You’ve done it before: Perhaps you have a weird rash or feel a little strange, something beyond the usual flu or back pain. You Google it, becoming part of the 75 million Americans who use WebMD each month to check symptoms, rule out conditions, or to go down a hypochondriacal rabbit hole (unsurprisingly, “cyberchondria” is a real thing).
According to a BMJ study, online symptom-checking websites provide accurate diagnoses roughly half the time, amounting to millions of people worrying unnecessarily or, worse, breathing a sigh of relief when something’s actually wrong.
But what if artificial intelligence could accurately diagnose you—and save you a trip to the doctor’s office?
It’s not a crazy idea. The UK has integrated an AI-powered healthcare system called Babylon into its National Health Service, separating patients with urgent needs from those with more run-of-the mill illnesses.
And it’s simple. If you type, “I have a cough and fever. What’s wrong with me?” Babylon’s AI will ask for more details and run through a list of other possible symptoms to determine whether you should go see a doctor or whether you can buy over-the-counter meds and send yourself to bed.
Babylon shows how AI can act as a middle-ground between self-diagnosis and a visit to the doctor. Apps powered by AI provide services for patients who are too busy to get to the doctor or who live in remote areas; it can also help patients who have real, but non-life-threatening conditions.
These apps work by harnessing the power of an extensive database of conditions, indicators, and images—essentially, tapping the brains of thousands of doctors at once. Natural language processing allows these programs to understand the symptoms patients describe via chat and to translate the corresponding medical data into instructions.
One of the advantages of diagnostic AI is its accessibility to reach people in remote areas. For example, Babyl, the Rwandan version of Babylon, offers remote appointments with clinicians, fills prescriptions, orders lab tests, and issues referrals.
In addition to the convenience, diagnostic AI are “very cost-effective, and developing such techniques will likely aid in fixing the unaffordability of healthcare,” according to Bart Wolbers, clinical health scientist and researcher at Nature Builds Health.
Part of the savings is in avoiding unnecessary appointments, co-pays, and tests; the other part involves catching diseases earlier, when it’s cheaper and more effective to treat them.
AI can diagnose skin cancer as accurately as dermatologists. Most of us are familiar with the ABCDEs of assessing moles: asymmetry, uneven borders, color variation, large diameters, or moles that evolve over could indicate melanoma. AI is particularly good at picking up on these signs, because databases contain thousands of images of cancerous and non-cancerous moles, allowing algorithms to classify moles more precisely.
A Belgian study presented last year at the European Respiratory Society International Congress demonstrated that AI can help pulmonologists diagnose lung disease by more accurately identifying symptoms and assessing lung function tests, which can be too small and subtle for human eyes. Computers, however, don’t have this problem.
In fact, that’s the main draw of AI over humans — the potential decrease in human error and ability to spot things we might not be able to detect because, well, we’re only human.
“Supercomputers equipped with AI can provide alternative suggestions to medical professionals, which potentially cuts the amount of time it takes for a patient to be diagnosed and increases available treatment options,” Bridget Rooney, a health investigator with the Mesothelioma Cancer Alliance, told The Daily Beast.
AI’s reach extends to clinical trials. In 2018 the Mayo Clinic partnered with IBM’s Watson to match patients with breast cancer to accessible clinical trials covered by their health plans. The matching program increased the enrollment of breast cancer sufferers in Mayo Clinic’s own clinical trials by 80%.
“Clinical trials allow patients who have exhausted their options to potentially benefit from experimental treatment, increasing the possibility of remission,” Rooney said.
Consistent use of AI diagnostics could ultimately reduce misdiagnoses. “AI can be much more precise in the long-run, preventing the 10-20% of misdiagnoses that currently occur,” Wolbers told The Daily Beast. “Humans will thus get better health care, especially in cases where diagnoses are difficult.”
But AI isn’t a perfect cure-all. Rare diseases can be tricky for AI because, according to Rooney, inaccuracies are more likely when there’s “not enough data or information to make a conclusive decision.”
But even in those cases, AI could free up physicians to see patients with more pressing needs by separating those who can self-treat from those who need to seek in-person care. Ada, an AI-powered medical app used in over 130 countries, helps doctors make more precise and personalized diagnoses.
“Ada reviews multiple pieces of data and provides a probabilistic assessment, suggesting possible causes for symptoms and the likelihood of each,” Bethany DuFresne, Head of Communications for Ada, told The Daily Beast. “Individuals are presented with possible causes, a diagram illustrating the number of people who experienced the same symptoms and associated conditions, and relevant information to help them make more informed decisions about what to do next.”
In a pilot study, 92% of UK health care providers found it “helpful” to receive information gathered by Ada before a patient visit, largely because the app saved them time in reaching a diagnosis. The doctor-AI tag-team approach also prevents doctors and patients from ceding autonomy or control to an algorithm.
While any technology that facilitates accurate diagnoses and accessible care seems like a no-brainer, AI’s efficacy is a double-edged sword—the very process that enables it is outside of the realm of our understanding.
“Due to AI having a reasoning process entirely of its own, humans end up in a position where they fundamentally cannot understand why AI makes a decision, even though the decision is correct. That’s problematic because humans will start working with tools that they don't fundamentally grasp, which is scary at best and dangerous at worst,” says Wolbers.
And it’s not just medicine where this problem exists—AI in other fields, such as criminal sentencing, pose the same problem. Even the judges who use them don’t understand how they work.
Another problem beyond transparency: AI, like people, is finicky and doesn’t work broadly and equally well on all people.
Justin Sherman, co-founder of Ethical Tech, a nonpartisan initiative at Duke University, pointed out that “certain skin cancer predictors are terrible at making accurate predictions on darker skin, since most of the photos used to train the system, to teach it what to look for, are of white skin. These questions of AI fairness are a serious ethical and policy matter, but they're especially of concern in healthcare.”
That problem could be fixed over time with more data—the more use these apps get, especially across diverse populations, the more data they will collect and the better they’ll get at helping everyone. But that takes time, which could result in missed diagnoses or other potentially dangerous advice..
Given doctor shortages, increased wait times for appointments, and skyrocketing costs, the health care systems in remote or ravaged areas, as well as countries like the U.S., resemble the one in the 2013 science fiction dystopia Elysium. In the movie, the wealthy live in a habitat above Earth where they have access to technologies that can cure any illness. Everyone else remains on the ravaged planet below, sick and dying of conditions that could easily be treated on Elysium.
AI might not be able to level the playing field between all patients just yet, but it’s a start. And who knows—maybe AI could help us achieve the holy grail of health care: a shift from reactive, curative to preventive care, which would save time, money, and lives.