By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Peter Attia MD309 ‒ AI in medicine: its potential to revolutionize disease prediction, diagnosis, and outcomes
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
- 0:00 – 8:58
Zak Kohane's path from biology to medical AI and his first look at GPT-4
Peter Attia welcomes Zak Kohane, who describes his unconventional path from biology and computer science through medical school, a PhD during AI's second heyday, and pediatric endocrinology. He recounts getting early access to GPT-4 and being stunned when it worked through a rare endocrine case.
- •Kohane left medical school to do a computer science PhD in the early 1980s, then returned to finish clinical training
- •He founded a biomedical informatics lab and later a department at Harvard
- •Peter Lee gave him early access to GPT-4 before ChatGPT was public
- •GPT-4 correctly walked through the workup of a rare ambiguous genitalia case
- •Line engineers saw scaling coming, while many top computer scientists did not
- 8:58 – 16:55
The Turing test and the moving goalposts of intelligence
Attia asks Kohane to explain the first generation of AI and the Turing test. Kohane describes early perceptrons and how definitions of intelligence keep shifting as machines master chess, faces and writing.
- •Turing proposed that a machine that can converse without being identified as a computer shows intelligent behavior
- •Chess, face recognition and creative writing were each dismissed as 'not really intelligence' once computers did them
- •Language models predict the next word using a model of about a trillion parameters
- •Kohane no longer finds the Turing test debate interesting and links this to Kahneman's system one and system two
- 16:55 – 24:56
Second-wave AI: rule-based expert systems and their limits
Kohane explains the 1970s–80s rule-based expert systems such as MYCIN and R1, which encoded human knowledge as rules. They were labor-intensive, brittle and hit a complexity barrier, though a few narrow tools like ECG reading survived.
- •Rules were written in English-like form from textbooks and expert introspection
- •MYCIN was an antibiotic advisor that needed constant manual updating
- •R1 configured computer orders for Digital Equipment Corporation
- •Beyond a few thousand rules, interactions became unpredictable (the complexity barrier)
- •ECG interpretation and serum protein electrophoresis programs were deployed in narrow areas
- 24:56 – 33:26
Three ingredients of the third wave: data, deep networks and GPUs
Kohane lists the three revolutions behind modern AI: large online datasets, multilayer neural networks, and GPU hardware originally built for video games. He dates the breakthrough to the 2012 image recognition contest.
- •ImageNet, PubMed and electronic health records supplied large annotated data
- •Deep multilayer neural networks overcame the perceptron's limitations
- •GPUs built for video games perform the parallel linear algebra neural networks need
- •Nvidia became the most successful ecosystem for deep learning
- •Around 2012 deep networks on GPUs clearly beat rivals at image recognition
- 33:26 – 42:58
Retinopathy recognition and the transformer architecture
Kohane describes the 2018 Google retinopathy paper, where doctors only labeled images and judged performance, as a wake-up call. He then explains the 2017 transformer paper and how it led to large language models.
- •Transfer learning from ImageNet improved retina recognition
- •Doctors' role in the retinopathy study was limited to labeling and judging
- •Word embeddings captured meaning from co-occurring words
- •The transformer added word position and scaled to roughly a trillion variables in GPT-4
- •Kohane says the 2012–2018 deep learning wave was already a new heyday before transformers
- 42:58 – 58:18
AI risks and regulation debates
Attia asks about the growing regulation debate. Kohane covers the doomer view, OpenAI's calls for regulation and the possible motive of regulatory lock-in, and the nearer dangers of human misuse, military autonomy, job displacement and bias.
- •Some scientists now fear existential threats from AI, while nobody openly argues for racing ahead heedlessly
- •Sam Altman called for regulation; Kohane offers sincere concern or regulatory lock-in as possible reasons
- •Kohane sees human misuse such as misinformation and autonomous weapons as the clear and present danger
- •AGI and self-awareness are loosely defined; Hofstadter's 'I Am a Strange Loop' is cited
- •Gain-of-function research is an example of potentially severe misuse
- •Worries also include bias, hallucinations, white-collar job displacement and the quality of medical advice
- 58:18 – 1:06:09
AI in radiology and image-based specialties
Attia moves to medicine, asking about radiology. Kohane says AI matches many experts in visual specialties but isn't replacing doctors, and notes that transformers add clinical context and that the real problem is a shortage of clinicians.
- •Pathology, radiology and dermatology programs perform as well as many experts
- •Overnight offshore reads could be replaced by AI
- •Multimodal transformers can combine images with patient history
- •Geoffrey Hinton's prediction that radiologists would be unnecessary was wrong
- •Primary care and some pediatric subspecialties are badly under-filled
- 1:06:09 – 1:12:04
Patients and doctors using chatbots: diagnoses, HIPAA and the Google reflex
Kohane notes that patients are already using GPT-4, citing a tethered cord syndrome diagnosis. He discusses doctors' lack of a 'Google reflex', HIPAA concerns with ChatGPT, and the use of AI for administrative tasks.
