Can you teach an algorithm to know when you are most likely to die? One Stanford University research team is answering yes, reporting in a new paper that they have taught an algorithm to predict patient mortality with startlingly high accuracy.
Having an algorithm know your expiration date can sound like a dystopian concept, but the Stanford researchers said that they created the algorithm to benefit patients and doctors by improving the end-of-life care for ill patients. The researchers cited past studies that found the overwhelming majority of Americans would prefer to spend their last days at home if possible, but only 20% get that wish realized. Instead of getting to spend their final days at home, up to 60% of patients spend their last days in the hospital receiving aggressive medical treatments.
By creating a deep learning algorithm to predict patient mortality, doctors can better inform patients about their end-of-life options before it is too late, allowing more patients to get their spiritual and cultural final wishes met, the paper argues.
Research: There’s an algorithm that can predict patient mortality for critically ill patients
To train itself and make its predictions, the algorithm was given the electronic health records of about 2 million patients from two hospitals between 1995 and 2014. From there, the researchers identified around 200,000 patients suitable to be studied, and selected a smaller group of 40,000 patient case studies to be analyzed. The algorithm was then given the following marching order: “Given a patient and a date, predict the mortality of that patient within 12 months from that date.”
The results were highly accurate. Nine out of 10 patients died within the 3-12 month window the algorithm predicted they would die in.
Relax, doctors won’t be losing their jobs to machines
But the algorithm is not going to be replacing doctors anytime soon. The algorithm could only predict when selected patients were going to die, but not why or how. “The scale of data available allowed us to build an all-cause mortality prediction model, instead of being disease or demographic-specific,” Anand Avati, a Ph.D. candidate at Stanford’s AI Lab and one of the author’s of the paper, said.
For palliative care physicians, the algorithm’s focus on the timeline is still useful since their work focuses beyond the initial patient diagnosis and why someone is sick. If patients are told about their mortality after the three-month window, it’s too late to start proper end-of-life care, while being told more than a year out is too early to prepare for palliative care.
But more and more professionals need to learn to work with AI
The researchers said that doctors are still needed to fairly interpret the algorithm’s probability scores for both ethical and medical reasons. “We think that keeping a doctor in the loop and thinking of this as ‘machine learning plus the doctor’ is the way to go as opposed to blindly doing medical interventions based on algorithms,” Kenneth Jung, one of the author’s of the paper, said.
Commenting on the AI-based system’s power, physician Siddhartha Mukherjee said, “Like a child who learns to ride a bicycle by trial and error and, asked to articulate the rules that enable bicycle riding, simply shrugs her shoulders and sails away, the algorithm looks vacantly at us when we ask, ‘Why?’ It is, like death, another black box.”
This article was first published on January 19, 2018.