Artificial Intelligence Solutions for Reliable Epidemic Forecasting

Infectious diseases, such as seasonal influenza, Zika, Ebola, and the ongoing COVID-19, can be spread, directly or indirectly, from one person to another leading to an outbreak, an epidemic, or a pandemic. Infectious diseases place a heavy social and economic burden on our society. Producing timely, well-informed, and reliable spatiotemporal forecasts of the epidemic dynamics can help inform policymakers on how to provision limited healthcare resources, develop effective interventions, rapidly control outbreaks, and ensure the safety of the general public.

Traditional approaches are mainly based on theory-based mechanistic models (e.g., an agent-based SEIR model) and statistical time series models (e.g., autoregressive models). Recent advances in deep learning have significantly improved the state of the art in computer vision, natural language processing, and many other fields. Although deep learning-based predictive models have gained increased prominence in epidemic forecasting, they are far from being well explored. One challenge is the lack of sufficient good-quality training data, particularly during new emerging epidemics. Another challenge is that existing models are seldom designed to consider both spatial and temporal correlations dynamically for capturing disease spread dynamics. A further challenge is that such models rarely consider epidemiological context as prior. Models in the aforementioned cases are prone to be overfitting and are unlikely to provide explanatory power for the underlying phenomena due to the black box nature. Given the challenges, my research focuses on deep learning-based methods that incorporate spatiotemporal features and theory-based mechanistic models for a better understanding of disease spreading and improving forecasting accuracy and explainability. The aims are 1) improving epidemic forecasting accuracy by proposing graph neural network-based frameworks that consider temporal and spatial signals using a novel large scale mobility dataset, 2) improving explainability and accuracy of deep learning-based forecasting models by combining deep learning models with theory-based mechanistic models to incorporate epidemiological context.

First, we proposed a mobility informed graph neural network-based framework to capture cross-location co-evolving disease dynamics for better spatio-temporal epidemic forecasting. The proposed frameworks leverage priors from domain knowledge and mobility data. The priors are employed to instruct the model learning with the aim to allow for easier interpretation of the model and forecasting results. We incorporated large-scale aggregated spatio-temporal mobility data into graph neural networks. The proposed model provides a natural representation of disease and human mobility dynamics to develop spatially explicit forecasts thus leading to better forecasting accuracy.

Second, we proposed TDEFSI that works towards enhancing deep learning models with theory-based mechanistic models with the aim of providing accurate predictions and gaining a mechanistic understanding from a learned model. TDEFSI combines deep learning models and mechanistic models in a sequential learning process. In TDEFSI, mechanistic models are used to generate context-specific synthetic training data and then deep neural networks are trained with that synthetic data. Accurate high geographical resolution forecasting was achieved by using high-performance computing simulations. Furthermore, the explainable power of the proposed framework was explored by what-if scenario analysis.

Third, we further proposed CausalGNN that uses a causal module to mutually provide and embed causal features to get epidemiological context. CausalGNN adopts a joint learning process that learns a latent space to combine the spatiotemporal and causal embeddings using graph-based non-linear transformations. The learned model employs a causal mechanistic model to provide epidemiological context thus leading to better forecasting accuracy and better understanding of the underlying phenomena. In addition, the learned model can generate meaningful disease model parameters leading to explainable predictions.