The robot will see you now: Health-care chatbots boom but still can’t replace doctors

Health-care chatbots have been used for years, but they exploded in popularity over the past year as the system became overwhelmed with covid-related questions and treatment

Scout’s job is just to ask questions. A lot of them.

How long have you had a fever? Are you feeling short of breath? When you bend your neck forward, is there pain that is so severe it makes you want to cry?

If so, Scout might recommend a trip to the emergency room. Scout is usually overly cautious about these things. It kind of has to be, as a robot.

Scout is a conversational chatbot made by health tech company Gyant and used by Intermountain Healthcare in Utah to tell patients what they should do when they feel sick. It may suggest getting some good rest, or setting up a doctor’s appointment or, oftentimes, making a trip to the ER or an urgent-care facility.

Chatbots such as Scout are often called “symptom checkers” because they are trained to ask about your ails and escalate serious issues to doctors when specific concerning conditions come up. These health-care chatbots have been used for years, but they exploded in popularity with health systems and insurance companies over the past year as the system became overwhelmed with coronavirus-related questions and treatment.

“All kinds of technologies that can be used remotely were greatly accelerated during the pandemic,” said Mildred Cho, a professor of pediatrics at Stanford’s Center for Biomedical Ethics. That includes telemedicine appointments, chatbots and even virtual dental screenings.

But as the use of chatbots in health care continues to accelerate, researchers and some health-care professionals are pushing for more data about how accurate the technologies are. Chatbots are dealing with sensitive and often urgent information, and getting the right answers is critical. Companies developing the technologies say they err on the side of caution and regularly review and test their results with trained medical professionals.

Many chatbots that evaluate patients are using artificial intelligence to understand the symptoms, match them to likely illnesses and map the probability of an outcome. Developers input thousands of pieces of medical literature into the computer systems, and they train models to understand what symptoms coincide with which diagnoses.

When patients tell a chatbot like Scout what ails they have, it can match the answers to possible diagnoses to understand which question to ask next or what to tell the patient to do. Gyant said its system has been trained on 17 million patient interactions, pulled from patients using the system.

Systems like these are fairly sophisticated, said Hamish Fraser, an associate professor of medical science at Brown University. They use complicated probability models to predict outcomes, and they have evolved over time from early versions that were simpler and similar to decision tables.

Some chatbots, including many that were used to screen for coronavirus tests, are still using technologies similar to simple decision trees. Have a persistent fever and lost your sense of smell? Yes, you qualify to get a test. But overall, the artificial intelligence technology is starting to become more advanced, and some companies are beginning to use true machine learning, the act of a system learning new things from examples, Fraser said.

Even as more health-care companies try out chatbots to triage symptoms, data on the efficacy of the different chatbots is limited and the technologies need significant advancement to come close to imitating a doctor’s brain, researchers say.

Many developers will insist that’s not the goal, anyway.

“When someone thinks about an AI chatbot, I think the first thing people leap to is, we are trying to replace a doctor,” said Andrew Le, co-founder and CEO of Buoy Health, which makes a health-care chatbot. “What we are really trying to do is replace the act of searching for symptoms, being scared and feeling not sure what to do.”

Retailers have been testing out this type of chatbot technology for years — an automated assistant can help you make an appointment with Apple’s technicians, or one could help you find the best shade of lipstick on Sephora.com. These smart assistants have even made their way into our homes in the form of Amazon Alexa and Google Assistant.

Tech giants, including Microsoft, Amazon and Google, have been aggressively pursuing avenues into the world of health-care technology, just as health-care systems are opening up to the idea of embracing more tech tools. But tech adoption in health care has been slower than in some other areas because of the sensitivity of patient data and the importance of every decision when dealing with people’s health.

So far, many health systems use the chatbots in limited capacities to act somewhat like a receptionist — the goal is to figure out what your ailment is and get you to the right place, whether that be an appointment scheduler, an emergency room or a telemedicine meeting. Or, throughout 2020, to a coronavirus test.

Intermountain started using Scout in March 2020, when questions started to roll in from people who were worried they could have covid-19, which had started to rapidly spread across the country. The Utah poison control center designated a phone line to handle coronavirus queries. Intermountain added a dedicated line to its usual ask-a-nurse call center, said William Beninati, a critical-care physician at the health system.

