Predicting High Utilizing EMS Patients for Proactive Care Interventions

Background: One of Memphis Fire Department's key priorities is providing high quality, responsive emergency care to Memphians. The High Utilizer Group (HUG) is a program implemented by MFD emergency medical service (EMS) that provides outreach and assistance to high frequency 911 callers, who typically have complex health needs and limited access to non-emergency healthcare. Patients enrolled in HUG receive preemptive care or an alternate emergency medical response to better provide the right care at the right time. The program also aims to increase ambulance availability for urgent medical emergencies by helping patients get connected to more routine services. In 2019, MFD EMS partnered with Data Science for Social Good to build a predictive model to identify potential high utilizers of EMS in an effort to provide more proactive care interventions.
Service Question: Can we identify potential high utilizers of EMS resources and proactively match them with appropriate intervention to ensure limited EMS resources are reserved for emergent calls?

Analytics: EMS currently uses call volume to identify and redirect high utilizer patients to primary care based on how often an individual has called: more than 3 calls in the past 7 days; more than 6 calls in the past 30 days or more than 9 calls in the past 90 days.
EMS used incident level data from 2016 through 2019 including electronic patient care records, computer-aided dispatch records and community paramedicine records. The challenge was to identify individual patients across data sources using a unique identifier. 
Results: EMS created a model predicting the risk of a patient calling EMS more than 9 times in the next 90 days. The model, known as a binary classification model, generated a list of the top 20 individuals most likely to call. It was optimized to maximize the proportion of correctly classified individuals among the top 20 with the highest predicted risk. The fraction of all high utilizers during the test period was also calculated using different list lengths.  
A random forest was chosen based on better performance on small list lengths compared to logistic regression and gradient boosted trees. With this model, EMS created a list of the top 20 individuals making nine or more calls. 9-1-1 emergency calls from these patients represented over 20% of total H.U.G.s. The baseline model is trained on two key predictive features, number of calls ever made and number of calls over the past year. 
Impact: EMS successfully used predictive modelling to identify potential high utilizer patients. By using advanced data analytics, the model was able to identify patients with greater accuracy when compared to current approaches. The model can generate a targeted list of patients on a weekly basis and can be integrated in current workflows by EMS paramedics. Patients identified for early interventions can be prioritized for preventive care, reducing 9-1-1 emergency calls. This allows EMS paramedics to provide high quality services to more patients.