Three specialists in medicine, biology and machine learning discussed how AI might predict and help prevent disease in an episode of Microsoft Research's podcast series The Shape of Things to Come, hosted by Microsoft Research leader Doug Burger. The guests were Microsoft chief scientific officer Eric Horvitz; Eric Topol, executive vice president of Scripps Research, whom the episode description identifies as a Microsoft AI health advisor; and Microsoft principal research manager Ava Amini.
Diagnosis with AI
Horvitz described a Microsoft project that moved AI testing beyond handing a model a complete case. Instead, the system was placed in a clinician's position: it received an initial set of symptoms and history, then had to choose questions and tests, weigh information value against cost, and decide when to commit to a diagnosis. He said several frontier language models performed well, though he cautioned that comparisons with human generalist physicians have limitations.
Topol said the work, published in Science, was the first in which an agentic system reached more accurate diagnoses than both physicians and other frontier models. By his account, accuracy almost doubled relative to physicians, and at lower cost. He also cited an estimate from Johns Hopkins of 800,000 deaths or disabilities a year from diagnostic errors.
Unused imaging evidence and whole-patient models
Topol argued that medicine has overlooked an area where deep learning is better proven: medical images such as retina scans, chest x-rays and electrocardiograms. He pointed to a randomized trial of over a hundred thousand women in mammography and 44 randomized trials in colonoscopy, and said those findings are not being put into practice. According to Topol, almost no women in the United States receive state-of-the-art AI reading of their mammograms; the alternative, he said, is paying for it within a very limited network.
He said the next phase would be whole-patient models combining health records, images, genomics, proteomics and other data, to forecast when disease might occur and test what-if scenarios. Such models, he said, could already give some timing of risk, and he expected them to mature before a virtual cell does. He added that the aim should go beyond telling people they are at risk, toward prevention.
Barriers: adoption and incentives
Topol said what is missing is implementation. In his account, the medical profession is slow to adopt change, health systems face financial constraints and operate on very low margins, and neither physicians nor patients are demanding these advances. He described the reasons given as excuses rather than good justifications.
Burger relayed an argument from a recent conference that the economic model of healthcare is at odds with prevention. Topol replied that the United States has disincentives because of structures such as one-year insurance coverage, whereas elsewhere there are incentives to keep populations healthy.
Amini, who co-leads Project Ex Vivo, a Microsoft collaboration with the Broad Institute on cancer cell states, said enthusiasm for biological AI models must now be matched by clear problems to solve and adoption by experimental scientists. She said biology's multi-scale complexity makes integrating data from molecules to patients a hard machine learning challenge. Asked whether biology could ever be mapped as precisely as a computer chip, Horvitz answered no, saying its critical details will never be understood at a precise atomic level. Amini was more tentative, saying she did not think science would ever fully get there, but argued such precision is not needed for useful discoveries.





