How finalist Vituity uses emergency room wait times to detect potential flu pandemics
This is the second post in our Q&A series featuring the five Hidden Signals Challenge finalists. These solutions demonstrate the exciting potential for open data to help identify biothreats in real time.
Our second Q&A features finalist Vituity. This team’s model, Monitoring emergency department wait times to detect emergent influenza pandemics, alerts authorities of spikes in emergency room wait times that can be attributed to emergent flu pandemics. The solution sources real-time data from a network of 203 emergency departments in 20 states and is updated hourly, allowing agencies to quickly intervene.
“We have seen firsthand how data products start with a simple question, but frequently morph over time into a platform (e.g., a spoon vs. a swiss army knife). Our goal in this submission was to keep our product simple and to alert public health authorities when we detect a spike in influenza-like illnesses.”
What inspired you to submit this concept?
Our team floated several ideas in the early stages, and they mostly pulled from our academic experiences. For example, Nate Sutton initially proposed that we could detect outbreaks of avian influenza by monitoring transpacific bird migrations with eBird, as he was familiar with these data from his previous research. In the end, we chose to develop a solution to estimate influenza-like illnesses based on wait times in emergency departments (EDs). This felt closer to Vituity’s core competency, as our partnership of physicians has been staffing emergency departments since 1971. It also allowed us to leverage our in-house database of historical ED wait times to build a machine learning model.
How will your concept enable city-level operators to make critical and proactive decisions?
The Centers for Disease Control has a robust monitoring program for influenza-like illnesses based on data streams from outpatient centers, mortality databases, and laboratory samples. Unfortunately, these streams all have a substantial lag in reporting that varies from one to three weeks. Our goal was to complement these existing systems where current information is missing with real-time data. This will make proactive interventions such as prophylaxis possible and timely— within hours, not days.
What sets your concept apart from existing solutions?
The closest analog to our solution was the ‘nowcasting’ of influenza-like illnesses with the Google Flu Trends project. This project has been archived after more recent analyses have shown the project was both overfitting and failing to accurately estimate the rate of influenza-like illnesses. We have avoided these and similar pitfalls and mistakes in our solution by basing our model upon more reliable data, as we believe that people leaving the comfort of their homes to seek emergency care is a more reliable signal of influenza-like illnesses than web searches or posts on social media.
How did you see your concept evolving during Stage 2 of the Challenge?
A large part of our effort in Stage 2 was focused on refactoring our ingestion framework to make it less brittle to underlying changes in hospital websites. Our Python modules are now functions that ingest the desired components of HTML structure, and they are coordinated from a configuration table in our PostgreSQL database. This is a more flexible approach that will have lower maintenance costs in the long run, and code changes won’t be required to bring hospital feeds back online.
We have seen firsthand how data products start with a simple question, but frequently morph over time into a platform (e.g., a spoon vs. a swiss army knife). Our goal in this submission was to keep our product simple and to alert public health authorities when we detect a spike in influenza-like illnesses. The simplest way to do this is to surface our model’s predictions as an application programming interface, and so we built a RESTful API. We expect relevant authorities to inspect any alert, and so we built a graphical user interface to surface our model’s predictions. This will allow the user to quickly go in and understand the spatial and temporal context of the current alert.
What is the biggest insight you’ve uncovered through the Challenge thus far?
We first hypothesized that we could leverage emergency department wait times as an indicator for influenza-like illnesses in late October 2017. At this time the activity of influenza-like illnesses was mild in the majority of the US, and low in a couple of southern states. By December the severity of this flu season was clear, and the issue of long wait times in emergency departments was brought to the forefront of national attention. The data we collect from emergency departments staffed by Vituity providers has confirmed this narrative.
Learn more about other finalists’ solutions on the Challenge blog and subscribe to our newsletter to continue receiving updates. The winner(s), who will receive up to $200,000 in cash prizes, will be announced later this spring.