State Space Gaussian Process Prediction

Learning accurate models of complex clinical time-series data is critical for understanding the disease and its dynamics. Modeling of clinical time-series is particularly challenging because: observations are made at irregular time intervals and may be missing for long periods of time. In this work, we propose a new model of clinical time series data that is optimized to handle irregularly sampled and missing observations. Our framework combines two models: the linear state-space and the Gaussian Processes (GP) models, into a novel dynamical model, named State Space Gaussian Process (SSGP). The model is learned using the expectation-maximization algorithm that iterates between inferences in the dynamic model and learning of the parameters of the underlying SSGP dynamic model. Experiments on real-world clinical time-series data show that the model outperforms alternative time-series prediction models.