- •A mother got a correct diagnosis from GPT-4 after multiple doctors and imaging studies failed
- •Kohane leads the coordinating center of the Undiagnosed Network
- •Most clinicians don't instinctively look up or query tools for diagnosis
- •Pasting patient notes into ChatGPT is a HIPAA violation, unlike HIPAA-covered cloud versions
- •Doctors mostly use it for administrative tasks like prior authorization letters
- 1:12:04 – 1:21:11
Augmenting clinicians and filling the primary care gap
Attia and Kohane discuss 3D imaging such as echocardiograms and a tiered view of which physicians are easiest to augment. Kohane argues nurse practitioners and physician assistants with AI could fill the primary care shortage without any new technological breakthroughs.
- •EchoCLIP trained on a million echocardiograms and reports
- •Visual-data specialties are most easily augmented, followed by those combining language and images
- •AAMC projects a shortfall of about fifty thousand primary care doctors by 2035
- •Current models are good enough; the barriers are evaluation, workflow, payment and reimbursement
- •Raising the bottom half of doctors to the level of the top half would be transformative
- 1:21:11 – 1:32:10
Predicting disease early: Alzheimer's, long-horizon risk and data sources
Attia asks whether AI could predict neurodegeneration decades ahead. Kohane says early detection is straightforward given the right data and discusses retina-based prediction, long-term risk models, electronic health record limitations, and subgroups hidden in existing data such as in autism.
- •Retinal images already reveal hypertension, age, sex and longevity estimates
- •Gait, speech, eye movements and skin could predict Alzheimer's
- •Attia wants lifetime or thirty-year risk models rather than ten-year ones
- •Israeli HMOs like Clalit and Kaiser have long continuous records; proxies can substitute for some measures
- •Kohane's autism EHR study found immunological subgroups that were knowable but not known
- 1:32:10 – 1:41:54
Patient data access, Apple Health and using GPT-4 on records
Attia raises how to get patient data to train models and how unfriendly health data is. Kohane points to the 21st Century Cures Act, Apple Health, and HIPAA-compliant GPT-4 on Azure, and illustrates GPT-4's abilities with image and data analysis examples.
- •Patients have a legal right to programmatic access to their data
- •Apple Health pulls records from hundreds of hospitals onto iPhones
- •GPT-4 can convert lab PDFs into spreadsheets, but HIPAA-compliant versions with patient consent are needed
- •GPT-4 correctly guessed shiitake mushroom toxicity from a New England Journal image
- •Dean George Daley wants AI infused throughout Harvard's medical education
- 1:41:54 – 1:49:45
Autonomous robotic surgery
Attia asks whether robots like the da Vinci could one day operate on their own. Kohane predicts robot-controlled routine prostatectomy within ten years, comparing it to autonomous driving, while warning about keeping doctors alert when AI does the work.
- •Generative models trained on surgical videos can already suggest corrections
- •Kohane calls a robot-controlled da Vinci in ten years a safe bet for routine prostatectomy
- •Surgery is framed as a pixel and movement problem with many degrees of freedom
- •His Tesla anecdote shows the need to keep humans attentive
- •Attia doubts complex operations like the Whipple procedure
- 1:49:45 – 1:54:44
AI and mental health care
Attia asks whether psychiatry will be the most protected medical field. Kohane cites the ELIZA program and says many patients may prefer a non-judgmental, empathetic non-human, such as for cognitive behavioral therapy.
- •Mental health resources are insufficient, even at Harvard
- •Rapport matters as much as knowledge in psychotherapy
- •ELIZA, a 1960s–70s pattern matcher, drew users who valued its non-judgmental nature
- •Kohane would not want it for himself but sees acceptance among a subset of patients
- 1:54:44 – 1:58:52
Business models and resistance from medical institutions
Asked what he most fears and hopes for in medicine, Kohane worries the establishment will use AI to cement current practices and block information. He is excited about new business models from young doctors and gives better than even odds one will emerge within ten years.
- •Hospitals have huge revenue but 1–2% margins, making them risk averse
- •Fear: information blocking and regulation that hinders new entrants
- •Excitement: new data-driven care businesses outside hospitals
- •Patient data access rights plus AI could allow companies to work around hospitals
- •Some companies may accept payment tied to performance
- 1:58:52 – 2:08:02
AI beyond medicine: social media, creativity and human greatness
Kohane fears AI will magnify social media chaos and bots, but is hopeful that it will widen creative expression, comparing it to the printing press. Attia asks what it means for human greatness if AI can write like a master, and Kohane imagines future AI celebrities.
- •Biggest fear is AI worsening social media vitriol and eroding trust
- •AI could let more people express themselves musically, visually and in film
- •The printing press analogy covers both benefits and wars
- •Attia asks whether an AI writing like Siddhartha Mukherjee would diminish admiration for greatness
- •Kohane envisions AI authors with brands and followings in fifty years
- 2:08:02 – 2:16:15
Digital immortality, Rewind AI and closing thoughts
Attia asks about training an AI on oneself to leave a legacy for his children. Kohane discusses Rewind.ai, which records his screen and audio, finds the idea of talking to the dead creepy but sees value in keeping a personal record. The conversation closes with Attia's surprise at the timeline and Kohane's suggestion for his clinic.
- •Kohane notes Mark Cuban sought trademarks and copyrights on his voice and likeness
- •Rewind.ai records screen and sound compactly and makes it searchable
- •Kohane thinks a good approximation of public Zak is technically feasible
- •He finds returning from the dead to advise children creepy but values a personal record against being taken out of context
- •He suggests Attia use video, sound and gait to follow patients, with consent
