Using nudges with machine learning algorithms to increase serious illness conversations


Patients with cancer often receive care that is discordant with their wishes, resulting in inappropriate and costly utilization. Evidence demonstrates that having serious illness conversations (SICs) early on can help clinicians align care with patient goals. However, many patients with advanced cancer die without ever having a SIC.  


Our goal is to test whether nudges can be combined with a machine learning algorithm that predicts the risk of 6-month mortality to increase SICs among clinicians and cancer patients. Each week clinicians will be provided with a list of patients they will see the following week who have a high risk of mortality. Using an opt-out approach, they will be asked to pre-commit to having these conversations and sent a text message reminder on the day of the visit. To assess the impact of these interventions while implementing this throughout Penn Medicine, we have designed a stepped-wedge cluster randomized controlled trial.


This intervention has been implemented across all practices at the Penn Abramson Cancer Center. We found the combination of machine learning mortality estimates with behavioral nudges led to a >4-fold increase in serious illness conversations (intervention, 15.2% vs. control, 3.6%; P<.001). It also increased the overall rate of serious illness conversations among all patients (intervention, 3.6% vs. control, 1.3%; P<.001). This is one of the first clinical trials to demonstrate the potential of combining machine learning with behavioral nudges and indicates the approach could be applied to other areas of medicine.


Penn Medicine Abramson Cancer Center


Penn Center for Precision Medicine