Introduction
Traditional contact tracing is primarily "forward-facing": once an individual tests positive, health officials isolate them and their recent contacts to prevent future transmission. However, for respiratory diseases like COVID-19, influenza, or RSV, many carriers are asymptomatic or infectious before symptom onset. Manual tracing relies on fallible human memory, which often fails in cases of community spread where the "who infected whom" link is invisible.
Forward tracing alone frequently misses the source of an infection. If the source remains unidentified, the entire branch of the outbreak they initiated continues to spread unchecked. This research introduces an automated framework to discover these hidden transmission paths using real-world mobility data and machine learning, enabling "bidirectional" tracing that looks both backward to the source and forward to all potential secondary infections.
Methodology
The researchers developed a multi-step pipeline to transform raw mobility data into an actionable public health tool. The process begins with constructing person-to-person contact networks from mobility trajectories and simulating disease spread using an agent-based SEIR model. The core innovation is the Infectious Path Centrality (IPC) metric, which identifies nodes central to the paths connecting known infectious individuals. This metric is then used to train a Graph Neural Network (GNN) for edge classification, distinguishing true transmission events from casual interactions.
Data and Experimental Setup
The study utilizes the Foursquare Mobility dataset from May through August 2020 in Austin, Texas. The data captures device "dwell times" at various Points of Interest (POIs), resulting in networks ranging from 20,000 to 31,000 nodes and up to 244,000 edges. This dataset is particularly valuable as it reflects real human behavior during various stages of pandemic-related lockdowns and reopenings.
Key Insights: The Power of Bidirectional Tracing
A critical realization of this work is that bidirectional tracing is exponentially more effective than forward tracing because it identifies "unseen branches." By finding the "parent" of an infection, the system can locate all "sibling" infections that would otherwise go undetected.
Standard metrics like betweenness centrality often fail in this context because they overvalue every high-degree node. In contrast, IPC acts as a powerful filter, highlighting specific sub-graphs where transmission is most probable. This helps overcome the massive class imbalance in contact networks, where the vast majority of interactions do not result in infection.
Results
The framework demonstrates significant improvements in both detection and mitigation:
- High Classification Accuracy: The GNN edge classifier achieved an F1-score of 94%, proving that IPC is a highly predictive feature for identifying transmission events.
- Outbreak Suppression: In simulations where only 30% of symptomatic individuals were tested, forward tracing only reduced the reproduction rate ($R_t$) from 1.33 to 1.27. Bidirectional tracing, however, slashed $R_t$ to 0.54—a 71% reduction that effectively halts the outbreak.
- Optimal Depth: Ablation studies indicated that looking 2 hops back into contact history provides the optimal balance between graph coverage and classification accuracy.
Limitations and Future Work
While promising, the study relies on simulated health labels (SEIR) due to the lack of public ground-truth infection data linked to GPS pings. Future research will focus on the model's robustness against "noisy" data (such as false test results) and its scalability to larger, denser urban environments like New York City. Additionally, the real-world success of such a system depends heavily on public compliance with quarantine notifications.
Summary updated