"Right now, the best solution is producing physicians who are comfortable with both data and professional judgment."

You cite Ayers disapprovingly, but he talks quite a bit about using both, comparing two models of integrating the two:

1) Use regressions and numerical techniques to get a prediction or suggested decision, then allow the physician to take that into account in her judgment
2) Use physicians' judgments as another class of data to feed into the formal model

The second approach generally outperforms the first, as doctor overconfidence and poor calibration means that most of the time when they override the model they make things worse.

"Right now, the NIH is divided up into disease-based institutes, but so much of today's research cuts across many institute boundaries. The NIH should reorganize"

That would seem to be best for advancing medicine and human health, but the disease-based organization stems in significant part from the fact that political mobilization for medical research is disease-driven. When someone has a relative die of cancer X, they try to support research into cancer X, rather than thinking the generalized, "my relative died of disease, I am going to encourage the development of better treatments for disease." So this could be politically difficult, and if successful it might weaken the strength of lobbying for more medical research spending.

I agree that medicine will not be fully deskilled for quite a while (although it' is extremely important to remember that less knowledgeable nurses and nurse practitioners already do just as well or better at patient care for much of medicine, and could be used much more often if not for regulatory restrictions), but I think there are large areas where the doctor's judgment should be overridden by relatively simple algorithms (where those algorithms include provisions for how to account for disagreements from physicians, e.g. by dividing objections into categories).

"However, medicine does have to become more data-driven, meaning we need physicians who have both good judgment and the ability to use computational tools or checklists or whatever to help them make evidence-based decisions in complex situations."

The argument is that if you let physicians have the final say they systematically override the formal models too often, and that the harm to patients from this outweighs the rare cases where the physician is right, so you need to constrain physician judgment with firm rules. This is an old result for medical expert systems: even if a doctor knows that in 90+% of cases of physician-expert system disagreement the computer is right, he or she will still consistently claim that in THIS case human judgment wins. This phenomenon means that adding a human veto often has a negative effect on decision quality.

The goal should not be to eliminate the role of doctors, et al., entirely by machines but to make their contribution to societal health more macroeconomically efficient and quite possibly to make their work more fulfilling at the same time. The trick is to replace those parts of the process that humans do poorly. Part of this is fast and accurate searching and testing of diagnostic options (this is more or less what is being talked about in the article). But this is only vaguely defined and overly simplistic in any case.

From an economics limits point of view we are have already reached both the upper and lower bounds of economic scaling for the medical profession. On the lower bound of individual work we can't squeeze more performance per hour of work out of doctors without pushing accidental death rates to very noticeable levels (you are already more likely to die from accidental death by prescribed medicine or medical treatment than traffic accident in a car). There are only some many hours (2000 hours = 8 hours/day @ 50 weeks a year) in a year any person can work and still productively deliver full value. There are only some many patients that can be seen per hour and still deliver value as an individual.

A standard technique to extend this limit is to create hierarchies of descending skill. This sets the upper-bound for human-delivered value. Lawyers and accountants use this in partnerships. The HMO structure is the medical version of this.

It's the lack of fungibility in professional services that ultimately limits how much this structure scales - and it's is the scaling solution of "last resort" for knowledge workers short of cutting humans out of the loop somehow. The fact that we are not getting medical value (and thus have a "healthcare crisis") from this very system is evidence we've hit the upper bound as well.

Using machines is how the same limits were transcended prior to the industrial revolution when the bottleneck was how much a single craftsman could deliver; the industrial revolution took over those bottleneck tasks. There are still craftsmen today but most are aided by the benefits of the industrial revolution and all are specialized is some way that involves not directly competing with what machines do better. Today with this idea for doctors it's computers, software and new testing equipment instead of looms and steam engines.

The primary process steps that humans have problems with are related to fatigue and attention. In terms of health and medicine this encompasses the surveillance phases of preventative, chronic and maintenance medicine. In each of these there is a need for medically-aware vigilance that would normally enable early cost-saving detection of symptoms and that would be prohibitively expensive to do with people: certainly doctors, even lower skilled medical and dubiously with the patient given medically-directed education. In both cases simple humans limits of fatigue and attention work against the efficacy of these.

