Healthcare is a Computational Problem

“The goal of healthcare is to keep people healthy and failing that to fix or alleviate the problem.“

This is our mission statement so to speak. The actual actions we take to get there, the verbs of medicine if you will, are:

When we look under the hood, we find that each of these actions are computational problems. This means they consist of data collection and systematic processing to arrive at a solution.

Let’s look at Diagnostics as an example. Diagnostics is the task of classifying illness, classification being a classic computational problem

Mystery Patient

A patient arrives in my emergency room, I know nothing about them save that they are there. What’s their diagnosis?

Absurd as this may seem, we can begin answering the question even at this stage. This isn’t guesswork, there’s data to draw on.

We know, for instance, the most common complaints in an emergency room. We also know the most common diagnoses for each of those complaints. We can use this and similar data to calculate a so-called “Pre-Test Probability” for all possible diagnoses. This is the probability of a condition before additional information is available. We can then create a ranked list of the most likely diagnoses. Admittedly, at this stage the list is both enormous and useless.

Next, I’m told by the triage nurse that the patient is a male. In a flash, the list of possible diagnoses is slashed and re-shuffled. With this fact we calculate a “Post-Test Probability” for each diagnosis on our list. This is the probability of a condition after additional information is available. In this case, the probability of hundreds of gynecologic conditions is zeroed out. Of course, the list remains long and impractical.

The triage nurse now tells me the patient is here for abdominal pain. Again, the list is trimmed and re-shuffled, the patient is unlikely to have an ankle sprain. I’m then told the patient is 45. Again, trim, re-shuffle — hundreds of pediatric conditions are excluded. We go on to add details shared by the patient, results of blood tests, imaging, and so on.

Each datum is computed upon in a specific way and the solution re-calculated. This is what we mean by a computational problem. The process is summarized in the diagram below.

DDx: Differential Diagnosis. A list of potential diseases given available information.

Each arrow in the diagram represents a computation. The computations are not arbitrary. They are not subject to the whims of the clinician, nor their best judgment, nor where they went to school, nor what they had for breakfast. Instead, the computation is systematic, quantitative, and empiric.

Of course, this is an oversimplification and only one, quite linear, possible approach. But you get the idea.

Cold Hearted

Some colleagues I’ve talked to object. They say this is too utilitarian a view. They say healthcare is a human problem, a spiritual problem, a question of empathy, compassion, and so on.

I agree, but only to a point.

We shouldn’t conflate what we do with how we do it. The core functions of healthcare should be performed with patience, empathy, and kindness. But patience, empathy, and kindness alone won’t cure a case of pneumonia. An X-ray and course of antibiotics will.

Clinical decisions should be computed (including accounting for such things as patient autonomy, informed consent, etc.), and thereafter carried out with all the humanity we can muster.

The Gap

Of course, I don’t do this when I diagnose my patients. I can’t. Neither do my colleagues because they can’t either. The human brain isn’t well suited to heavy, exacting computation. That’s where computers come in and it’s why diagnostics should ultimately be automated.

In fact, the rest of medical decision making, such as prognostication, treatment planning, and so on are all computational problems too.

That’s why I believe the inevitable endpoint of healthcare is almost complete automation.

Today, too much of the core functions of healthcare are left to clinical judgment, expert opinion, and a thousand cognitive biases. The gap between how practice and the ideal are hundreds of thousands of medical errors, injuries, and deaths.

The point is urgent and the stakes are high.

Recognizing that healthcare is a computational problem allows us to focus on lasting solutions: developing automated clinical decision making systems.

TL;DR

Healthcare problems are computational problems. The sooner we recognize this the sooner we can focus on building computational solutions and ultimately transition to automated clinical decision making.