| Dynamic Bayesian networks (DBNs) are graphical models to represent stochastic processes. This dissertation investigates the use of DBNs to predict patient outcomes based on temporal data, the effectiveness of DBNs on nonstationary multivariate time series data, and the assumptions on the parametric nature of DBNs along with two related hypotheses: (1) Given the assumption that the dataset was generated by stationary and first-order Markov processes, patient-specific DBNs, each of which models a single patient, would predict patient mortality more accurately than DBNs that model an entire patient population. (2) The predictive performances of patient-specific DBNs would improve by relaxing the stationary and first-order Markov assumptions.; Both hypotheses were tested on two datasets: A dataset of 6704 intensive care unit patients and a dataset that was generated through a nonstationary process simulation. The hypotheses were not supported by the results that were evaluated through receiver operating characteristics analysis.; In light of this evidence, a new class of DBNs, which is called dynamic simple Bayes (DSB) models, is developed in this dissertation. The DSB approach further restricts the parametric nature of DBNs with a set of conditional independence assumptions; that is, all temporal variables in any time period t are conditionally independent given the temporal variables in the next time period t + 1. Unlike conventional DBNs, temporal arcs of the DSB models are not in the direction of time flow. Test results suggest that DSB models are superior to conventional DBNs in predicting the next-day patient mortality and in predicting future outcomes on nonstationary multivariate time series data.; The results of this dissertation imply that relaxing parametric restrictions (e.g., relaxing assumptions on the Markov orders of processes, on the stationary characteristics of probability distributions, or on the conditional independencies between variables) may lower predictive performances of DBNs in multivariate time series data. The results further suggest that the DSB approach would be the preferred baseline for modeling multivariate time series with large sample space and relatively small sample size. |
