Research

AI-generated text

Researchers examine AI’s promise and limits in medicine

In a Microsoft Research podcast, experts discussed how AI could improve diagnosis, medical imaging and biological discovery while identifying barriers to clinical use.

Researchers examine AI’s promise and limits in medicine

Artificial intelligence could improve diagnostic accuracy, help interpret medical images and accelerate biological research, but translating results into patient care still faces substantial obstacles. That was the focus of a discussion hosted by Microsoft Research leader Doug Burger on an episode of The Shape of Things to Come podcast.

His guests were Ava Amini, a principal research manager at Microsoft Research; Eric Horvitz, Microsoft’s chief scientific officer; and Eric Topol, a physician-scientist, executive vice president of Scripps Research and Microsoft AI health advisor. Their conversation ranged from diagnosis and whole-patient models to virtual cells, aging and the design of biological systems.

Diagnosis and imaging

Horvitz described research published in Science in October 2026 in which language models were not simply asked to diagnose a fully documented case. Instead, starting with initial symptoms and medical history, the system could ask questions, select tests and decide when it had enough information to make a diagnosis. The task also involved weighing the information value of tests against their cost.

Topol said the agentic system achieved nearly twice the diagnostic accuracy of the physicians in the comparison, at a lower cost. Horvitz emphasized, however, that the comparison with generalist physicians had limitations; the results principally demonstrate the potential for such methods to assist clinicians in the future.

Topol argued that, amid the attention given to generative AI, healthcare has not made enough use of deep learning’s established capabilities in medical image analysis. He cited a randomized mammography trial involving more than 100,000 women and 44 randomized colonoscopy trials. In his view, published findings often fail to reach patient care because healthcare systems change slowly, may have limited resources and are reluctant to invest in tools they do not yet fully trust.

Whole-patient models and prevention

The participants said a next step could be models that combine electronic health records with raw medical images, genomic and proteomic data, and information about environmental exposures. Over time, these could improve estimates not only of disease risk but also of when illness might occur and how interventions might change the outcome.

Amini said a fundamental difficulty is biology’s multiscale nature: processes involving molecules, cells, tissues, organs and the whole organism would need to be linked coherently. According to Topol, some existing models can already provide clues about disease timing from limited data, but detailed patient data collected over long periods remains scarce at scale.

Topol expects mature whole-patient models to arrive before virtual cells capable of broadly predicting responses to different interventions. The aim, he said, should be to support prevention rather than merely report risk. He noted, however, that short-term insurance and financial incentives in the United States often work against long-term prevention.

The researchers did not consider precise knowledge of every atomic detail necessary for useful biological models. Amini said the more important question is what outputs are needed to enable a discovery or improve treatment; Horvitz argued that substantial benefits are possible despite uncertainty if causal relationships become better understood.

Clinicians’ work and research limitations

Topol said AI tools that document medical conversations are already spreading in the United States and can reduce the time clinicians spend on administration. Digital notes can also support tasks such as arranging tests or reminding patients about their care. Patients, meanwhile, are turning to chatbots with questions and for help interpreting results. Topol hopes that time saved could help restore the doctor-patient relationship.

In research, AI could help identify promising options before laboratory experiments begin. Amini said this could reduce resource use and the time needed for discovery, but real-world experiments cannot be fully replaced by virtual ones. Horvitz highlighted the potential for models to help determine which experiments are likely to yield the most new information within available resource limits.

A lack of suitable data is another constraint. Topol pointed to the role of the Protein Data Bank in the development of AlphaFold. Amini added that its protein structures are static, while proteins function dynamically. In her view, more data is needed to measure how molecules, cells and tissues function and respond to interventions.

As an example, Amini cited Project Ex Vivo, a collaboration between Microsoft and the Broad Institute that connects cancer-cell states with functional outcomes, including resistance to immunotherapy or targeted treatments.

Immunity, aging and designed biology

Topol identified detailed measurement of the immune system as an important missing layer in whole-patient models. Discussing an article published in Nature on October 27, 2026, he said studies of the thymus and immune cells could help clarify relationships involving cancer and responses to immunotherapy. Horvitz suggested that a better understanding of the underlying processes of aging could advance prevention across several age-related diseases at once.

The discussion also addressed synthetic biology. In earlier work, Amini studied genetic circuits in E. coli bacteria that respond to different sequences and combinations of chemical inputs. The researchers discussed AI-assisted design of biological systems for specified functions as a future possibility, while the creation of complex new organisms remains a longer-term prospect.

Horvitz expects AI to help turn some fatal cancers into chronic diseases and delay Alzheimer’s disease within ten years. Topol hopes for even broader disease prevention. Both framed these as expectations for the future, while the discussion emphasized that benefits in patient care will require data, experimental validation and practical implementation.