Leveraging Predictive Analytics to Support Emergency Operations in NYC
Written by Jennah Gosciak
New York City Emergency Management (NYCEM) is the primary city agency responsible for coordinating citywide planning and response to major emergencies that impact New Yorkers. These include unexpected incidents, such as major transit strikes or building collapses, as well as more common seasonal ones, like coastal storms and severe winter weather events. While NYCEM uses numerous data-driven tools to prepare for and respond to disasters, the agency is still exploring whether and how predictive tools may support their work.
The Challenge
As a Rubinstein PiTech PhD Impact Fellow, I worked with NYCEM’s Office of Strategic Operations to improve preparedness and response efforts using machine learning. The project involved several parts:
Understanding how NYCEM units make decisions and identifying problem-areas where predictions may have the greatest impact.
Studying the different agency-developed tools involved in decision-making – indices such as the Heat Vulnerability Index, the Flood Vulnerability Index, or the Urban Risk Index – and conducting sensitivity analyses.
Identifying hazards that might be well-suited to a prediction-based framework and prototyping a solution.
Identifying the Problem
During the first few weeks of the fellowship, I met with several teams across the agency, e.g.the Community Engagement Bureau and the Critical Infrastructure Unit, to understand the kinds of challenges they face and begin to identify if a machine learning tool could be useful. These conversations revealed tools that NYCEM units regularly use to make decisions, such as the Heat Vulnerability Index (HVI) or the Urban Risk Index (URI), which I was then able to study . My goal was to understand how the tools might differ from the risk scores obtained from a machine learning model, and whether they are sensitive to changes in the inputs and the scale.
Figure 1. The interactive portal for the Heat Vulnerability Index on the Environment & Health Data Portal
Narrowing the Project
From our initial meetings, my NYCEM supervisor, Luke Boyce, and I identified three directions for the project:
Improving outreach to affected communities
Anticipating incidents and staffing needs
Improving the assessment of vulnerability and risk throughout the city
Additionally, we narrowed down the types of hazards that our work would address and focused specifically on extreme heat impacts. Extreme heat is a consistent and growing problem in New York City, contributing to an estimated 500+ deaths each summer. NYCEM prepares for heat season regularly by sending out alerts (e.g., the Beat the Heat campaign) via the City’s emergency public messaging program, NotifyNYC, and activating cooling centers.
Working towards a Solution
Figure 2. A prototype tool that visualizes heat risk scores obtained from a predictive model, as opposed to an index (Note: this tool uses fake data).
After narrowing down the scope of the project, I focused on building simple models to predict the impacts of extreme heat exposure and comparing them with static, index-based tools like the HVI. Defining labels for the models proved difficult, as measuring the impact of heat exposure is far from clear and existing data sources -- whether public or private – may be insufficient. Prior work on morbidity and mortality undercounts the true effect of heat exposure, while sensor data and weather models miss factors like access to air conditioning and unique characteristics of the built environment. With these limitations, I identified four proxy outcomes that could serve as potential labels: (1) 311 complaints for open hydrants, (2) 311 complaints for power outages, (3) Con Edison power outages, and (4) EMS calls for heat exhaustion and heat stroke.
Figure 3. A photo with my supervisor at NYCEM, Luke Boyce
Using these proxy outcomes, I evaluated the predictions from several simple models. Their risk scores diverged significantly from existing risk scores like the HVI, underscoring the limitations of relying exclusively on these tools for decision making. At the same time, the proxy outcomes are imperfect – highlighting the importance of collecting high-quality data on heat impacts in order to make well-evidenced decisions.
Impact and Path Forward
Jennah Gosciak
Ph.D. Student, Information Science, Cornell University
My initial results demonstrate that, while valuable for their intended purpose, existing tools have limitations, and thinking about preparedness efforts as a prediction problem may be beneficial. This fall, I will be continuing to work with NYCEM to refine the modeling process, explore more complex approaches (e.g., models that leverage the underlying spatial structure of the city), and search for more reliable proxy outcomes. I also will consider the equity implications of different models, particularly if used in high-stakes decisions to inform heat-related interventions. Based on feedback from NYCEM staff, I ultimately hope to develop a user-friendly tool that complements existing tools like the HVI and influences decision-making.