Modeling and Inferring Latent Diffusion Networks for Active Surveillance and Prediction of Infectious Diseases

This project will aim to develop and empirically validate a novel computational framework for active surveillance in the control and prevention of infectious diseases.

The large ranges of infectious disease diffusion make comprehensive surveillance challenging or even impossible due to limited surveillance resources during outbreaks. Existing disease surveillance relies on case reporting from existing health agencies; however, this approach is somewhat passive, and can lead to unsatisfactory results because of error-prone observations.

Active surveillance is a crucial strategy for addressing the shortcomings of passive surveillance when resources are limited. It proactively determines key locations for surveillance by identifying and utilizing the latent diffusion network inferred from clinically confirmed historical data. Computationally speaking, the problem of inferring the latent diffusion network from historical data for key location selection is very challenging as during the disease diffusion process, the interrelationships between two locations are hidden; what we have in the historical data is only the number of confirmed infected cases at all locations in the past. Moreover, given the limited resources, the goal of diffusion network inference is not merely characterizing the structure of the entire network but identifying key locations from which current/future infected cases at all locations (including unobserved ones) can be predicted.

This project will develop a group sparse Bayesian learning (GSBL) framework to address the above problems, i.e., selecting key locations from the inferred latent diffusion network for active surveillance. Under the GSBL framework, we will tackle the following computational challenges:
1) How to model the spatially sparse distribution of infectious disease cases;
2) How to infer latent diffusion networks in a timely fashion;
3) How to characterize the complicated diffusion patterns of infectious diseases.
To address those challenges, we will correspondingly propose three novel methods:
1) GSBL with Bernoulli likelihood (GSBL-B);
2) Scalable GSBL (GSBL-S);
3) Nonlinear model embedding for GSBL (GSBL-N).

We will carry out both experimental evaluation and onsite validations on multiple scales. Specifically, we will consider two representative real-world disease diffusion scenarios: the influenza outbreak in Hong Kong and the malaria elimination initiative in China.

The outcomes of this project will contribute to academic research in computer science and epidemiology, and to the control and prevention of infectious diseases in practice. Results derived from this project will provide a scientific foundation for public health authorities in the development of policies dealing with infectious diseases