Artificial intelligence has had many promising applications in medicine, but has also had some disappointing failures in realizing that promise.  IBM’s Watson was supposed to provide expert diagnostic advice by learning from interactions with doctors during examinations, as well as scouring existing medical literature to include the torrent of new papers covering new discoveries, procedures, and lessons learned that practicing doctors are too busy to read.  Despite years on the job at Cleveland Clinic and other places, Watson never quite secured a place at the bedside with doctors.  What is the key to succeeding with medical AI?

For starters, medical AI developers should target health insurers and quality management agencies, such as the Center for Medicare and Medicaid, not doctors directly.  Most experts are threatened by the idea of anything challenging their domain of expertise. To the extent that doctors control the decision making at hospitals, AI would need to provide advantages too compelling to ignore, smear, or subvert to get past the gatekeepers.  Some surgeons claim that surgical robots are better than the average surgeon, but not as good as a great surgeon.  What they don’t say is that it is better than a great surgeon on a bad day, and can help assure the quality and consistency of any surgeon using it.  Some doctors only use surgical robots for marketing reasons because their patients shop around for surgeons who use them.

Secondly, most forms of artificial intelligence depends on large amounts of high quality data to learn.  Unfortunately medical data is fragmented and riddled with errors.  Privacy concerns have frustrated many attempts to aggregate large quantities of patient data, and some studies have made the public even more uneasy about how easy it is to identify patients even with data that supposedly has the personally identifying information stripped out.  Even so, most medical systems are converging on a few large electronic health record (EHR) providers.  Epic, for example, handles the electronic health records for more than 250 million patients, and that data is aggregated in their data warehouse.  The UK centralized all patient medical records, to include their genetic, health, and physical data and provided it to medical researchers to advance personalized medicine and other medical research.

Unfortunately recent studies shows that most patient records are riddled with errors or omissions, and well-meaning government initiatives to improve healthcare quality have forced doctors to spend more time doing administrative tasks and much less of it interacting with their patients.  Many small practices closed or merged to form larger practices to cope with the administrative overhead increases.  Trained medical scribes might help with the quantity and quality of this data, but there will always be human error and omissions due to many causes, to include tacit knowledge.  The doctor may be thinking something and assume she will remember it or doesn’t even think to tell the scribe to record it because it is so obvious to her.

Success with surgical robots has been arguable, if only because many surgeons think they are superfluous.  Robotics and sensor technology promise to make surgical robots react more quickly and efficiently than a human surgeon, sense directly through diagnostic tools and imaging, and reach places an unaided human surgeon could not reach without a much larger opening.  So called “keyhole surgery” is in high demand by patients because it inflicts much less trauma and patients can heal faster.  Autonomous surgical robots have allegedly performed full surgeries on pigs with the surgeon observing.  Even if these devices cannot perform the full range of surgeries a human can perform, or does not do them as well, they could provide vital lifesaving capabilities to people in rural areas or without access to the best surgical care.

But diagnostics is the most promising application of AI for medical care.   For example, artificial intelligence developers sharpened capabilities in areas like face recognition, and have transferred some of that technology to recognizing other types of patterns.  AI can now read a CT scans just as well as a human radiologist, and such intelligence can be built directly into MRI machines, X-rays, CT scans, ultrasound, and other diagnostic devices directly to provide a high quality assessment on the spot.  AI relies upon high quality data, and in this application it can drink directly from the source.  

Another recent study examined how medical specialists identify problems by looking at things like X-rays.  It turns out that they don’t follow some sort of procedure or decision tree; they use recognition built from thousands of supervised interactions with demonstrable feedback on their performance.  Unfortunately the small amount of automation used in healthcare today has allegedly denied junior doctors and specialists the opportunity to learn alongside senior experienced doctors.  This training pipeline is also extremely labor intensive and requires years to produce a competent specialist.  Fortunately recognition is exactly what deep learning excels at replicating at scale.

But despite the hype about robots or AI displacing human workers, and even the advice from a medical AI expert that they “stop training radiologists immediately,” at least this last type of medical AI we discussed here will actually increase the demand for trained medical professionals, not replace them.  The American Hospital Association warned that the demographics of today’s medical professionals will lead to a dire gap in the coming years, with far more retiring doctors and even more nurses than are coming in through the training pipeline.  Midwest hospitals are advertising for nurses in places as far away as Australia and the Philippines.   

So the medical professionals we have in the future can reorient from bespoke and artisanal problem identification and focus more on leading teams of experts to more holistically treat the condition and eventually prevention.  With this kind of expert diagnostic availability more radically and universally available, millions of people will get much, much earlier detection of their problems and seek treatment while they have a chance.  Medical professionals will be in more demand than ever as a result.

This shift also suggests that medical schools should shift their focus to include more emphasis on leading teams effectively and human qualities required to interact with patients effectively.  Hospitals will need that ability to lead and inspire both to synchronize the entire healthcare experience of a patient from admittance, through the surgery, and on through recovery and rehabilitation, but also to get patients to comply with treatment.  

So human medical professionals will only become more important and valuable due to their intuition, leadership, and human skills as medical AI makes them more accessible, efficient, and affordable to many more people worldwide.  Are doctors up for this shift?