What Does Modern Healthcare Look Like In The Age Of AI?

Dan Riskin, MD, MBA, FACS, is Clinical Professor of Surgery at Stanford University and CEO and Founder of Verantos.

gettyI met her in the intensive care unit, after the damage was done. Years earlier, a heart murmur had been noted at a routine visit. It went into the doctor’s note and nowhere else. She changed care providers, the note didn’t follow, and the murmur was forgotten until she arrived in the emergency department unable to breathe. She needed emergency heart surgery for a problem that could have been safely fixed years earlier.

That isn’t uncommon in healthcare today. A patient can receive a gene therapy that was science fiction a decade ago while their referral to a specialist travels by fax. The treatments are extraordinary. The system around them has not kept up. So, it is fair to ask: In the age of AI, what should modern healthcare look like?

I received my first AI grant from the National Institutes of Health in 2010 and filed my first AI patent applications in 2011, back when only the most technically oriented doctors had heard of AI. In 2012, I wrote in Forbes that healthcare has a revolution roughly every 50 years, including public health in the 1870s, penicillin and effective medications in the 1920s, and evidence-based medicine in the 1970s. I wrote that the next revolution would occur in the 2020s, built on newly available electronic data, rapidly improving AI technology and increasingly available compute power. Those enabling technologies have since matured.

In some clinics today, using ambient AI scribes, a doctor walks in already briefed on the chart, spends the visit listening rather than typing and sends the patient home with a plan in plain language. Had the woman in the ICU been seen this way, her murmur likely would have been known and considered by every doctor she saw before it became an emergency. But that is the smaller prize: a better conversation about care we already know how to give. The bigger prize is rewriting the standard of care itself, by letting AI learn from what happens to millions of patients every day: which treatment works best for whom.

Why rewrite the standard of care? Because the evidence behind the current approach has limits. A clinical trial can show that a treatment is safe and works in general. It cannot say which of more than 200 approved options for high blood pressure or more than 150 for diabetes is best for a specific patient. So, doctors reach for what seemed to work on the last patient.

Google Maps learns from every drive and Spotify from every song; medicine rarely learns from routine experience. After surgery, the woman in the ICU had a dozen reasonable options for her weakened heart, and the honest answer to “which is best for her” was that no one knew. Yet millions of patients before her had been treated with those same options. Had healthcare learned from these interactions, I believe the answer for her would have been known, her likelihood of recovery better and her own care a lesson for the next patient. That is how medicine moves from guessing to knowing.

Here is the hard part, and where data and AI become inseparable. AI does not know medicine; it learns from the information it sees and makes decisions based on the records it is shown. In routine care, the clinical record was written to treat the patient and bill the visit, not to answer a research question. I shared this in testimony before Congress in 2014, using this same patient as an example, highlighting records not moving, facts trapped in narrative text and critical information stored in disparate and unavailable systems.

Since then, the size of those gaps has been measured multiple times. A systematic review of 67 studies covering 41 conditions found that the coded fields of the electronic health record identified, on average, only 62% of the patients who actually had a condition. In the TRUST study, funded by the FDA to help inform regulatory guidance, I analyzed records from more than 50 hospitals and more than 100,000 patients. The type of data most real-world analyses rely on was accurate about 60% of the time, was less than half complete and only one fact in nine could be traced to the original clinical document. These findings are further supported by research published in the Journal of Public Health and the Journal of Clinical Epidemiology. Learning from data like that can make a harmful treatment look helpful, or the reverse.

This is not a cleanup problem. Turning routine care into data an AI can safely learn from requires discipline. This means drawing the facts out of the notes, linking across every place a patient was treated and every type of data, inferring what no single record states outright, tracing each fact to its clinical source of truth and repairing it where it is wrong, measuring the result at every step and running the study with the discipline of a trial.​ In the TRUST study, this approach raised data accuracy above 90%, completeness above 95% and traceability to nearly 80%.​​

Done well, learning from routine care transforms healthcare. Done poorly, it is dangerous: Bad healthcare data is misinformation, and an AI trained on misinformation gives confident answers that can hurt or kill people.​​

Governments and industry are beginning to see the problem and agree on an approach. At Congress’ direction, the FDA published guidance in 2024. It defined reliable data as accurate, complete and traceable, and required that data be relevant to the question asked. Regulators in Europe, Japan, China and Canada are adopting a 2025 international guideline built on the same dimensions. I published the first peer-reviewed implementation of the FDA standard in 2025, and it moved from study to practice: a peer-reviewed Pragmatic Registry of patients with moderate-to-severe asthma, built by applying advanced AI to key clinical concepts sampled and measured for accuracy against clinician review, patients linked across sources and data traced to its clinical source of truth.​

The principles are settled and the path is demonstrated. What remains is for the rest of healthcare to follow: building datasets with that rigor, training AI on them and carrying what is learned into care. Train AI on data built that way, and the woman I met in the ICU becomes a lesson the system never forgets, instead of a mistake it keeps repeating. ​​​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?