Deep learning algorithm identifies candidates for palliative care—and predicts death

A novel deep learning algorithm could hold the key to proactive palliative care—and predict patient deaths, according to a paper published by the Stanford University School of Medicine in California.

Just one-fifth of Americans are able to spend their final days at home, PhD candidate Anand Avati and co-authors wrote, but studies have indicated that around 80 percent of the U.S. population wishes they could. This gap is bridged somewhat by palliative care—the specialized approach to relieving pain and stress in patients with life-limiting illnesses.

And though the number of dedicated palliative care teams at larger hospitals is growing—from 53 percent in 2008 to 67 percent in 2015—less than half of the 8 percent of admitted patients eligible for palliative care actually receive it.

“Though a significant reason for this gap comes from the palliative care workforce shortage, and incentives for health systems to employ them, technology can still play a crucial role by efficiently identifying patients who may benefit most from palliative care, but might otherwise be overlooked under current care models,” Avati and colleagues wrote.

Clinicians tend to under-refer eligible patients for palliative care, they said, due to overoptimism, time pressures or treatment inertia. This can be a direct path to more aggressive treatment in a patient’s final hours, which often contradicts that individual’s wishes.

The researchers outlined a method for identifying palliative care candidates using deep learning and electronic health record (EHR) data, which is currently being piloted and pending Institutional Review Board approval. Their algorithm identifies a host of factors upon a patient’s arrival onsite—including a predicted death date.

The system is based on a deep neural network trained on EHR data from previous years, Avati et al. explained in the study. The model uses three-to-12-month mortality projections to predict a patient’s possible need for palliative care.

“Our predictions enable the palliative care team to take a proactive approach in reaching out to such patients, or conduct time-consuming chart reviews of all patients,” the authors wrote.

Indeed, predicting a patient’s need for palliative care manually can be both expensive and time-consuming. Palliative care clinicians are often faced with either caring for their current patients or spending hours reviewing medical charts to identify potential palliative candidates.

Avati and colleagues’ approach uses deep learning to screen patients upon admittance to the hospital, quickly identifying which individuals might benefit from future palliative care.

“This frees the palliative care team from manual chart review of every admission and helps counter the potential biases of treating physicians by providing an objective recommendation based on the patient’s EHR,” the authors wrote.

Avati and colleagues said their current model is being piloted for daily, proactive outreach to newly admitted hospital patients, and they plan to continue collecting objective outcome data on the algorithm’s rate of success.