Using Machine Learning to Improve Emergency Medical Dispatch Decisions
Karen Lavi (Friedrich Miescher Institute, Part of Novartis Research Foundation);Ritvik Kharkar (UCLA);Mathew Kiang (Harvard Public Health Schoool);Christoph Hartmann (Boston Consulting Group);Paul Van Der Boor (McKinsey & Company);Adolfo De Unanue (Instituto Tecnologico Autonomo de Mexico);Leigh Tami (Office of Performance & Data Analytics, City of Cincinnati);Anson Turley (Cincinnati Fire Department);Cedric Robinson (Cincinnati Fire Department);Brandon Crowley (Office of Performance & Data Analytics, City of Cincinnati);Eric Potash (Center for Data Science & Public Policy, University of Chicago);Rayid Ghani (Center for Data Science & Public Policy, University of Chicago)
Emergency medical services (EMS) provide out-of-hospital acute medical care and transport to definitive care for those in need. For many medical incidents, minimizing out-of-hospital time is crucial and directly linked to patients’ chance of survival. Therefore, ideally, a transport unit should be sent to every medical incident; But in reality resources are limited. Thus, it is immensely important that medical transport dispatches are as accurate as possible—- sending medical transport units as quickly as possible when necessary and not sending them unnecessarily.
In this paper, we describe our work in partnership with the City of Cincinnati in building a live dispatch system that predicts which incidents will result in hospital transport and require medical transport dispatch. In addition to using historical data on past incidents, our system incorporates weather, temporal, and spatial data. Compared to the current approach being used in Cincinnati, and while using the same available resources, we find that a prediction model uses this enriched data increased dispatch accuracy by 25%—getting faster to ~3,000 patients that need hospital transport. Based on these results, this live predictive model is now being tested for deployment in the Fire Department of Cincinnati.