AI in Medicine: Beyond the Status Quo

Rafeed Alkawadri, MD | Human Brain Mapping Program

Years before AI became a business headline, we were asking what the brain’s electrical signals could tell us that we were failing to see. Our work on spike detection, brain–computer interfaces, and cortical mapping during sleep grew from that curiosity. At Yale surgical conferences, I remember turning gigabytes of data into slides that made the findings easier to grasp. The pleasure was in seeing an idea become useful.

Seeing companies develop these possibilities feels natural to me. I enjoy academia’s freedom to follow a question wherever it leads; industry brings resources to build and scale, within the constraints of a product and a market. I have never regarded its arrival as a defeat for academic research.

My thinking about health systems has changed, too. For years, I admired the preventive ambitions I associated with the Eastern bloc. I still value prevention, while increasingly appreciating the Western combination of free inquiry, democratic institutions, public investment, and enterprise. AI is an international achievement. I want an environment where discovery can flourish and reach patients.

What troubles me is how much of medicine’s future may depend on its willingness to reconsider its past. AI can already outperform clinicians on specific tasks. In a study of 50 physicians, an exploratory AI-alone assessment achieved a median diagnostic-reasoning score of 92%, compared with 74% for physicians using conventional resources. Those were written cases, not patient outcomes. My greater interest is what AI might reveal beyond familiar diagnoses: markers hidden in signals we cannot readily interpret ourselves. Our neural-network study, using six minutes of intracranial EEG, offers a proof of concept for sensorimotor mapping from intrinsic brain activity.

Medicine’s human obligations remain. A frightened patient needs someone who listens, understands what matters, and helps make a difficult decision. The patient–clinician relationship has measurable benefits, modest though the trial evidence is. I do not believe we can replace that relationship. I want AI to give us more room to practice that part of medicine well. Caution about that relationship is reasonable.

Other reasons for delay are harder to accept. Some computational steps in our work finish in seconds; recording takes minutes, and the clinical evaluation takes longer. Less burdensome testing could improve safety, subject to validation. I expected that prospect to be welcomed. I have also heard concerns about established procedures, the livelihoods attached to them, and decades spent setting standards. I understand why change feels personal. Yet a professional legacy should leave room for a better method to succeed, and payment should reward patient benefit even when fewer procedures are needed.

Funding can reinforce the same attachment. My concerns about NIH review include favoritism and a view of novelty that can undervalue useful improvement. Novelty matters, but so does making an existing method faster, reproducible, or clinically reliable. Discovering something a good physician could not previously measure matters, too. Replacing that physician is the wrong test of AI’s value. NIH’s 2025 reforms aim to reduce reputational bias; I welcome the effort and remain concerned about its translation into actual reviews. The wider problems deserve a separate discussion.

In epilepsy surgery, important guidance exists, including consensus on timely referral. Decisions remain complex, and invasive procedures deserve scrutiny. Familiarity alone should not make a method seem safer or better supported. Academic medicine’s corporate structures can improve coordination while restricting exploration; pharmaceutical, device, and other funding relationships also call for independent judgment. The association between industry sponsorship and favorable conclusions in drug and device research gives us reason to insist on transparency.

Two further obstacles deserve more attention. Clinical data are scattered across institutions, with differences in recording practices and populations. Privacy complicates sharing, and performance cannot be assumed to transfer. Research across 20 institutions shows one way to collaborate without exchanging patient records. An independent sepsis-model evaluation also shows why local validation matters. AI can reproduce existing inequities. Building dependable infrastructure takes work that a striking demonstration can easily conceal.

There is also the gap between possessing a tool and knowing how to work with it. In the physician study, access to AI did not significantly improve performance. That does not establish why, but it cautions against assuming access is enough. Clinicians need training, time, and workable routines for questioning and using its advice—the human factors addressed by DECIDE-AI. These are practical reasons for delay, even where enthusiasm is real.

Our AI editorial emphasized do no harm. I hold to that principle. It should also make us examine the harm of avoidable delay and the limitations of the care we already provide.

I suspect this change may be generational. As the internet became familiar to those raised with it, AI may become part of how a new professional generation learns to ask questions. Experience need not prevent that openness; youth does not guarantee it. I hope established colleagues will help lead the change, and I hope I am wrong about how long it will take. The prospect still excites me. I would like our generation to be remembered for what we helped make possible.