Artificial Intelligence Solutions for Reliable Epidemic Forecasting
Abstract:
Infectious diseases, such as seasonal influenza, Zika, Ebola, and the ongoing COVID-19 pandemic, place a heavy social and economic burden on our society. Producing timely, well-informed, and reliable spatio-temporal forecasts can help for public health decision making, effective interventions and optimal mitigation resource allocation.
However, producing \textit{reliable} epidemic forecasting using deep learning models is challenging due to the lack of \textit{explainability} and \textit{consistency} of deep learning models, where explainability is the ability of a learned model to provide explanation for the underlying phenomena and consistency is the ability of a model to consistently produce correct predictions between multiple generations of a trained model.
My research focuses on deep learning-based (DNN) methods that incorporate physical knowledge, spatio-temporal features, and multi-source data analysis for a better understanding of disease spreading and improving forecasting accuracy. The aims are improving explainability of deep learning models using ground truth-based and casual theory-based explainable factors for epidemic forecasting, and improving consistency of deep learning models using ensemble techniques with theoretical proofs and empirical demonstrations.
First, I am exploring methods to integrate ground truth-based explainable factors into graph neural network-based models for forecasting spatio-temporal epidemic dynamics with post-hoc explanations of the underlying phenomena. I leverage dynamic human mobility information to enhance a model leading to explainable accurate epidemic forecasting. This is one of the first works of graph neural networks adapted to epidemic forecasting.
Second, I am investigating methods to combine casual theory-based explainable factors into deep learning models for accurate spatio-temporal epidemic forecasting. I propose to train deep learning models with casual theory generated synthetic training data. The casual theory generated data is context specific and can represent unique properties of a given region thus provides implicit explanation of the learned model. The learned model enables accurate high-resolution forecasting even when the observations are only available at an aggregated level. The combination of theory and data-driven machine learning is an important and emerging approach to scientific problems that are data sparse. This is one of the first work to combine epidemic models from mechanistic domain and AI domain.
Third, I am incorporating physical knowledge and causal mechanism process into DNN-based forecasting models as an explanation generation module in an effort to develop self-explainable machine learning methods for reliable forecasting. The trained models can capture the underlying causal mechanisms of disease spreading. This will spark more interest in applying AI-based techniques for reliable forecasts.
Finally, I am inquiring deeply into ensemble methods to improve consistency and correct-consistency of deep learning models for producing reliable predictions between multiple training runs. I explore the problem in a theoretical way and demonstrate the value of theorems and proposed methods on real-world datasets. This is the first work to formally define consistency of deep learning models in a general way, which leads to trustworthy AI systems.
Committee:
- Anil Vullikanti (committee chair)
- Madhav Marathe (co-advisor)
- Jiangzhuo Chen (co-advisor)
- Jundong Li
- Stephen Gardner Eubank
- Adam Sadilek