Researchers have trained an AI system to accurately predict if a person would die in three to twelve months. If you could know the day, or even year, of your death, would you choose to know ahead of time?

2019

AI DEATH PREDICTOR

In 2016, researchers at Stanford University were able to train an AI system to accurately predict if a person would die in three to twelve months. This remarkable application is included in this book as representative of the wide variety of roles that AI and deep learning will play in coming years.

Palliative care usually involves providing relief for a patient’s pain, stress, and other symptoms when the patient has a terminal diagnosis and no cure is expected. Knowing when such specific care is warranted may have beneficial effects for the sufferer, family, and caregivers—and help determine when such care would be most effective. To create the AI “Death Algorithm,” the Stanford team used information from about 170,000 patients who had died with, for example, cancer and heart and neurological diseases. Various information from medical records—including a patient’s diagnosis, medical procedure, medical-scan codes, drugs prescribed, etc.—was used as input to “teach” the AI system. Then a deep neural net was trained, with various internal weights adjusted for the neuron units. The deep neural net made use of an input layer of 13,654 dimensions (e.g., codes for diagnoses and drugs), 18 hidden layers (each 512 dimensions), and a scalar output layer.

In the end, nine out of ten people predicted to die within three to twelve months did die within this time frame. Also, 95 percent of those whom the algorithm determined would outlast twelve months did live longer lives. However, as physician Siddhartha Mukherjee explained in a recent New York Times article: “[The deep learning system] learns, but it cannot tell us why it has learned; it assigns probabilities, but it cannot easily express the reasoning behind the assignment. 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.” Nevertheless, research on these AI death predictors continues, and in 2019 a team of experts at University of Nottingham showed that machine learning could outperform traditional methods at predicting premature deaths, based on demographic, biometric, clinical, and lifestyle factors.