Together with the Vanderbilt Initiative for Smart Cities Operation and Research, the Nashville Fire Department and the city’s IT agency created a tool that uses predictive modeling to forecast emergency response times.
MetroLab Network has partnered with Government Technology to bring its readers a segment called the MetroLab Innovation of the Month Series, which highlights impactful tech, data, and innovation projects underway between cities and universities. If you’d like to learn more or contact the project leads, please contact MetroLab at email@example.com for more information.
In this month’s installment of the Innovation of the Month series, we explore how the Vanderbilt Initiative for Smart Cities Operation and Research has been working on emergency response with the Nashville Fire Department and the Information Technology Services Department for the Metropolitan Government of Nashville and Davidson County.
MetroLab’s Ben Levine and Stefania Di Mauro-Nava spoke with Abhishek Dubey, senior research scientist at the Institute for Software Integrated Systems and an assistant professor in the Electrical Engineering and Computer Science department at Vanderbilt University; Geoffrey Pettet, graduate research assistant at Vanderbilt University; Colleen Herndon, project manager at the Metro Information Technology Services Department for the Metropolitan Government of Nashville and Davidson County; and Ayan Mukhopadhyay, graduate research assistant at Vanderbilt University to learn more.
Stefania Di Mauro-Nava: Could you please describe what the Integrated Safety Incident Forecasting and Analysis project is? Who is involved in this effort?
Abhishek Dubey: The Integrated Safety Incident Forecasting and Analysis project is a collaborative project undertaken by the Vanderbilt Institute for Smart Cities Operation and Research (VISOR) team, the Nashville Fire Department (NFD), and the Metro Nashville and Davidson County Information Technology Services (ITS) Department as part of a National Science Foundation grant. The objective of this research is to understand and improve the resource coordination and dispatch mechanisms used by first responders in Nashville. The analysis and tools created in this project seek to improve NFD’s ability to respond quickly and effectively to emergencies such as traffic collisions with injuries to humans, medical emergencies, and fire incidents that occur in the Nashville area.
The primary idea behind this work is to move toward a proactive, rather than a reactive, emergency response strategy. For this work, we have used historical data collected by NFD over the last three years and more, recording the location, time and type of various safety incidents that include motor vehicle incidents, fire incidents and health-emergency related incidents along with the response times for the dispatched emergency vehicles. Additional information in the data includes the severity of the incident, the location of the incident, and the response time for the incident. We integrated this data with other crucial parameters that affect incident occurrence, like pedestrian traffic, road characteristics, traffic congestion and weather, and developed models for different parts of the city. These models predict with high accuracy the likelihood of when and where incidents will happen in the future.
Geoffrey Pettet: Given the incident forecasting tool and information about the past response times, we can analyze the effect of creating new fire stations, relocating fire stations or adding new response vehicles to existing stations (as shown in figures 1 and 2). It also allows us to provide city planners with historical incident data occurring at different time periods for analysis (shown in figures 3 and 4).
Additionally, we have developed a novel online framework to suggest optimal automatic dispatch decisions for emergency responders by integrating predictive models of incident prediction and the city environment, including events and changing traffic conditions. This framework includes an online survival model for incident prediction and a recurrent neural network model for learning environmental features affecting dispatch. These are used by a Monte Carlo-based dispatching algorithm that simulates different responses to the predicted incidents and chooses the best current dispatching action (figure 5). Further, the online framework provides the capability to pre-emptively optimize the placement of responders and perform forward-thinking dispatching decisions in real time.
At Vanderbilt University, Abhishek, along with Gautam Biswas and Yevgeniy Vorobeychik (now at Washington University in St. Louis) led a team including Ayan Mukhopadhyay, Chinmaya Samal, and me. Colleen Herndon with Metro Information Technology Services coordinated the project from Metro’s side, with collaboration from Rusty Lacy and Jeanne Martin with the Nashville Fire Department.
Figure 1: Planners can add, move, or remove fire stations to see their effect. Courtesy of VISOR group.
Ben Levine: Can you describe what motivated the city and university to address this particular challenge?
