The Daily Pennsylvanian is a student-run nonprofit.

Please support us by disabling your ad blocker on our site.

The Perelman Center for Advanced Medicine and the Pavilion Hospital of the University of Pennsylvania on Feb. 24 Credit: Mehak Dhaliwal

A recent Penn Medicine study found that “reminders” triggered by artificial intelligence for clinicians improved the quality of end-of-life care for terminally ill cancer patients.

The study, published in January, aimed to provide more efficient and personalized palliative care and increase the frequency and efficiency of “serious illness conversations” so that patients could have more agency over their treatment. The AI-based algorithm found high-risk patients who would likely die within six months and sent emails to their clinicians to discuss end-of-life goals. 

This study is the largest to examine machine learning in palliative oncology, with over 20,000 patients and over 40,000 patient encounters. Results included a decrease in chemotherapy among patients who died, but the study did not impact other end-of-life metrics like hospice or ICU visits. 

The algorithm used approximately 150 electronic health record variables associated with prognosis to predict which patients had the highest likelihood of proximate death. This was done in a way that doctors cannot do manually and accurately, according to Ravi Parikh, the associate director of the Penn Center for Cancer Care Innovation.

“Our hypothesis was, by integrating AI to engender better prognoses, we can then target nudges to clinicians to encourage them to have the conversations that they know they ought to have, and that they want to have earlier than they're currently doing,” Parikh said. 

He went on to suggest that some of the program’s goals were targeting resources to patients who need them most and doing so in a timely manner. 

Parikh said that the algorithm helps integrate conversations as a standard practice of palliative care, framing them as an ongoing process. This helps make conversations shorter and more digestible for the patient and their family. 

Serious illness conversations were first outlined by a Harvard-developed conversation checklist. The conversations include topics such as family involvement in decision-making, medical interventions, and treatment priorities. Penn subsequently became its predominant user in national oncology settings, said Parikh. 

Pallavi Kumar, the director of Oncology Palliative Care, explained that these priorities look different for different patients. Kumar discussed how doctors can aid with deciding how and when to begin hospice, connecting with others in their final days and outlining their goals.

Kumar and her colleagues start a relationship with a patient by asking for information preferences, such as whether they would prefer data points or the big picture, and what degree of detail they would like. 

“I would love to [think] about how we can use these smart tools to lift some of the emotional weight off the shoulders of the clinicians and utilize other disciplines like nurses [and] social workers,” Kumar said. 

Parikh is excited to see how artificial intelligence can become better implemented and used for positive patient outcomes.

“[Implementing algorithms to change healthcare] is the last mile problem in clinical care,” Parikh said.