B0474
Title: Deep neural network with a smooth monotonic output layer for dynamic risk prediction
Authors: Zhiyang Zhou - University of Wisconsin-Milwaukee (United States) [presenting]
Abstract: The fundamental concern of risk prediction is the relationship between predictors and the survival function. The recent success of survival analysis has already been extended to dynamic risk prediction, where the model considers repeated measurements of time-varying predictors. However, existing approaches usually involve strong model assumptions (e.g., additive effects and/or proportional hazard) or discretize the time domain and approximate the survival function by a step function, which may lead to biased prediction. To tackle these issues, we present a deep neural network with a novel output layer termed the Smooth Monotonic Output Layer (SMOL). The resulting network involves no discretization and specifies no parametric structure for the underlying relationship between predictors and the time to event. Attaching SMOL to a neural network, one may infer/learn the cumulative distribution function for a continuous random variable, directly and nonparametrically. We conduct experiments on datasets from the Lifetime Risk Pooling Project (LRPP). LRPP pools together individual data from twenty community-based studies on cardiovascular disease and involves around three hundred thousand participants with long-term follow-ups of longitudinal risk factors (e.g., blood pressure and cholesterol). Extensive results show that our proposal achieves state-of-the-art accuracy in predicting the individual-level risk of atherosclerotic cardiovascular disease.
