Exploring the new challenge of combining the art of healing with the new science of machine learning.
A hard question to answer
Similar to diagnosis, prognostication is an extremely challenging, yet crucially important part of providing excellent clinical care. Unfortunately, we’re often poor prognosticators. As an Intensivist, I’m often faced with questions from families asking about what will happen when the breathing tube is removed. My brain searches and comes up with a very rough estimate: “minutes to hours,” “hours to days,” or “days to weeks,” usually with the additional, “it’s out of our hands,” leaving a grey fog in the room.
Machine learning and prognostication
Oncologist Siddhartha Mukherjee recently wrote in the New York Times Magazine about a cat who could prognosticate the upcoming demise of terminally ill patients in a nursing home. The story of the cat, Oscar, made it all the way to the NEJM in 2007. Mukherjee describes not a cat for every cancer ward, but a deep learning network, also known as artificial intelligence or machine learning, to facilitate prognostication.
Researchers at Stanford fed a deep neural network clinical data from 200,000 patients (160,000 to learn a “dying algorithm,” then 40,000 to test it on). The results were impressive—the algorithm identified 9/10 patients predicted to die within a 3-12 month time window, and 95% of patients assigned low probabilities survived longer than 12 months.
The art of talking to patients and their families
Talking to patients and their families about their ultimate demise feels like art rather than science to me. Part of me is comforted by saying, “it’s out of our hands,” and I sometimes even say, “it’s beyond science at this point.” However, as technology advances, the future may put this type of prognostication back into the world of tests, statistics, and numbers. It will be a new challenge to combine the art of healing with the new science of machine learning in medicine.