Introduction
During the COVID-19 pandemic, institutions like universities faced a critical challenge: identifying infected individuals before they could drive community transmission. Traditional methods, such as random surveillance testing, are resource-intensive and often miss asymptomatic cases. Conversely, manual contact tracing is frequently hampered by delays and the limitations of human memory. Researchers at the University of Notre Dame addressed this "needle in a haystack" problem by developing an "adaptive" testing program that leverages social network analysis and machine learning to proactively identify high-risk individuals.
Methodology
The researchers modeled the university population as a heterogeneous social network. In this model, students are nodes and edges represent administrative connections, including shared dorm rooms, floors, classes, and athletic teams. To predict risk, the team deployed an ensemble of two machine learning models:
- Node Risk (NR) Model: This model uses a risk propagation approach. When a student tests positive, their risk score is transmitted to neighbors in the social network. The weights for different connection types (e.g., roommate vs. classmate) were optimized using an evolutionary strategy.
- Link Prediction (LP) Model: This model estimates the probability of unobserved contact tracing relationships. By analyzing historical data and metrics like Katz centrality, it identifies likely close contacts that manual tracing might overlook.
These models were trained on a massive dataset from the 2020–2021 academic year, including over 145,000 test results and comprehensive administrative records.
Results
The adaptive testing program significantly outperformed random surveillance. The adaptive cohort demonstrated a positivity rate of 0.53%, compared to 0.37% in the general surveillance group. This advantage became even more pronounced over time; within a five-day window, the adaptive program's detection rate rose to 1.49%, a 122% increase over the 0.67% found via random testing.
Speed of intervention was another major benefit. The models identified close contacts in an average of 0.94 days from exposure, effectively halving the 1.92-day average required for manual contact tracing. This rapid identification allowed for earlier testing and isolation of potentially infectious individuals.
Discussion and Insights
A key finding was the effectiveness of administrative data as a proxy for viral transmission paths. Even without direct biological or location data, class schedules and housing assignments provided a robust framework for predicting risk. Furthermore, the program benefited from a "nudge" effect: students flagged as high-risk had a 95% testing response rate, significantly higher than the 78% seen in general surveillance.
While the study was conducted in a specific university setting with a relatively homogeneous demographic, the results suggest that these models could be adapted for other infectious diseases or corporate environments using Enterprise Social Networks (ESNs). Future work will focus on accounting for inter-community transmission and refining the social graph to capture interactions not documented in administrative records.
Summary updated