# COVID-19 hotspot detection in a university setting

## TL;DR

This paper presents a data-driven approach to detect potential COVID-19 hotspots in universities using student testing data and statistical modeling.

## Contribution

The novel contribution is a simulation-based method to identify hotspots by comparing observed infection rates to expected rates under no in-class transmission.

## Key findings

- An XGBoost model was used to estimate individual student positivity probabilities based on demographic data.
- Simulation experiments showed the method can detect groups with higher-than-expected infection rates.
- The approach was demonstrated using anonymized data from a university's Fall 2020 semester.

## Abstract

The onset of the COVID-19 pandemic commenced an era of widespread disruptions in the academic world, including shut downs, periodic shifts to online learning, and disengagement from students. In an effort to transition back to in-person learning, many universities and schools tried to implement policy that balanced student learning with community health. While academic administrators have little control over some aspects of COVID-19 spread, they often choose to use temporary shutdowns of in-person teaching based on perceived hotspots of COVID-19. Specifically, if administrators have substantial evidence of within-group transmission for a class or other academic unit (a “hotspot”), the activities of that class or division of the university might be temporarily moved online. In this article, we describe an approach used to make these types of decisions. Using demographic information and weekly COVID-19 testing outcomes for university students, we use an XGBoost model that produces an estimated probability of testing positive for each student. We discuss variables engineered from the demographic information that increased model fit. As part of our approach, we simulate semesters under the null hypothesis of no in-class transmission, and compare the distribution of simulated outcomes to the observed group positivity rates to find an initial p-value for each group (e.g., section, housing area, or major). Using a simulation-based modification of a standard false discovery rate procedure, we identify possible hot spots—classes or groups whose COVID-19 rates exceed the levels expected for the demographic mix of students in each group of interest. We use simulation experiments and an anonymized example from Fall 2020 to illustrate the performance of our approach. While our example is based on hotspot detection in a university setting, the approach can be used for monitoring the spread of infectious disease within any interconnected organization or population.