Do you have pain or burning with urination? Yes/No. Do you have diabetes or an immune disorder? Yes/No… So goes the online app that “interviews” a patient and determines if she has an uncomplicated urinary tract infection that could be treated through an “e-visit”. Computerized algorithms such as these are starting to appear everywhere, in outpatient urgent care, in inpatient quality checklists, and on the web, promising “immediate care.” These algorithmic diagnostic assistants are not perfect, but the best ones do possess some sort of competence in performing the healing task. How much of the doctor’s touch will they be able to replace?
As pure calculators, pure computers that simply follow instructions, none of the apps have “comprehension” of what they do. As machines, they cannot reflect on their roles or duties or results; they simply calculate by following instructions, branching and looping, at lightning speeds. A certain portion of a human provider’s skillset also involves nearly mindless memorization of algorithms: steps to follow in pursuing a diagnosis or treatment plan, remembering best-practice guidelines, or knowing the pattern of clicks to use in an EMR. How much of the providers skills involve algorithmic competence without global comprehension, and thus are susceptible to computerized replacement? A recent editorial in JAMA addresses this question in the context of errors that Machine Learning algorithms make. Can clinicians perform competently without holistic comprehension of a patient’s situation?
Two particular thinkers have illuminated our world’s dependence on competence without comprehension, Charles Darwin and Alan Turing. First, Darwin realized that blind evolution could create increasingly complex biological systems which appear to have been intelligently designed, but don’t have a designer. Genes do not comprehend anything, and yet can be remarkably competent in solving the environmental problem of survival and self-propagation. Next, Turing realized in 1931 that a simple mechanical device, the simplest computer, which executes only three instructions, increment, decrement and branch, could calculate anything. Today’s powerful smartphones and mainframe servers are still just Universal Turing Machines, mindlessly executing simple algorithms, albeit at mind-boggling speeds. Turing’s insight consisted of the realization that complex, and even seemingly “intelligent” algorithms, can be built up using processes with absolutely no comprehension – increment, decrement, branch.
“Both Darwin and Turing claim to have discovered something truly unsettling to a human mind – competence without comprehension.” -Daniel Dennett, From Bacteria to Bach.
As clinicians, these insights are particularly challenging. How much of what we do involves complex comprehension and not simply algorithmic competence? So far, our computer assistance involves only straightforward diagnoses such are urinary tract infections. But it’s not hard to imagine that in a few years these tools will recognize complex disease interactions and even psychological tendencies toward treatment effects based on answers to screening questions. How far will we be able to push competence without comprehension? Unsettling indeed.
Many levels of medical providers today seeminglhy care for patients quite well without understanding all the complexity behind their actions: medical assistants, nurses, perhaps even physician assistants and nurse practitioners. But as we follow this line of thinking, we realize that even the most subspecialized expert physician uses abstractions of pathophysiological processes to treat diseases. I treat heart failure while employing concepts like “neurohumoral” response and “atrial stretch”. At some point my comprehension breaks down at an organ, or cellular, or molecular, level, depending on how closely I follow the literature for each disease. There is an argument to be made that ALL medical providers utilize an algorithmic competence at some point, as we approach the limits of our knowledge. Even more unsettling.
One of the fundamental issues facing medicine in the coming generation will be attempting to understand how and why a human’s synthetic comprehension adds value to the merely competent algorithmic approach. Are we simply better at emotional empathy than an algorithm? Or is it our ability to use “intuitive” knowledge? What actually is contextual comprehension, and why is it valuable?