309 ‒ AI in medicine: its potential to revolutionize disease prediction, diagnosis, and outcomes

KEYPOINTS

• 01

  AI's near-term promise: lift the bottom half of doctors to the level of the top half.

  Kohane, a physician and computer scientist who chairs biomedical informatics at Harvard Medical School, says merely making the bottom 50% of doctors as good as the top 50% would be transformative—before even counting 'superhuman' capabilities.

• 02

  AI has come in waves.

  Kohane majored in biology, got 'nerd-sniped' by computing in the late 1970s and did a computer science PhD in the early 1980s—AI's second wave of rule-based 'expert systems' written in English-like rules, such as Digital Equipment Corporation's R1, which configured minicomputers. These were brittle and lacked data, so the field stalled.

• 03

  Gaming GPUs and data made deep learning work.

  Video games' need to render high-resolution, 24-bit color at 60 frames per second produced cheap, massively parallel chips that could train multi-layer neural networks. Around 2012, deep networks on GPUs crushed competitors in an annual image-recognition contest—when 'everybody woke up'.

• 04

  Transformers and scale produced large language models.

  Google's 2017 'Attention Is All You Need' paper introduced the transformer, noticed mainly by computer scientists. Kohane wrote a 2018 JAMA editorial on Google's diabetic-retinopathy AI. Researchers saw performance keep scaling with data and model size, and GPT-3.5 (ChatGPT, December 2022) and GPT-4 became the tipping point—like when everyone knew what Google was by 2000.

• 05

  He saw GPT-4 early and was stunned.

  In October 2022, Peter Lee emailed him about GPT-4 before ChatGPT's public release, leading to their book on AI in medicine.

• 06

  Visual specialties will be augmented first, then language-heavy work.

  Radiology, dermatology, pathology and ophthalmology—fields interpreting images—are most easily augmented, followed by work combining language and visual data. Geoffrey Hinton predicted around 2016 that radiologists would soon be replaced; that hasn't happened, and the real future job of doctors in 10 years is an open question.

• 07

  Doctor shortages make AI urgent.

  The Association of American Medical Colleges estimates a shortfall on the order of 50,000 primary care doctors by 2035. Patients are already using GPT-4, Gemini and Claude in unanticipated ways. Kohane imagines low-cost AI-assisted concierge-like care versus the $5,000–$20,000 a year concierge services cost in Boston.

• 08

  Better prediction needs long-horizon data, which is scarce.

  Peter calls 10-year cardiovascular risk 'ridiculous', preferring lifetime or at least 30-year risk models with many more inputs. Kohane notes Framingham and its offspring cohort followed LDL over decades; systems like Kaiser Permanente have 20–25 years of detailed records, enough to approximate such models now—don't let perfect be the enemy of good.

• 09

  Novel signals—retina, gait, voice—could predict disease.

  Peter would take a retinal exam over HbA1c 'all day every day'. AI might analyze eye movements or gait to predict neurodegenerative disease 20 years early, and cameras can detect hypoglycemia risk. Kohane contributed to a 2005 Nature study showing about a 30% drop in gene transcription after age 40, with one 90-year-old who looked young.

• 10

  Patients now have a legal right to their data, enabling AI tools.

  The 21st Century Cures Act requires programmatic patient access; Apple Health connects to about 800 hospitals. With consent, GPT can extract data from years of lab reports—Epic is using GPT-4 with Stanford—replacing clumsy optical character recognition. Still, Kohane pays hundreds of dollars to extract records for rare-disease patients.

• 11

  Why AI suddenly felt different from past hype.

  Kohane notes that in the 1980s nobody worried about AI threats because it simply wasn't powerful enough; the Terminator's Skynet was fiction. With GPT-3.5 and GPT-4, AI became a household noun almost overnight. He sees data as one of three key ingredients—alongside compute and algorithms—and electronic health records now supply medical data at scale, though it's messy and fragmented across systems.

• 12

  AI can settle old clinical debates quickly when data exist.

  Kohane recalls 1990s arguments over how often and at what dose to give certain drugs; with enough real-world data, such questions can be answered rapidly. He also describes an NEJM 'image of the week' of a 72-year-old's back that AI could plausibly interpret—examples of pattern problems that are fundamentally data problems.