In brief
John Halamka, president of Mayo Clinic Platform, argues that AI in medicine should be built like good engineering: modular, validated and trusted before it is used on patients. Predictive models in imaging and cardiology are already catching disease earlier than humans. He holds that generative AI is not yet safe for diagnosis or treatment, and that AI should augment clinicians rather than replace them.
The central tension is speed versus safety. Mayo has built a privacy-protected data platform that has produced 180 models, yet generative AI has no agreed way to certify consistent quality. Halamka's answer is curated data, guardrails, risk-based use cases and emerging regulation. Meanwhile, shrinking doctor supply and global access gaps create pressure to deploy AI widely and soon.
Questions
What is the Mayo Clinic Platform and what does it do?
It is a Mayo Clinic initiative, started in 2019 under CEO Gianrico Ferruccia and led by Halamka from 2020, that lets companies and researchers build AI models on de-identified Mayo data. The data stays in a controlled cloud container and is never sold or taken out.
Can AI really detect diseases before doctors can?
Halamka says Mayo has an algorithm that can spot pancreatic cancer on a CT scan about two years before a human can see it, often at stage zero instead of stage four. Mayo also built a model that reads a single-lead Apple Watch ECG to estimate how well the heart is pumping.
Is it safe to use ChatGPT for medical questions?
Halamka says there are no objective quality measures yet, so risk matters. Asking a chatbot to summarize what trusted sources like Mayo, Stanford or PubMed say is reasonable, but relying on its general internet training is worse, since that data includes claims like ivermectin curing COVID.
Will AI replace doctors?
Halamka says no for now: he expects AI to be 'augmented intelligence' that supports clinicians. Patients also want respect, listening and help with care planning from a human, and he expects AI to ease burnout, for example by cutting a radiation-planning task from 16 hours to 45 minutes.
How does Mayo protect patient privacy when training AI?
It uses what Halamka calls 'data behind glass': the data is fully de-identified, kept in an audited cloud container, limited in what can go in and out, and blocked from being linked with outside datasets. He admits that with enough data and computing power re-identification could become possible, so policy controls are needed too.
Facts
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Halamka says a 1905 self-winding clock built by an engineer in Armonk, New York used a Singer sewing machine motor and mercury switches to rewind its weights.
“he took a Singer sewing machine motor, a simple standards-based modular component at the time, and some mercury switches.”
0:01:00 · John Halamka · Self-winding clock · 1905
Halamka says the company behind the self-winding clock was called International Time Recording and was later renamed International Business Machines.
“And he called the company International Time Recording and decided that it was such a successful idea that he would expand into other devices. And he renamed the company International Business Machines.”
0:01:39 · John Halamka · IBM origins
Halamka says Stanford held what may have been its first entrepreneurial conference in 1983, at which he and Michael Dell were speakers.
“So in 1983, Stanford ran what I think may have been their first entrepreneurial conference where it was, you know, there are students that start companies.”
0:07:46 · John Halamka · Stanford entrepreneurship · 1983
Frederick Terman was Stanford's provost and, per Halamka, had Hewlett and Packard as students in the 1930s.
“So Terman was the provost of Stanford University, and he had two students in the 1930s, one named Hewlett, the other named Packard.”
0:06:08 · John Halamka · Frederick Terman · 1930s
Halamka says Frederick Terman founded the Stanford Research Park and is widely seen as a seminal figure in the creation of Silicon Valley.
“he founded the Stanford Research Park and is widely acknowledged as being one of those seminal figures that brought Silicon Valley to be.”
0:06:50 · John Halamka · Stanford Research Park
Halamka describes MYCIN as the first AI system running on Stanford's medical computing exchange to give doctors rules-based expert advice on antibiotic prescribing.
“Mycine was the first AI system running in the Stanford University Medical Exchange or Sumax that would provide doctors with expert antibiotic prescribing advice.”
0:13:18 · John Halamka · MYCIN · 1980s
Halamka says the idea of the Mayo Clinic Platform originated in 2019 with Mayo Clinic's CEO.
“So in 2019, Gianrico Ferruccia, the CEO of Mayo Clinic, had a notion.”
0:18:21 · John Halamka · Mayo Clinic Platform · 2019
John Halamka joined Mayo Clinic in January 2020 as president of Mayo Clinic Platform.
“I joined Mayo Clinic in January of 2020 as president of Mayo Clinic Platform.”
0:19:09 · John Halamka · Halamka career · 2020
Mayo Clinic de-identified 150 years of data on 10 million patients to make it available on its platform.
