Deep learning helps identify palliative needs of hospitalized patients

Register now

Stanford University researchers have developed a pilot program using artificial intelligence to identify hospitalized patients with a high risk of death in the next three to 12 months by leveraging electronic health record data.

Using a deep neural network, the EHR data of patients at Stanford are automatically evaluated by an algorithm, the predictions of which enable the palliative care team to take a proactive approach in reaching out to such patients, instead of waiting for referrals from treating physicians or conducting time-consuming chart reviews.

The algorithm is trained using diagnostic codes, encounter and demographic information, as well as medication and procedure codes, which generates a report that highlights the most critical factors in the patient’s EHR data that explain the high-probability decision.

“We demonstrate that routinely collected EHR data can be used to create a system that prioritizes patients for follow up for palliative care,” concludes a recently published paper from the Stanford researchers. “In our preliminary analysis we find that it is possible to create a model for all-cause mortality prediction and use that outcome as a proxy for the need of a palliative care consultation.”

Also See: Online tool puts cancer patients at the center of their care

Ken Jung, a research scientist in the Stanford Center for Biomedical Informatics Research, points out that 80 percent of Americans want to die at home, but just 20 percent actually do. In fact, he says more than 60 percent of deaths in the U.S. occur in an acute care hospital, with most patients subjected to increasingly aggressive interventions in their final days.

“Events can take on a momentum of their own,” adds Jung, who notes that “oftentimes, doctors are a little overly optimistic about the prognosis of their own patients” and that referrals from treating physicians to the palliative care team do not happen when they should—causing patients to “fall through the cracks” without documenting their goals of care to maintain a high-quality life, as much as possible.

By predicting the probability of death in the next three to 12 months, Jung contends that Stanford’s deep learning model is able to automatically identify patients with palliative needs early enough in the process, bringing them to the attention of the palliative care team so that patients’ wishes for end-of-life goals are in alignment with the actual services they receive.

“The first-line care teams aren’t always recognizing the need or acting on them in a timely manner—and, the palliative care team are also busy doctors that don’t have the time to manually review every patient to evaluate them for this intervention,” observes Jung.

At the same time, he emphasizes that the deep learning algorithm isn’t making any decisions about what care patients do or do not receive.

“What it’s doing is flagging patients to receive intensive review by doctors who can then have conversations with the first-line care team, and then—if all the clinicians involved agree that it’s appropriate—they will go ahead and initiate contact with the patients and their families,” Jung concludes.

Ultimately, he says the goal is to identify patients who will benefit from this kind of palliative care consultation.

For reprint and licensing requests for this article, click here.