Medicine is one of the fields that AI promises to disrupt. After all, researchers have found that algorithms can predict heart attacks and diagnose pneumonia more accurately than doctors. But it’s hard to know why these black box algorithms arrive at their decisions–and if doctors can’t see the logic behind their conclusions, why should they trust them?
As researchers teach algorithms to help diagnose real patients, it’s an open question that’s not going away anytime soon. But a new study out of Stanford University takes a novel approach. The researchers used AI to increase the number of patients who get palliative care–the medical attention needed when someone is nearing the end of their life or has a very serious illness–while also ensuring that medical professionals could trust it. How? By keeping the doctors in the loop.
Palliative care is fundamentally about improving quality of life when a person has a serious, usually life-threatening illness, by helping them cope with pain and stress. But patients who need it don’t always receive it; a 2015 report showed that access to this type of care is inadequate across the country due to a lack of resources and personnel. Nigam Shah, an associate professor of medicine and biomedical data science at Stanford and coauthor on the paper, points out that this is important because an estimated 7% to 8% of patients who are admitted to hospitals need palliative care, while only 3.5% receive it. That leaves an estimated 1 to 1.8 million patients without palliative care, and even when it is available it may be offered too late because doctors will overestimate a patient’s chance of survival.
The Stanford researchers believed that machine learning could help with identifying more people that need this type of care–but to do so, they needed a proxy for when someone could benefit from it. The proxy they chose was predicting whether a patient will die in the next 3 to 12 months.
In the recently published study, the researchers unveiled an algorithm trained to analyze diagnoses, prescriptions, demographics, and other factors within electronic health records during that 3 to 12 month period before a patient passed away. Once trained, the algorithm was able to flag still-living patients in a hospital’s system that might be good candidates for palliative care. When Stanford Hospital’s palliative care team assessed 50 randomly chosen patients that the algorithm had flagged as being very high risk, the team found that all of them were appropriate to be referred.
Because the algorithm doesn’t make any decisions about care that patients receive–instead, it simply highlights people that may have been overlooked–the researchers believe it could help doctors give this end-of-life care to more people who really need it.
“The human doctor is always in the loop, and the program is not an automated clinical decision system,” says Anand Avati, a PhD student who studies machine learning at Stanford. “The current flow of referrals that are initiated by the treatment team continues as is. The program identifies among the rest of the patients, proactively, those who might benefit from a consult and might have otherwise slipped through the cracks.”
Avati, Shah, and the other researchers recently started a pilot program at Stanford Hospital to test whether the algorithm could improve the statistics on how many patients get the end-of-life care they need.
Their research also reveals how AI might be used in medicine so that doctors aren’t forced to trust it. Instead, it presents a model for how algorithms could work side by side with doctors, augmenting their skills so that more people get the care they need. “The program only helps the doctor be more operationally efficient,” Avati says.