| Description |
Patient data are collected over time at varying time intervals to update patient status and to support medical decisions, leading to a wide variety of patient time-series data. These data are used for a variety of predictive analytics tasks such as outcome predictions (e.g., mortality, diagnosis, and adverse events), medical expense predictions, and perceived provider and clinic performance predictions. Leveraging temporal patient data for predictive analytics tasks requires addressing different challenges including effective temporal data representation (e.g., time abstraction strategies), understanding and handling of missing data (e.g., imputation of missing data), discovering temporal patterns of predictive power (e.g., change points, structural patterns), selecting and developing prediction methods (e.g., model-based, instance based) that are synergistic of the model features. In this dissertation we attempted to find effective strategies that assemble varying suites of methods to tackle each of these challenges in three different studies for three different healthcare applications including early prediction of mortality in intensive care unit (ICU), acute kidney injury (AKI) prediction, and medical expense prediction. For temporal abstraction we adopted fine-grain and coarse-grain abstraction methods. To address the temporal classification challenge, we propose a similarity-based method and adopt a model-based method. To detect innovative temporal patterns, we propose change point detection methods (with regard to each study) and hyper parameters for convolutional neural networks, and also adopt structural pattern recognition methods. The results show that fine-grain abstraction is effective on any kind of healthcare application. Since most healthcare research uses coarse-grain temporal abstraction to represent temporal data, this dissertation suggests that it is essential to engineer detailed fine-grain temporal features into building predictive models using patient data. Change point detection methods are able to successfully extract temporal patterns where the healthcare application needs to detect fluctuations in patient health status. Moreover, structural pattern recognition is useful for healthcare applications where local trends in patients' data do not follow global trends. Finally, as a new field of predictive methods, convolutional neural networks show a remarkable ability in learning features that detect various kinds of temporal patterns. |
