Exploring Temporal Frameworks for Constructing Longitudinal Instance-Specific Models from Clinical Data
The prevalence of the EMR in biomedical research is growing, the EMR being regarded as a source of contextually rich, longitudinal data for computation and statistical/trend analysis. However, models trained with data abstracted from the EMR often (1) do not capture all features needed to accurately predict the patient's future state and to ground clinical decisions; and (2) are not normalized to a standardized timeline. This dissertation demonstrates the advantages of instance-specific predictive models and event-based frameworks for normalizing population and patient-specific data, by evaluating a modified Lazy Bayes' Rule algorithm adapted to structured and unstructured longitudinal clinical datasets. The results of the evaluations indicate the superior performance of the instance-specific model over its global equivalents, in the context of staging clinical data via event-based change.