Healthcare AI Should Make Medicine Cheaper and More Accessible, Not Just More Advanced

When I look at recent medical research from leading American institutions, I am often impressed by the intelligence, dedication, and technical sophistication of the researchers. Yet I have also felt increasingly uneasy about the direction in which our research system pushes these talented people. Too much intellectual energy is devoted to making already advanced medicine even more advanced, while comparatively little is devoted to making adequate medicine dramatically cheaper, simpler, and more accessible.

This is not an argument against frontier medical research. We absolutely need high-field MRI, photon-counting CT, advanced PET imaging, molecular diagnostics, precision oncology, and sophisticated medical AI. A small number of outstanding laboratories should continue pushing these technologies to their limits. Breakthrough research is inherently uncertain, and government support is essential because private markets will often underinvest in discoveries whose benefits may lie decades in the future.

The question is not whether we should fund frontier research. The question is how much of our limited research budget should be devoted to it. I believe the balance has become distorted. The academic incentive system rewards novelty, technical sophistication, publications, citations, and grants. A researcher who improves an already excellent imaging system by another few percentage points can often demonstrate the improvement elegantly and publish it. A researcher who spends five years trying to make a CT scanner ten times cheaper, small enough for a local clinic, easy enough to operate without dedicated technical staff, and inexpensive enough to maintain may create much greater social value but have a harder time producing prestigious papers along the way.

The researchers are not necessarily making irrational choices. They are responding rationally to the incentives we have created for them. Government R&D policy therefore needs to change the optimization problem. Instead of asking only, “How can we achieve the highest possible diagnostic performance?” we should also be asking, “What is the minimum amount of technology, infrastructure, labor, and money required to obtain clinically sufficient information?” This distinction matters enormously. 

Consider medical imaging. A state-of-the-art CT scanner is an extraordinary machine. For cardiac imaging, vascular studies, major trauma, subtle lesions, and complicated diagnostic questions, its capabilities can be indispensable. But many patients do not need that level of performance. A frail elderly patient with suspected pleural effusion, pneumonia, a large mass, a fracture, renal stones, or bowel obstruction may need an answer to a relatively simple clinical question. The physician may not need the best image that modern physics can produce. The physician needs an image good enough to make the next decision.

Dentistry provides an interesting analogy. Cone-beam CT systems have become compact and relatively inexpensive. In many clinics, operating the equipment is straightforward. The patient enters the scanner, follows simple positioning instructions, the scan is performed, and the patient returns to the examination room. We should be asking whether a similar philosophy can be extended to a much broader range of medical imaging.

Imagine a low-cost CT scanner designed from the beginning not to compete with a million-dollar hospital scanner but to serve local clinics, nursing homes, rural communities, and perhaps mobile medical units. It might rotate more slowly and therefore be more sensitive to motion. It might have lower temporal and spatial resolution. It would not be appropriate for cardiac CT or many sophisticated contrast studies. That is acceptable. Its purpose would be different.

Its design goals might instead be a purchase price comparable to an automobile, a small footprint, ordinary electrical requirements, minimal cooling, automated calibration, remote maintenance, automated patient positioning, and little or no need for dedicated on-site technical personnel. Artificial intelligence could guide the patient through positioning and breathing instructions, detect motion, perform image-quality control, optimize reconstruction, reduce noise and artifacts, and automatically transmit the study to a remote radiologist when necessary.

The research objective would no longer be maximum image quality. It would be maximum clinically useful information per dollar of total social cost.

That last phrase is important because the price of the medical test is only a fraction of the real cost of healthcare.

I learned this very clearly while helping my parents receive medical care. For a frail elderly person, a ten-minute physician visit can consume most of a day. Someone has to prepare the patient, help them move, arrange transportation, accompany them, wait at the hospital, and bring them home. Often the family member providing this assistance is already in his or her fifties or sixties and is not physically well suited to lifting and supporting an elderly parent. The hospital bill may actually be one of the smaller components of the true cost.

Yet our healthcare accounting system largely ignores these costs.The relevant equation should not be simply the price of the physician visit or CT scan. It should include the medical charge, transportation, caregiver labor, family time, waiting time, and the physical burden imposed on the patient.

Once healthcare is evaluated this way, many seemingly “expensive” services become economical. I would personally be willing to pay substantially more for a remote physician visit if it saved an elderly patient and family from spending an entire day traveling to and waiting at a hospital. Paying a physician twice as much for an effective remote consultation could still reduce the total social cost of the encounter dramatically.

This also changes how we should think about medical AI. Some of the most exciting AI research attempts to diagnose cancer, interpret sophisticated imaging, or compete with expert physicians. This work is scientifically valuable and should continue. But I do not believe replacing the expert oncologist or radiologist should be the first priority of medical AI.

A much more immediate objective should be to make sure that physicians spend their time doing work that actually requires physicians. AI can organize patient histories, summarize records, schedule visits, coordinate caregivers, route mobile equipment, monitor stable patients, identify which patients need attention today, arrange remote consultations, and determine which level of medical resources should be deployed. A physician should not spend valuable time searching through records, coordinating appointments, or performing tasks that a less specialized worker could safely perform under supervision.

