Imagine this scenario, one that’s not uncommon for people diagnosed with incurable cancer: You and your cancer doctor decide that you should try chemotherapy to prolong your life. Six months later, that chemotherapy and several other treatments not only haven’t slowed the cancer but have caused burdensome side effects — some so bad you needed to be hospitalized.
Finally, at this point, the doctor asks, “What matters most if you were at the end of your life?”
Would having that conversation about your end-of-life wishes earlier have helped you make the most of your last days? And if your doctor had believed you may not survive past six months, would you have wanted to know?
When conversations about goals and end-of-life wishes happen early, they can improve patients’ quality of life and decrease their chances of dying on a ventilator or in an intensive care unit. Yet doctors treating cancer focus so much of their attention on treating the disease that these conversations tend to get put off until it’s too late. This leads to costly and often unwanted care for the patient.
This can be fixed, but it requires addressing two key challenges. The first is that it is often difficult for doctors to know how long patients have left to live. Even among patients in hospice care, doctors get it wrong nearly 70% of the time. Hospitals and private companies have invested millions of dollars to try and identify these outcomes, often using artificial intelligence and machine learning, although most of these algorithms have not been vetted in real-world settings.
In a recent set of studies, our team used data from real-time electronic medical records to develop a machine learning algorithm that identified which cancer patients had a high risk of dying in the next six months. We then tested the algorithm on 25,000 patients who were seen at our health system’s cancer practices and found it performed better than relying only on doctors to identify high-risk patients.
But just because such a tool exists doesn’t mean doctors will use it to prompt more conversations. The second challenge — which is even harder to overcome — is using machine learning to motivate clinicians to have difficult conversations with patients about the end of life.
We wondered if implementing a timely “nudge” that doctors received before seeing their high-risk patients could help them start the conversation.
To test this idea, we used our prediction tool in a clinical trial involving nine cancer practices. Doctors in the nudge group received a weekly report on how many end-of-life conversations they had compared to their peers, along with a list of patients they were scheduled to see the following week who the algorithm deemed at high-risk of dying in the next six months. They could review the list and uncheck any patients they thought were not appropriate for end-of-life conversations. For the patients who remained checked, doctors received a text message on the day of the appointment reminding them to discuss the patient’s goals at the end of life. Doctors in the control group did not receive the email or text message intervention.
As we reported in JAMA Oncology, 15% of doctors who received the nudge text had end-of-life conversations with their patients, compared to just 4% of the control doctors.
Interestingly, doctors who received text messages doubled their rates of conversations even with patients who had a lower risk of dying. This suggests that the combination of the algorithm and the nudge may have changed normal practice patterns, spurring doctors to have these important conversations with patients sooner, and possibly forming new habits.
There are several reasons why this might have worked. It’s hard for doctors to tell which patients may die sooner than others. One study found that doctors’ prognoses for patients with advanced cancer were close to their actual survival only about 20% of the time. With the growing amount of data collected by electronic health records, machine learning algorithms can help doctors identify and prioritize patients who might benefit the most from an end-of-life conversation.
The bigger reason is that these are difficult conversations, and it is much easier to put them off. Getting a nudge can help doctors initiate it earlier with their patients.
There has been a lot of hype about the potential for artificial intelligence and machine learning to improve health care. So far, there’s been limited evidence that these approaches alone are successful. But combining machine learning with a simple nudge could make the difference between a missed opportunity and a patient who gets to spend her final days at home with her family.
Ravi Parikh is an oncologist at the Crescenz VA Medical Center in Philadelphia and an assistant professor of medicine and medical ethics and health policy at the University of Pennsylvania Perelman School of Medicine. Christopher Manz an oncologist at the Dana-Farber Cancer Institute in Boston and an instructor in medical oncology at Harvard Medical School. Mitesh Patel is a general medicine physician at the Crescenz VA Medical Center, director of the Penn Medicine Nudge Unit, and associate professor of medicine and health care management at the Perelman School of Medicine and The Wharton School.