More than six years ago, I wrote an article about whether technology would eventually make most doctors obsolete, as futurists like Ray Kurzweil and others would certainly argue. Since then, we haven’t moved much closer to this scenario. In the face of recent news coverage and scientific papers about IBM’s Watson Health, I’ll take the opportunity to revisit my claim from 2012, which is that we are facing a false dichotomy here.
First of all, the reality of working in a hospital is still (maybe more than ever) full of redundancies, repetitions, routines and friction. For patients, it is an experience of waiting, being sent around, and talking to all kinds of service providers who are constantly in a rush.
It would be great if at least some processes were intelligently computer assisted: not having to remember a long list of pager and phone numbers, not having to write patient charts by hand and then copying them once a week, getting rid of unnecessarily complicated processes for ordering tests.
To think ahead a bit more, a virtual assistant could await dictation while doing a physical examination, and report test results as they come in, instead of having to search for them in fragmented software masks and diverse paper folders. Test results could be quickly tagged and sorted with adjectives like “caution”, “improving”, or “relevant”.
Electronic health record (EHR) interfaces should enable quick and seamless attachment of new external information, new symptoms, and updated findings on physical examination — in the way one would add a photo or business card to Evernote, only that it’s a secure account owned by the patient, and health care providers are temporarily allowed to use it. Think of a blood glucose measuring pen that scans a QR code on the patient’s wrist before making a small prick on the finger. For those requiring continuous measurement of heart rate, blood pressure, blood glucose or ECG, such devices obviously exist and are improving, but again there are no EHR interfaces.
An ideal EHR is easily writable and readable by computers and humans alike. Only with high quality EHRs can machine learning algorithms be properly trained.
An impressive display of AI capabilities, AlphaZero can dominate any other player in a complex game after a day of self study. But crucially, this is limited to any two-person, zero-sum game of perfect information. In contrast, predicting patient outcomes is more like a blown up version of poker, not only with plenty of face-down cards, but in fact not even the number of cards is known.
Healthcare data takes structured and unstructured forms, and includes demographics, clinical activities (screenings, diagnoses, treatment assignments), events reported by the patient or a relative (with varying degrees of reliability), data from medical devices, medical notes, physical examinations, lab results, biomarkers, images, genetic data and electrodiagnosis. All of this data needs to be organised and made accessible in an EHR. Inevitably, EHRs will always contain irrelevancies and information that can safely be ignored.
Only once the basics of ensuring good data quality are taken care of, can we start thinking about systems that will move from narrow to broad clinical decision support. Any AI system will only be as good as the data it is being fed.
Last fall, several news outlets reported that even after years of development, expectations in IBM’s Watson Health have not been fulfilled so far (see here, here and here). Sometimes treatment recommendations were incorrect, and sometimes the system would come up with information that the clinicians already knew.
Nevertheless, one can also find in the literature high rates of concordance for treatment decisions between human experts and e.g. Watson for Oncology. Improvements can be expected, although it’s important to bear in mind that technology becoming better is not a law of nature: it always requires motivated individuals and teams to move things forward.
We could choose an all-or-nothing viewpoint: “computers take over health care”, versus “This could never happen in my field”. Instead, in analogy to Garry Kasparov’s introduction of Advanced Chess, we should aim at Advanced Medicine as a symbiosis of creative and empathic humans, together with knowledgable brute force computers. A combination of heart and brain, if you will. Intelligence Augmentation (IA), rather than Artificial Intelligence (AI), is a goal that might be much more attainable than trying to surpass experts in their respective fields right away.
Why should it develop into a cooperation? There are those particular services in a society that one wouldn’t want to be left to a machine. Those include child care, education, health care, and care of the elderly. One way to look at it is, the more we let machines do the not-so-fun routine jobs, the more we are free to care for each other (at least for those who find fulfilment in doing so), explore the Universe and think about the meaning of life. For those who have read the highly recommended book by Yanis Varoufakis, “ Talking to My Daughter About the Economy”, I’m referring to the Star Trek outcome versus the Matrix outcome.
This also requires the realisation that, in contrast to production and industry, health care and all those other deeply personal services should not and cannot be subjected to the dogmas of cost saving and efficiency maximising. Instead of asking if the same level of care can be achieved with less personnel, the goal should be to have as many creative, empathic humans as possible in the game.