthcare
The Future of Autonomous Robotic Surgery
AI's Impact on Mental Health Care
AI's Transformation of Medical Industry Business Models
Broader Societal Impacts of AI: Creativity vs. Misinformation
Digital Immortality and Legacy: Emulating Personality with AI
8 Key Concepts
Turing Test
A test proposed by Alan Turing to determine if a computational entity can maintain a conversation without revealing it's a computer, leading others to mistake it for a human. The goalposts for what constitutes 'intelligent behavior' in this test have continuously moved as AI capabilities advance.
Rule-based Systems (Second Gen AI)
AI systems from the 1970s programmed using human-like rules (e.g., 'If X, then Y with probability Z'). These systems were labor-intensive to maintain, prone to unexpected interactions between rules, and brittle when encountering information outside their narrow focus.
Deep Neural Networks
Multi-layered artificial neural networks where the output of one layer propagates to the next. This architecture, combined with large datasets and powerful processing units, overcame limitations of earlier perceptrons and enabled significant breakthroughs in pattern recognition.
Graphical Processing Units (GPUs)
Parallel processors originally developed for rendering high-resolution video game graphics. Their ability to perform linear algebra efficiently made them critical for training deep neural networks, accelerating AI development significantly.
Transfer Learning
A technique where a neural network is first trained on a broad, general dataset (e.g., recognizing cats and dogs) to learn fundamental patterns, and then fine-tuned on a more specific dataset (e.g., medical images). This approach significantly improves performance compared to training only on the specific dataset.
Transformer Architecture
A specific type of neural network architecture (introduced in 2017) that improved natural language processing by considering the exact position and ordering of words in a sentence, not just their co-occurrence. This breakthrough enabled the development of large language models like GPT-4.
Regulatory Lock-in
A potential consequence of AI regulation where well-funded companies influence the creation of complex compliance requirements. This can inadvertently create barriers for smaller, less-resourced companies, limiting competition and innovation.
Multi-modality AI
AI systems capable of integrating and interpreting data from multiple sources or types, such as combining image data (e.g., X-rays) with textual clinical history. This allows for a more comprehensive understanding and improved diagnostic accuracy compared to single-modality systems.
12 Questions Answered
The initial concept, embodied by the Turing Test, proposed that if a machine could converse indistinguishably from a human, it was intelligent. However, as AI capabilities advanced (e.g., beating chess masters, recognizing faces), the definition of 'intelligence' has continuously shifted, making the Turing Test's goalposts a moving target.
Early AI, including rule-based expert systems, relied on manually programmed rules derived from human experts. These systems were labor-intensive to build and maintain, struggled with the complexity of interacting rules, and were brittle when faced with situations outside their narrow, predefined scope, lacking common sense.
The three revolutions were: 1) the availability of vast online datasets (e.g., ImageNet, PubMed, electronic health records), 2) the development of multi-level deep neural networks that could learn complex patterns, and 3) the advent of powerful Graphical Processing Units (GPUs) that enabled the parallel computation needed to train these large networks.
AI can now perform as well as many human experts in interpreting visual medical data (e.g., X-rays, MRIs, pathology slides, dermatology images). This is due to large annotated image datasets and deep neural networks, though it primarily augments rather than replaces doctors, as image recognition is only one part of their job.
The transformer architecture, introduced in 2017, is a type of neural network that significantly advanced natural language processing. It improved AI's ability to understand text by recognizing the importance of word order and position within a sentence, leading to the development of highly capable large language models like GPT-4.
Concerns range from the immediate dangers of individuals using AI for harmful purposes (e.g., spreading misinformation, creating toxins) and military applications, to potential job displacement in white-collar industries, and even existential threats posed by highly advanced AI. There's also debate about 'regulatory lock-in,' where regulations might inadvertently favor large, well-funded companies.
AI can augment nurse practitioners and physician assistants by assisting with tasks like ordering appropriate studies, suggesting medications, assessing chronic conditions, and handling administrative burdens (e.g., prior authorizations). This could enable paraprofessionals to operate at the level of entry-level doctors, filling critical gaps in care.
Yes, it is considered realistic. AI can already detect subtle nuances in data like retinal images to predict conditions like hypertension, age, and longevity. With access to comprehensive, long-term patient data (e.g., from electronic health records, gait analysis, speech patterns), AI could accurately predict neurodegenerative diseases far in advance, enabling earlier intervention.
The biggest barrier is the medical establishment's resistance to change and potential 'information blocking.' Hospitals, despite large revenues, operate on thin margins and are risk-averse, making them hesitant to adopt disruptive technologies. Data accessibility, despite legal frameworks like the 21st Century Cures Act, remains a practical challenge due to user-unfriendly data formats and captive systems.
