It just goes to show me! Yesterday I wrote a really nice introduction to some reviews of articles on artificial intelligence and today the whole thing is gone, vanished in a cloud of electronic smoke. Machines can do some pretty amazing stuff, way beyond the ability of human beings when it comes to remembering, sifting and calculating data. But they’re machines, not humans. If the laptop I’m sitting at again today were really intelligent it would have known better than to flush yesterday’s work down the electronic toilet. So I’ll start over.
- Marc Ringel, MD
ARTIFICIAL (INTELLIGENCE?)
I was inspired to choose artificial intelligence, also called machine intelligence, by a book I just read, God, Human, Animal, Machine: Technology, Metaphor, and the Search for Meaning by Meghan O’Gieblyn. As an earnest seminary student, the author spent years steeped in questions about god and ethics viewed from within a fundamentalist theological framework. Doubt led her to dive into philosophy, through which she developed a wide and deep understanding of the underpinnings of western metaphysics and ethics. O’Gieblyn became a “tech evangelist” for a while, attempting to satisfy her obsession for ultimate meaning with the concept of the “singularity” predicted by Ray Kurzweil and others, when machines would become smarter than people and take the lead in probing the mysteries of the universe.
If you’re not going to read O’Gieblyn’s fascinating tome, here’s the underlying theme. No matter how hard we strain against our limitations, we will always live in a world of human dimensions. We have to believe something about how the world works, even when couched in the “objective” language of pure science or the “hard facts” of machines. It is up to us to assure that the principles we build our world on are human ones. For me, the need to humanize is particularly pertinent to healthcare, the world I’ve practiced in, written, spoken and politicked about my whole adult life.
I’m not going to do any sort of sweeping survey of artificial intelligence here, just a review of seven of the dozens of articles I’ve collected in the last few years about the role of AI in healthcare. It’s a moving target. We can expect ever more powerful machines and programs to make ever more sophisticated contributions to the patient care enterprise. Here’s the bottom line. Healthcare is still a consummately human endeavor. As you can see in the articles cited below, electronic devices already assist us in caring for patients. In the process, though, we must not lose sight of the ultimate goal, which is to serve and to heal each other.
Characteristics of publicly available skin cancer image datasets: a systematic review
The Lancet Digital Health 2022
For most AI you need to start with lots and lots of data that have been labeled by humans, then used to instruct the software to recognize patterns. Language translation programs are trained, for example, on as many bilingual documents as the designers can get their hands on. The authors of this study found 21 open access atlases of 106,950 skin lesions labeled with diagnoses, a good place to start teaching an automated program to tell one skin ditzel (a medical slang term for lesion) from another.
View Article
Human–computer collaboration for skin cancer recognition
Nature Medicine 2020
…speaking of dermatology…The message in this 13-page article is simple. Artificial intelligence can enhance the ability to make accurate dermatologic diagnoses, particularly in support of less expert professionals. But it does not dispense with the need for humans to assure that skin lesions are precisely identified which, needless to say, is the first step in choosing how to treat them.
Development and evaluation of a machine learning-based point-of-care screening tool for genetic syndromes in children: a multinational retrospective study
The Lancet Digital Health 2021
…and speaking of pattern recognition…When I was an intern in pediatrics I learned of the bona fide diagnosis FLK, which stands for “funny looking kid.” We applied that term to infants whom we suspected were born with some sort of congenital problem but didn’t know what it was. I expect, based on changes in politics and language, that the term, “FLK,” is way unacceptable today. Besides, we have considerably better scientific tools now, including DNA analysis, to identify variations from standard (however you define it--a subject for another essay) than we did when I was in training in the age of dinosaurs. Nowadays clinicians are much less likely to need to reveal their ignorance by resorting to a term like “FLK.”
The face is especially likely to bear signs that something in the genome is different. This article reports on a machine learning program that tested to be about 90% accurate in pairing images of faces with congenital syndromes. The best thing about this software is that it can be used remotely, in locations where professionals who are highly trained in pediatric genetics may be very hard to come by. Not to be underplayed is the fact that the program did just as well with Hispanic and African, and just a little less well (80% accurate) with Asian children, as it did with White ones. Early diagnosis and intervention can make an enormous difference in quality of life over the course of these kids’ lifetime.
