Reliable and accurate forecast of epidemics in real-time has been critical in helping public health agencies and the society to react and respond to emergent infectious disease outbreak. In my thesis, I utilize the concepts, analytical tools and methodologies from physics, mathematics and statistics, along with computational tools to model and forecast epidemic evolution.

In my first project, we studied a mechanistic framework that simulates the infectious disease spreading as a reaction-diffusion process on data-driven multi-scale human mobility networks. We investigated the heterogeneity of contact patterns across different age groups in different countries and states using micro-demographic data, and demonstrated that it has significant impact on disease spreading using computational simulations. My second project focuses on developing a statistical method that extracts human behavior patterns from novel digital data source to infer the influenza circulation, so as to improve the forecast. We developed a now-casting and forecasting framework that provides estimates for influenza incidence in Italy up to 4 weeks ahead. The results showed quantitatively the value of incorporating retail market data in forecasting models, acting as a proxy that can be used for the real-time analysis of epidemics. My third project introduces a new forecasting framework that combines mechanistic and statistical methods. The data-driven, stochastic, spatially structured mechanistic model dynamically fuses highly accurate machine learning now-casting results as augmented ground truth of surveillance data, so as to enhance the model selection process and improve the predictions. We performed forecasts exercise from 2013/14 to 2018/19 influenza season in the United States. The forecast results in all seasons strongly proved that augmenting ILI activities only one week ahead through machine learning nowcasting results can significantly improve the forecasting performance for both short-term time series predictions and long-term forecasts of season peaks. Furthermore, the forecasting framework does not lose the capability to explicitly estimate key epidemiological parameters relevant for public health decision-making that cannot be captured by statistical models or machine learning models. My last project focuses on the real-time forecasting for the COVID-19 in the United States using the mechanistic framework developed in the first project and the model selection techniques in the third project. We integrated the transmission dynamics, the human mobility mitigation, the vaccination coverage, and the variants of the virus specified to the disease into the mechanistic model to project the outbreak evolution. We also enhanced the model selection process to overcome the reporting issues of the data. During the second wave of COVID-19 in the US, the model provided early warnings of the turning points of the epidemic roughly half month in advance. We generated weekly point predictions as well as confidence intervals that are able to quantify the uncertainty of both the epidemiological settings in the model and the reporting noise of the data.

Dissertation Committee

Alessandro Vespignani (chair)
Sternberg Family Distinguished Professor, Network Science Institute at Northeastern University

László Barabási
Robert Gray Dodge Professor of Network Science, Network Science Institute at Northeastern University

Dmitri Krioukov
Associate Professor,Network Science Institute at Northeastern University

Nicola Perra
Associate Professor in Network Science, Business School of Greenwich University

‍