There are studies that claim that preventative medical does work well but from what I've seen there is a central flaw in these. The physics says it should always be better and cheaper to detect problems early (as long as you believe the body works by cause-and-effect rather than magic). But this is as long as you can detect the changes. Most studies I've seen presume one or two visits to the doctor per year as "surveillance" for prevention. The problem is that more is prohibitively expensive but this number sample points per unit time is probably almost meaningless in terms of measurement noise for all but the most dominant pre-disease changes. If you could measure more often, the noise floor of the measurement can effectively be lower. For example if you could measure a key parameter every day you'd drop measurement noise floor by a factor of 20 for that parameter. Recent evidence of scale-free distributions in metabolic networks would imply that you can apply an 80-20 rule to what to measure: 80% of metabolic operation involve 20% of the metabolic components (enzymes, products, etc.) so you can enormous diagnostic mileage from measuring only dozens of parameters instead of millions. Since these 20% are interconnected you get extra information about one when measuring another and it's the early failures of the interconnected system that you're interested in anyway.

Economics dictates that automation is the only avenue to have sample rates this high and to measure multiple parameters of interest. And it's the body that will dictate which to measure, not what is cheapest to measure. It must be affordable by anyone and be cheap to operate. Say a hand-sized device costing $100 per month to operate and maybe as much as $500 at initial purchase would seem to be a reasonable target price. This device would ideally be non-invasive but that might not be possible initially - blood samples might be required for some parameters but those might not have to be daily - maybe weekly. On the other hand, the fact that diabetics can self-administer single or even multiple blood tests daily is proof this can be done even with blood samples.

There is the possibility that extremely sensitive gene arrays could allow single point measurement detection of pre-disease states (using blood samples) but from what I personally know about the company holding most of the patents in this area (I used to work there), innovation toward such low cost system is not a goal in any way, shape or form and won't be happening from them or others while their patents are in force in any case. After that? Let's just say that rust never sleeps but it will take a while to get there - longer than we have time for.

So what to do with all this data collected? No, we do not send it to some central computer to process! Creating a new bureaucracy of inefficiency is not what's needed here. And we also need to move toward more individual privacy in this country, not away from it. No. Instead the computation should be distributed because being tethered to your doctor (or hospital or insurance company or government) will only limit mobility and limit adoption. A computer (such as an iPhone or hand calculator) has enough computing power to handle the analysis and diagnosis. The internet can provide updates like it does for desktop computers. But what would such a computing device do?

Keep in mind that the AI program (now it would be called an "expert system") call "Mycin" was created in the 1960s on contemporary hardware of the time and it performed surprisingly well at the time. Newer hardware could to far better as today even the smallest computers in dish washers are literally 1000x faster than what Mycin ran on.

This is not to be all rosy about technology as a savior: such systems have their limits. But the point is to design with those limits in mind just as we are keeping human doctor's limits in mind: i.e. such automated diagnostic tools can be limited and less than perfect if they default to escalation to human doctors. E.g. Computer: "From you symptoms I can't determine a diagnosis with greater than 80% certainty; you must see a doctor to get a better opinion".

This would allow most medical surveillance to be done by individuals where the value delivered by human medical personnel is the worst and allow human medical personnel to deal with the exceptional aspects of medicine where they deliver the most value to patients.

Of course, if individuals have such machines, doctors will have better machines (naturally more expensive but only for escalation). The situations like the TV show "Mystery Diagnosis" really should never be occurring.

There is the question of doctor-patient "relationship" value that this doesn't get addressed by this. I'll simply acknowledge that both some doctors and patients find it valuable and it need to be considered somehow. This value isn't enough to justify avoiding automation, however. That is a certainty.

As a retired academic clinical pathologist with many years of experience in the application of numerical results to clinical practice, I feel compelled to provide a comment on how numbers appear to be evaluated in medicine. In science,of course, numbers are valued because they provide a degree of certainty about measurement. In medicine, unfortunatly, numbers frequently provide uncertaintity rather than certainty. This is particulary true in regard to a series of values for an anylite taken over a number of of days. Such values change both as a result of the reproducibility of the method of analysis and the course of the illness. The effects of anylitical variation on results is only rarely appreached by most clinicians. Nephrologists are a group clearly understand this problem, in part, because the values which they evaluate are used to determine when patients should be dialyzed. Regretfully, they are one of the few groups of specalists who are aware of this issue. I have often suggested that numerical values be removed from laboratory reports and be replaced by N for normal, an up arrow for an increased value and a down value for a decreased value. Removing numbers from medicine would help to recognize that it is an art rather than a science.