“So we took the entire Mayo corpus, 150 years of data on 10 million patients, de-identified it, which in itself is tricky.”
0:19:52 · John Halamka · Mayo data platform
Halamka says about 80 percent of healthcare records are free text such as histories, operative notes and diagnostic reports.
“About 80 % of the records, just the nature of the way records are recorded, is text.”
0:20:50 · John Halamka · Health records
Mayo collected about 125,000 voluntary, consented single-lead ECG recordings from Apple Watches to build a heart-function model.
“So we got 125 ,000 Apple watch submitted telemetry lead one ECGs from patients, voluntarily all concepted.”
0:23:57 · John Halamka · Apple Watch ECG
Halamka says a clinical trial showed Mayo's Apple Watch heart-function algorithm had an area under the curve of 0.92.
“And then we were able to, in a clinical trial, show that this algorithm has an area under the curve of 0.92, right?”
0:23:57 · John Halamka · Algorithm accuracy
The FDA has granted software-as-a-medical-device approval to Mayo's algorithm that analyzes single-lead Apple Watch ECG data, per Halamka.
“and FDA has now given it full software as a medical device approval to be used as part of diagnosis for patients.”
0:24:38 · John Halamka · FDA approval
Halamka says Mayo has an algorithm that can detect pancreatic cancer on CT scans about two years before a human can see it.
“So example, Mayo has an algorithm that can diagnose pancreatic cancer two years before a human can see it.”
0:16:14 · John Halamka · Pancreatic cancer
Halamka says radiology is the leading adopter of AI in healthcare partly because different humans reading X-rays are not very consistent.
“radiology is by far the leading adopter of AI. And why? Well, inter-rater reliability of humans looking at x-rays isn't so great.”
0:14:59 · John Halamka · Radiology AI
Mercy Health de-identified 15 million patients and created a cloud container for AI training and validation in partnership with Mayo Clinic.
“They have taken their 15 million patients, de-identified them, created a cloud container for the purpose of AI training and validation.”
0:42:25 · John Halamka · Mercy partnership
Mayo Clinic planned in 2024 to train its own large language model on de-identified Mayo data, hoping curated medical data would reduce hallucination.
“So we will, in 2024, using de-identified Mayo data, train a novel large language model ourselves with the notion that maybe if it's trained on just really highly curated medical data, it will have less hallucination.”
0:31:39 · John Halamka · Mayo LLM · 2024
Biewald says GPT scores only about 25 percent on a truthfulness benchmark built around common misconceptions, which explains the figure Halamka cited.
“there is this data set called like the truthfulness data set where it only gets 25 % accuracy”
0:36:56 · Lukas Biewald · AI truthfulness
Mayo received 355 internal generative AI proposals, narrowed them to 42 and chose eight areas to support with engineering and compute.
“We got 355 applications. Of those, we downsized to 42. Ultimately, we chose eight general areas.”
0:50:49 · John Halamka · Mayo AI projects · 2024
Halamka says Mayo built a radiation-therapy planning model that takes 45 minutes versus about 16 hours for a physicist on a complex head and neck tumor.
“It can take 16 hours for a physicist to develop a model for a complex head and neck tumor. We have developed a model that does it in 45 minutes.”
0:53:52 · John Halamka · Radiation planning
Halamka says medical innovations take about 20 years on average to spread widely across a specialty.
“that in medicine, it takes about 20 years on average for an innovation to disseminate widely across the specialty.”
0:48:00 · John Halamka · Innovation adoption
Under Mayo's 'data behind glass' approach, data is audited, inputs and outputs are restricted, and linkage with external datasets is prohibited.
“you restrict the inputs and the outputs, and you prohibit linkage with other external data sets.”
0:59:00 · John Halamka · Data privacy
Halamka notes that the US health privacy law HIPAA was written in 1996 and does not mention the genome.
“So HIPAA was written in 1996. It doesn't mention the genome.”
1:00:00 · John Halamka · HIPAA · 1996
Halamka says he is the second person whose genome was sequenced in the Personal Genome Project and that his data has been public for about 15 years.
“I'm the second human sequenced in the Personal Genome Project.”
1:01:04 · John Halamka · Personal Genome Project
Halamka says his wife was diagnosed with stage 3A breast cancer in December 2011 and is healthy today.
“My wife was diagnosed with stage 3A breast cancer in December of 2011, and she's totally fine today.”
0:43:34 · John Halamka · Personal story · 2011