Grant: Discovery Grants Program - Individual

Risk prediction, the core of survival analysis, is widely adopted in many areas including but not limited to medicine, public health, economics, and engineering. Classically it is based on risk factors observed at a single time point, whereas it has recently evolved to dynamic risk prediction, i.e., the risk is predicted from all the historical information. Utilizing more information, dynamic risk prediction is likely to result in an enhanced precision. It will hence help decision-makers more accurately identify the population at risk and facilitate the allocation of preventive strategies. However, existing approaches for dynamic risk prediction usually involve strong and unrealistic assumptions which may lead to biased prediction. My long-term objective is to create a convenient access to accurate dynamic risk prediction. In the context of, e.g., wildfire management in Canada, such an objective may be accomplished by integrating a web-based calculator for dynamic risk prediction into the Canadian Wildland Fire Information System. Taking the first step, my team and I will concentrate on the following two short-term objectives: (O1) to apply deep learning and functional data analysis to dynamic risk prediction accommodating mixed types of inputs and enjoying fewer restrictions and (O2) to theoretically bound the prediction error for various network architectures including those developed for objective (O1). As for objective (O1), we will resort to deep learning to extract information from longitudinal observations. Taking use of the yield of neural networks, techniques borrowed from functional data analysis are anticipated to play a central role in constructing individualized survival functions. Objective (O2) plans to express prediction errors in terms of network hyper-parameters such as the width and depth, which is not trivial at all: it remains open to theoretically bound the error for prediction via deep learning except for simple network architectures. When applied to a medical dataset with patients' time to event and follow-ups of time-dependent risk factors, our prediction models can highlight significant factors for the event and further provide evidence for precision medicine. Also, the findings from our study may contribute to an efficient engineering risk management by precisely figuring out the potentially risky projects. Another spotlight is the theoretically constructed error bound which can make the prediction more reliable, facilitate the comparison among our models and competitors, and inspire us on the optimal network architecture. This proposal will be boosted by the participation of highly qualified personnel (HQP). Also, it will provide HQP with opportunities to reach the frontline of research on both statistics and deep learning and will endow them with skills beneficial to their future development in both academia and industry.