Abstract COVID-19 remains a global threat in the face of emerging SARS-CoV-2 variants and gaps in vaccine administration and availability, and organizations must be prepared to detect and mitigate its risk to their people and activities. In this report we share key lessons learned from an adaptive COVID-19 testing program implemented at a mid-sized university in the Midwest. The program utilized two simple, diverse, and easily interpretable machine learning models to quickly and accurately predict which students were at elevated risk for contracting COVID-19 and should be called proactively for testing. Our adaptive testing cohorts produced positivity rates that were 26% higher than the random cohort: 0.58% positivity (95% CI 0.47% to 0.68%) from 19,171 tests, and 0.46% positivity (95% CI 0.41% to 0.51%) from 64,003 tests, respectively. Within 14 days of their selection, 2.94% of the adaptive cohort tested positive, compared to 1.27% of the random cohort. Close contacts who were predicted by the adaptive testing models received a COVID-19 test within an average of 0.94 days (95% CI 0.78 to 1.11) of the source testing positive, while those who were manually contact traced were tested in an average of 1.92 days (95% CI 1.81 to 2.02). These results suggest that machine learning strategies can improve surveillance testing effectiveness, especially in a university setting, by effectively distributing testing resources and potentially reducing community transmission.
The Efficiency Problem in Surveillance Testing
Surveillance testing is a cornerstone of pandemic management, but it faces a fundamental efficiency problem. Randomly sampling a population (the "general surveillance" approach) is egalitarian but often misses the specific clusters where the virus is actively spreading. Conversely, manual contact tracing is precise but slow, labor-intensive, and limited by the memory and honesty of the individuals involved.
Researchers at the University of Notre Dame sought to bridge this gap by using "adaptive testing." The goal was to use existing institutional data—like class rosters and housing assignments—to predict who was likely to be infected before they showed symptoms or were officially named as a close contact.
Methodology: Modeling the Campus as a Social Network
The researchers modeled the university as a heterogeneous network, where students are nodes and their interactions (roommates, dorm-mates, classmates, teammates) are edges. They deployed two distinct, complementary machine learning models to identify high-risk individuals:
- Node Risk Prediction (NR): This model used a "risk propagation" approach. It assigned a base risk score to students based on symptoms or recent positive tests and then allowed that risk to "flow" through the network to neighbors up to two hops away.
- Multi-relational Link Prediction (LP): Instead of looking at individual risk, this model predicted the likelihood of a "contact tracing relationship" existing between two students. It was trained on historical data to learn which network connections were the strongest proxies for viral transmission.
Students with the highest risk scores from either model were invited for "adaptive" testing. The study utilized a rich dataset from the 2020-2021 academic year, including dormitory addresses, course enrollments, and results from over 145,000 COVID-19 tests.
Results: Enhanced Detection and the Incubation Gap
The adaptive testing program significantly outperformed traditional random surveillance. The adaptive cohort demonstrated a 0.58% positivity rate, compared to 0.46% for the random cohort—a 26% improvement in detection efficiency.
A crucial observation was the "incubation gap." While the adaptive cohort was more likely to test positive on day zero, the gap between the two groups widened significantly over time. This highlights the "pre-infectious" window where models can identify risk before the virus is even detectable by a PCR test.
By day 14, 2.94% of the adaptive cohort had tested positive, compared to just 1.27% of the random cohort. This suggests the models were successfully identifying individuals in the early stages of infection or those with high exposure risk.
Results: Accelerating Contact Tracing
Speed is the most critical factor in breaking transmission chains. The adaptive models identified close contacts who received tests in an average of 0.94 days after the source case tested positive. In contrast, manual contact tracing took 1.92 days on average.
Saving nearly 24 hours in the transmission chain is critical for stopping outbreaks in high-density environments. As shown in the figure above, individuals flagged by the adaptive model reached near-total testing coverage much faster than those relying solely on traditional methods.
Discussion and Limitations
Beyond the raw data, the researchers noted a "behavioral nudge" effect: the adaptive cohort had a 95% response rate compared to 78% for the general cohort. Being informed that they were selected via a data-driven risk model may have encouraged students to take the testing request more seriously.
However, the approach has limitations. The models rely on institutional data and cannot account for "invisible" networks like off-campus parties or informal social gatherings. Furthermore, self-reported symptom data proved too inconsistent to be a reliable predictor on its own. The authors emphasize that machine learning is a supplement to, not a replacement for, manual contact tracing, which provides necessary nuance and detail.
Conclusion
This real-world application demonstrates that machine learning can significantly improve the effectiveness of public health surveillance. By identifying high-risk individuals faster and more accurately than random sampling, institutions can better distribute limited testing resources and reduce the window for community transmission.
Summary updated