External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients
JAMA Internal Medicine 2021
“Pattern recognition” does not just refer to patterns you can see, like picking a bottle of catsup out of a picture. (N.B. The standard example, used by virtually everybody when discussing visual pattern recognition is training an AI system to pick cats out of pictures.) To a computer, images are just arrays of data, arranged in pixels. Any kind of data can have patterns lurking within it, including first signs of a medical problem. Pattern recognition has been tested and employed widely in early identification of sepsis, an infection of the bloodstream that can cause death in up to half of patients who bear that diagnosis. The earlier sepsis is detected and treated, the greater the likelihood of cure. This article reports on a computerized early warning system employed at the University of Michigan Hospital that processed vital signs, test results, symptoms and clinical observations, calculated and reported every 15 minutes an Epic (the name of the computer system provider) Sepsis Model (ESM) score, an estimate of the risk of sepsis. To make a long story short, the system did not help human clinicians diagnose sepsis any earlier. What it did accomplish was to exacerbate “alarm fatigue,” that is, inattention among staff caused by being subjected to too many computerized warnings.
Quantitative Methods for Analyzing Intimate Partner Violence in Microblogs: Observational Study
Journal of Medical Internet Research 2020
The servers that host social media are lousy with data, petabytes and petabyte of it. (A petabyte is more than a million gigabytes.) Some of the richest people in the world have made their fortunes mining the data they collect from social media users. These data can be turned to other purposes than making billions of dollars. In this report machine learning was used to compare natural language patterns (that is, what people actually speak and write) in two Twitter datasets, #WhyIStayed and #WhyILeft, comprised of 8767 tweets each. There’s nothing very earthshaking here. It’s hard to imagine how society could use directly the insights generated by the computer system to address partner violence without violating privacy. But the information gathered will be quite useful to social scientists who are trying to understand intimate partner abuse, not to mention to the people on the frontlines actually dealing with it.
Construction of Genealogical Knowledge Graphs From Obituaries: Multitask Neural Network Extraction System
Journal of Medical Internet Research 2021
My medical specialty is family medicine. We family docs pride ourselves in seeing a patient in the broadest possible context, which means not just integrating our understanding of all the different organ systems that our colleagues in other specialties may claim as their exclusive territory, but also taking into account the family and the community in which our client resides. A few decades ago there was even an effort by some family medicine mavens to organize medical records into family charts which would contain the folders of each individual in a family, as well as a genealogy graph of the whole clan. (You can imagine how impractical this was when charts really were made of paper.) As critical as family history may be to understanding both medical and sociological aspects of a patient, gathering and recording these data is hard to accomplish; time consuming and often inaccurate. With the help of AI, the authors of this study sifted the data in 1700 obituaries published online in Minneapolis and St. Paul and came up with some pretty rich genealogical information. It’s too late for this intelligence to help me in my office practice but it should be of great use to researchers doing genetic studies.
The US Government Will Pay Doctors to Use These AI Algorithms
Wired 2020
Machines have gotten pretty darn good at recognizing when a retina is showing the effects of diabetes, the second greatest cause of blindness (after macular degeneration) in American adults. Software can also evaluate with good accuracy a head CT scan to rule in or out an acute stroke caused by a blood clot. The earlier this kind of stroke (as opposed to one caused by bleeding into the brain) can be diagnosed the sooner a medical team can administer a drug that dissolves the clot, thereby preserving precious brain tissue. The US Centers for Medicare and Medicaid Services is so enthusiastic about these augmentations to human practitioners (duly vetted by the Food and Drug Administration) that it is actually paying clinics for automated retinal screening and, with lots of restrictions, for AI-augmented emergency brain CT scans. Final decisions about what is to be done always rests, as they should, in the hands of human professionals.
Where do these articles leave us? Where (at least in my hands) the facts always lead. We are only just scratching the surface of what technology could do for healthcare, but only if technical knowhow is thoughtfully applied and continuously evaluated. We must never lose sight of the fact that innovation is not and end in itself, but a supplier of tools to be wielded by human beings.