Pettet: Responding to incidents quickly and effectively is a critical and costly operation, so NFD has spent considerable effort in analyzing their placement of responder stations and response patterns. They found that there are a few areas that could be improved. For example, they discovered an issue called the vortex problem: Many incidents occur in downtown Nashville, which is also where most of the hospitals in the area are located. Ambulances from outlying regions will take their patients to the hospitals, and on their way back will be dispatched to accidents in the downtown area.
Colleen Herndon: This causes ambulances to spend the majority of their time in the downtown core, reducing their time at the more outlying stations, and leading to an increased response time for incidents occurring in those areas. This is just an example of the type of problem that a principled, proactive dispatching approach can help address. NFD was motivated to pursue this project not only for the reasons mentioned, but also because the results of this analysis can provide additional justification for requests to increase and/or reallocate the department’s fleet of first responding vehicles.
Figure 2: Given a station arrangement defined by planners as in Figure 1, the system returns the estimated average response time change given expected incident distributions. Courtesy of VISOR group.
Levine: What kind of data are you exploring and why? What have been some of your initial findings and is this changing how you view the issue?
Ayan Mukhopadhyay: We are integrating data from different sources to explore this problem. The main source is the Fire Department database, which has detailed records of incidents that they have responded to. These include the incident type, where and when the incident occurred, which responder was dispatched to it, and how long it took for the responder to arrive on scene. After cleaning the data, this amounts to approximately 20,000 traffic incidents in the Nashville area over the course of two years. We combined this with historic weather data from Dark Sky and traffic congestion information from HERE to build an incident prediction model.
One interesting finding is that weather and congestion have a lower impact than we originally thought. The primary feature affecting traffic incident occurrence is previous incidents in the area. In other words, incidents cascade along roadways, causing more incidents in the future. This implication not only affects how we construct our incident models, but also further justifies the problem we are tackling: responding to incidents quickly not only ensures increased chances of saving lives, but possibly reduces secondary factors like congestion. This analysis has value beyond the scope of this specific analysis. It also provides insights for Metro and state roadway planners and engineers who can similarly analyze our results and incorporate improvements in areas identified as problematic, as well as use the information for planning new routes.
Figure 3: The dashboard allows planners to see the distribution of past incidents at different times. Courtesy of VISOR group.
Di Mauro-Nava: How are your findings being used and implemented at the Nashville Fire Department?
Herndon: The incident prediction models have been integrated into a dashboard that the Fire Department is currently testing. This dashboard is available on GitHub and YouTube.
This tool will allow management to see what the likely incident distributions are over the county for different time periods, and plan accordingly. A feature that NFD management has expressed particular interest in is the ambulance station exploration tool. This lets users move, remove and add stations and vehicles around the county, and gives them detailed information on how those changes will affect incident response times. This insight can be applied when planning and justifying the need for new stations and new vehicles, taking what traditionally had been knowledge gained through many years of field experience to knowledge gained from the use of this tool.
We are currently still in the experimentation and exploration phase of the dispatching tools, but once they are complete, management can see how the dispatching recommendations compare to their current strategies, and possibly integrate them to improve response times once they are proven.
Figure 4: Planners can also examine the distributions of the past incidents’ severity and types. Courtesy of VISOR group.
Levine: What was the most surprising thing you learned during this process?
Herndon: Metro has collaborated with Vanderbilt on multiple projects in recent years, including Connected Nashville: A Vision for a Smarter City. We feel that the partnership has been and will continue to be extremely successful. The one thing that stood out in this particular project is just how much pragmatic application this project’s analysis could provide. This project has not only served as a proof of concept, but is now poised to give NFD a tool that can be incorporated into their planning processes and, over time, into their daily operations.
Figure 5: Monte-Carlo Tree Search (MCTS) forms the foundation of the dispatching framework. Courtesy of VISOR group.
Levine: Where will this project go from here?
Dubey: Vanderbilt will continue to work closely with NFD, to gather feedback and provide refinements within the tools to maximize the value for the department. We will also continue to incorporate real-time traffic congestion into the models. We are optimistic that the success of this project could lead to even deeper analytical projects with NFD. Our incident models can also be shared with other departments, such as Planning and Public Works, to help identify the root causes of the incidents and help prevent them in the first place.
Looking for the latest gov tech news as it happens? Subscribe to GT newsletters.