The same principle can transform home care. A future elderly-care system could combine AI monitoring, remote physicians, local caregivers, visiting nurses or junior physicians, and mobile diagnostic equipment. Most stable patients would remain at home. A local caregiver could visit when physical assistance is needed. A junior physician or nurse could perform a basic examination while consulting a senior specialist remotely. A mobile imaging unit could visit the neighborhood when imaging is necessary. Only patients who genuinely require advanced facilities would travel to a major hospital.

AI would not replace the specialist. It would make the specialist far more productive. This is particularly important in palliative and end-of-life care. The objective of medicine changes when a patient is approaching death. Repeated blood tests, ECGs, sophisticated imaging, and hospital transfers may produce additional information without improving the patient's remaining life. The relevant question becomes simple: Will this test change what we do for this patient? If not, comfort, dignity, and time with family may be more valuable than another measurement.

AI could help enormously here, not by making a more sophisticated diagnosis but by simplifying logistics and continuously matching the intensity of care to the patient's actual goals and condition. There should also be a clear escalation pathway. A low-cost scanner in a local clinic should never pretend to replace advanced hospital imaging. If the image is uncertain or suspicious, the local physician should explain that uncertainty to the patient. AI could help quantify risk and explain the limitations of the test. A remote specialist could review the case. If a high-resolution CT, 3T MRI, PET scan, biopsy, or advanced specialist consultation is genuinely needed, the patient should then move to the next level.

Healthcare should therefore become more hierarchical rather than uniformly high-end: inexpensive and accessible medicine for common problems, remote expertise when uncertainty increases, and expensive advanced medicine when its additional information is actually valuable. Pricing should reinforce this structure. Advanced specialists and sophisticated hospital resources are scarce. They should not necessarily be made artificially cheap for every minor concern. Patients who medically require advanced care should of course be protected by insurance. But when someone with a low-risk condition wants substantially more expensive testing primarily for additional reassurance, some of that additional cost should reasonably be reflected in the price. Otherwise the system encourages everyone to consume the highest level of medicine regardless of its marginal benefit.

Government research policy can accelerate this transition enormously. I would like to see a substantial fraction of public medical R&D explicitly devoted to affordable healthcare technology and delivery. Whether the appropriate number is 30, 40, or even more than 50 percent should ultimately be determined by evidence, but the allocation should be large enough to change the behavior of universities, companies, and researchers.

The evaluation criteria for these grants should also be different. Success should not primarily mean another publication or a small improvement in a benchmark. Researchers should receive credit for reducing equipment cost, maintenance requirements, installation space, electricity consumption, operator training, patient travel, caregiver time, and total cost per clinically useful diagnosis.

Imagine a government challenge that asks engineers not to build the world's best MRI, but to build an MRI that a small rural clinic can afford, install, operate, and maintain. The machine would only need to answer a defined set of clinically important questions reliably. Once that performance threshold was reached, additional research funding would reward reductions in price, weight, maintenance, energy consumption, and operator skill rather than ever higher image quality.

That would create a very different research culture. Artificial intelligence makes this transition particularly timely. Cheap hardware no longer has to mean proportionately poor performance. AI reconstruction, motion correction, automated positioning, quality control, remote maintenance, scheduling, and specialist connectivity can compensate for some of the limitations of simpler hardware. We can combine inexpensive physical machines with increasingly powerful software.

Government tax policy could reinforce the same objective. Companies developing demonstrably lower-cost medical equipment and healthcare-delivery systems could receive enhanced R&D incentives. Clinics installing qualified low-cost diagnostic systems in underserved areas could receive investment tax benefits. Payment systems such as Medicare could reward providers when they reduce unnecessary hospital visits, transportation, hospitalization, and total patient cost while maintaining appropriate clinical outcomes.

The goal should not be cheap medicine for its own sake. Nor should it be lower-quality medicine imposed on patients who need sophisticated care. The goal should be something more rational: the right level of medicine, delivered at the right place, at the right time, with the least unnecessary consumption of human and financial resources.

For decades, medical progress has often meant adding capability: better scanners, more tests, more specialists, more monitoring, and more complex treatments. That model has produced extraordinary achievements. But aging societies now face a different problem. We cannot simply multiply expensive medicine indefinitely.

The next great medical breakthrough may therefore look surprisingly unimpressive in an academic presentation. It may be a modest CT scanner in a small rural clinic. It may be an inexpensive low-field MRI that requires almost no maintenance. It may be an AI system that decides that an 85-year-old patient does not need to travel to a hospital today. It may be a remote specialist who can supervise ten local physicians instead of seeing every patient personally.

None of these technologies may look as spectacular as the latest high-field scanner or an AI system claiming expert-level cancer diagnosis. But if they allow millions of people to receive adequate medical care more quickly, comfortably, and affordably, their contribution to human health could be much greater. 

Perhaps medical AI should not begin by asking how to make the world's best medicine even better. It should begin by asking how to make good medicine available to everyone.