Continuous learning and inference of individual probability of SARS-CoV-2 infection based on interaction data

Shangching Liu, Koyun Liu, Hwaihai Chiang, Jianwei Zhang, Tsung-Yao Chang

Abstract

This study presents a new approach to determine the likelihood of asymptomatic carriers of the SARS-CoV-2 virus by using interaction-based continuous learning and inference of individual probability (CLIIP) for contagious ranking. This approach is developed based on an individual directed graph (IDG), using multi-layer bidirectional path tracking and inference searching. The IDG is determined by the appearance timeline and spatial data that can adapt over time. Additionally, the approach takes into consideration the incubation period and several features that represent real-world circumstances, such as the number of asymptomatic carriers present. The CLIIP approach is validated using an individualized bidirectional SEIR model to simulate the contagion process. Compared to traditional contact tracing methods, our approach significantly reduces the screening and quarantine required to search for potential asymptomatic virus carriers by as much as 94%.

Introduction: Beyond Binary Contact Tracing

Traditional contact tracing is a reactive, binary process: an individual either had contact with a confirmed case or they did not. This approach struggles with two major hurdles. First, asymptomatic spread allows many carriers to remain "invisible" to traditional systems until they have already infected others. Second, resource scarcity means that testing and quarantine resources are finite; exhaustive tracing often leads to over-quarantining healthy people or missing "superspreaders" who move frequently through high-traffic areas.

The authors aim to move from binary "contact tracing" to probabilistic "contagious ranking," allowing health officials to prioritize the most likely carriers for testing and intervention.

The CLIIP Framework

The core of the research is the CLIIP (Continuous Learning and Inference of Individual Probability) framework. As illustrated in the system architecture, the model continuously updates individual risk scores based on evolving spatial and infection data.

The CLIIP framework flowchart illustrating the integration of spatial data, infection states, and machine learning.

The framework relies on three main technical pillars:

  • Individual Directed Graph (IDG): The system builds a graph where nodes are individuals and edges represent interactions.
  • Bidirectional Tracking: When a new case is confirmed, the model performs backward tracking (searching the incubation period to find the likely source) and forward tracking (identifying who that person might have infected).
  • Machine Learning (LightGBM): Instead of simple proximity rules, the authors use a LightGBM (Gradient Boosting Machine) model to extract features from the IDG—such as contact duration and the infection status of "center-surround" nodes—to calculate a continuous probability of infection.

Modeling Interactions with Directed Graphs

A key innovation of CLIIP is the use of directed edges in the interaction graph. The direction points toward the person who left a location later, based on the insight that the virus can persist in the environment; the person arriving later is at higher risk of being infected by the person who was there earlier.

The Individual Directed Graph (IDG) models interactions over time, tracking potential transmission paths across different time frames.

By making the interaction graph directed based on "staying time," the model accounts for indirect transmission at waypoints (like an elevator button or a shared desk), which traditional "proximity-only" tracing often misses. To handle the massive computational load of city-scale data, the system employs Monte Carlo Tree Search (MCTS) to quickly find important features within the IDG without exhaustive searching.

Simulation and Performance

The study validates the model using a combination of real-world city GPS data and simulated epidemiological states. Since "ground truth" for every asymptomatic carrier in a real city is impossible to obtain, the authors developed an enhanced SEIR (Susceptible-Exposed-Infectious-Recovered) simulation. This simulation "labels" the GPS data, creating a synthetic but realistic environment where the true infection status of every individual is known for validation purposes.

Key results from the simulation include:

  • High Accuracy: The CLIIP model achieved an average Area Under the Curve (AUC) of 0.96, indicating extremely high precision in distinguishing between infected and healthy individuals.
  • Massive Resource Savings: The ranking method allowed the system to capture the same number of infected individuals while testing only ~6% of the population that traditional tracing would require. This represents a 94% reduction in screening and quarantine resources.
  • Feature Importance: The most critical feature for predicting infection was contact time at high-risk locations, followed by the density of infected/exposed people in an individual's immediate graph neighborhood.

Conclusion

By providing a probability-based ranking rather than a simple binary result, CLIIP enables "precision quarantine." This approach allows for the identification of asymptomatic "bridges" (superspreaders) who are actively moving and infecting multiple clusters. While the model requires high-resolution spatial data, it offers a powerful tool for optimizing public health responses in the face of limited resources. Future work could incorporate additional factors like mask-wearing habits and local weather conditions to further refine the probability scores.

Summary updated

Abstract

This study presents a new approach to determine the likelihood of asymptomatic carriers of the SARS-CoV-2 virus by using interaction-based continuous learning and inference of individual probability (CLIIP) for contagious ranking. This approach is developed based on an individual directed graph (IDG), using multi-layer bidirectional path tracking and inference searching. The IDG is determined by the appearance timeline and spatial data that can adapt over time. Additionally, the approach takes into consideration the incubation period and several features that can represent real-world circumstances, such as the number of asymptomatic carriers present. After each update of confirmed cases, the model collects the interaction features and infers the individual person's probability of getting infected using the status of the surrounding people. The CLIIP approach is validated using the individualized bidirectional SEIR model to simulate the contagion process. Compared to traditional contact tracing methods, our approach significantly reduces the screening and quarantine required to search for the potential asymptomatic virus carriers by as much as 94%.
BibTeX
@inproceedings{liu2021continuous,
  author = {Shangching Liu and Koyun Liu and Hwaihai Chiang and Jianwei Zhang and Tsung-Yao Chang},
  title = {{Continuous learning and inference of individual probability of SARS-CoV-2 infection based on interaction data}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-81809-0},
}
Figure