Autonomous robotic surgery, particularly for routine procedures like prostate removal, is considered a very safe bet within the next 10 years. AI is already being trained on surgical videos to identify critical steps and potential errors, leveraging advancements in robotics and autonomy seen in other fields like self-driving cars.
AI could provide non-judgmental, empathetic support for a large group of patients, particularly for cognitive behavioral therapy. Early AI programs like ELIZA demonstrated that some individuals prefer non-human interaction for therapy. AI could offer accessible mental health support, remembering past conversations and showing verbal empathy, potentially filling gaps in underserved areas.
Digital immortality refers to the ability to create an AI trained on an individual's vast personal data (conversations, writings, screen activity) to emulate their personality and responses after their death. Technologies like 'rewind.ai' are already recording screen activity and audio, suggesting that creating a 'public approximation' of an individual's persona might be technically feasible in the coming decades, though ethical concerns remain.
12 Actionable Insights
1. Leverage HIPAA-Compliant AI for Admin
Physicians should use HIPAA-compliant AI platforms (e.g., Microsoft’s Azure Cloud GPT) to automate administrative tasks like writing prior authorization letters. This saves significant time and improves efficiency in healthcare administration.
2. Use AI for Diagnostic Support
Clinicians should adopt a ‘GPT-4 reflex’ to input patient histories and reports into HIPAA-compliant AI for diagnostic assistance in complex or undiagnosed cases. This can help identify missed diagnoses and augment physician acumen.
3. Patients: Access Your Health Data
Patients should proactively access their health data (e.g., via Apple Health or patient portals) and, with consent, use HIPAA-compliant AI to analyze it. This empowers individuals to gain personal health insights and potentially aid in diagnosis.
4. Augment Mid-Level Providers with AI
Healthcare systems should augment nurse practitioners and physician assistants with AI tools to perform at the level of entry-level doctors, especially in primary care and underserved specialties. This strategy can help address critical shortages in medical expertise and access.
5. Apply AI to Medical Imaging
Utilize AI, particularly multi-modal transformer architectures, to interpret complex visual medical data (e.g., CT scans, MRIs, echocardiograms) combined with clinical text. This provides expert-level analysis, improving diagnostic accuracy and efficiency in visual-based specialties like radiology.
6. Predict Disease Early with AI
Invest in and use AI models that combine diverse data (retinal images, gait, speech patterns, eye movements) for early prediction of diseases like Alzheimer’s, potentially decades in advance. This enables interventions at stages where reversal or significant halting of progression might be possible.
7. Build New AI-Driven Health Businesses
Entrepreneurs should create alternative healthcare businesses by leveraging patients’ legal right to access their data, using AI to clean and analyze it. This can disrupt traditional hospital models and offer new, data-driven insights and care delivery methods.
8. Stay Vigilant with AI Assistance
When AI augments medical decision-making or procedures, doctors must actively maintain vigilance and avoid over-relying on AI suggestions. Continuous critical assessment is essential to prevent errors and maintain clinical judgment.
9. Expand Creativity with AI Tools
Individuals should experiment with AI tools for creative expression in areas like music generation, picture generation, and filmmaking, regardless of prior skill level. This can significantly expand personal creativity and human expression.
10. Focus on AI’s Practical Output
When evaluating AI, prioritize its actual output and capabilities rather than engaging in philosophical debates about its ‘intelligence’ or ‘self-awareness.’ This pragmatic approach helps in assessing its utility and impact.
11. Beware AI’s Social Harms
Recognize that AI can magnify social network problems, such as spreading misinformation and creating convincing bots, making it harder to distinguish reality from AI-generated content. Exercise critical thinking and verify information from online sources.
12. Archive Personal Data with AI
Consider using personal AI tools (e.g., Rewind.ai) that record screen activity and audio to create a searchable, compressed archive of one’s digital and verbal interactions. This can serve as a personal memory aid and provide context for past statements.
6 Key Quotes
It's a noble profession, but it's not a science. It's an art. It's not a science. And I thought I was going into science.
Zak Kohane
The goalposts around the Turing test keep getting moved. So I just have to say that I no longer find that an interesting topic because it's what it's actually doing.
Zak Kohane
Perfect is the enemy of good. Waiting 30 years to have the perfect data set is not the right answer to help patients now.
Zak Kohane
What do you call the medical student who graduates at the bottom of his class? Doctor.
Zak Kohane
I would take a retinal exam over a hemoglobin A1C all day, every day. I'd never look at another A1C again if I could see the retina of every one of my patients.
Peter Attia
I think it's actually pretty creepy to come back from the dead, to talk to your children.
Zak Kohane