In less than three hours, the line was overwhelmed. Intermountain added staff, but it wasn’t enough. No matter what they did, the call center kept getting overwhelmed. So, the team turned to artificial intelligence.

Scout came in and started asking patients basic coronavirus screening questions to find out if they were eligible to get a test for the virus; at the time, tests were severely limited.

The technology for the covid screening was pretty straightforward, Beninati said. It was basically a set of conditional logic that relied on set questions and answers. Now, the health system has worked with Gyant to implement a more advanced artificial intelligence system to check patients’ symptoms and direct them to the right form of care.

The symptom checker, added in June last year, asks patients around 30 questions to narrow in on, say, how bad your headache is, how long it has lasted and if it comes paired with any other symptoms. Scout tries to rule out emergency or more serious situations before suggesting the most common answer: It could be a migraine. Or perhaps a tension headache. If it gets worse, set up an appointment.

The hospital is hoping to eventually add a way to have a “doctor’s visit” with Scout, where the chatbot would offer care that would be quickly reviewed and verified by a real, human doctor.

Developing this hits on one of the biggest challenges facing health AI developers — data. Health-care data is sensitive, regulated and often not aggregated in an easy format to use for software. Developers often have to rely on clinical texts and literature — similar to what medical students use — and some publicly available databases until they can collect their own information and case studies from users.

Gyant works with teams of physician reviewers to make sure the technology knows which symptoms need to be escalated to a doctor right away. But that can still be challenging because the chatbot gains more experience with common conditions, such as stomach flus, than illnesses that are less common but can be more dangerous, like meningitis. When it isn’t sure, the technology is taught to tell the patient to seek help.

“It’s challenging to make the right trade-off calls, but in general, we err on the side of caution,” Gyant CEO Stefan Behrens said. “We’d rather be too conservative.”

That seems like the safest course of action, but health-tech researchers also point out that overdiagnosing can be an issue when it causes crowding at emergency rooms or urgent-care centers. Those visits also can come at significant costs to patients and providers.

Companies are careful to say the chatbots will not actually diagnose patients but rather point them to the correct next step, whether it be an emergency room or a good night’s rest, said Sven Laumer, a professor of information systems at the Friedrich-Alexander University of Erlangen Nuremberg.

“They do not want to take over this responsibility,” Laumer said. “This is a big challenge for all digital health-care services.”

Before using Scout, for instance, a user has to acknowledge a screen agreeing to use the bot at their own risk. The message spells out the chatbot is “not a diagnosis or a substitute for professional medical advice.”

Still, when chatbots are used in lieu of phone calls, more as a way to triage patients and direct them to the right place, research has found that patients tend to like them and find them pretty easy to use, even when people are feeling sick.

Sometimes it goes even further — the researchers behind mental health chatbot Woebot found that many users were establishing a sort of bond with the bot within three to five days and were turning to “talk” with it regularly.

Woebot uses artificial intelligence to detect what users are saying to find the most appropriate responses. But unlike some other bots, it does not generate responses in real time. That’s too risky with something like mental health, said Woebot founder and president Alison Darcy.

Instead, Woebot pulls from canned responses that have been written and vetted by mental health clinicians.

“We could never risk Woebot saying something that undermined someone’s emotional well-being,” Darcy said.

She has seen it happen before. A Woebot engineer was trying out a different chatbot technology developed outside the company and sent it a picture of him playing the guitar.

“Are you playing a guitar?” the bot asked. “Yes,” the engineer responded. “But I’m not very good.”

That’s when the bot, clearly trying to pull context clues from the conversation, made a risky choice.

“You know what I learned?” it asked. “That you aren’t very good. And I learned that you aren’t very good. And also that you aren’t very good.”

Darcy confirmed then that it’s better to have a bot with curated responses, that checks in to make sure it is responding to the correct prompts. (“It sounds like you are talking to me about grief,” Woebot will ask. “Is that right?”)

Chatbot technology is advancing, and hospitals and health systems are increasingly keen to try out the technology as they warm up to virtual care after some success using it during the pandemic.

But as the world slowly opens up again and people get more comfortable with in-person activities — and with significant technological advancements still in the works — it’s safe to say that your doctor will continue to be a human. At least for now.

This article was originally published on The Washington Post on July 26, 2021.
Written by: Rachel Lerman

Shop & Compare Medicare Plans today for the Special Enrollment Period (SEP) open until August 15th, 2021.