Applying an Instance-specific Model to Longitudinal Clinical Data for Prediction.

Abstract

Dynamic Bayesian Belief networks (DBNs) have been commonly used to represent temporal data in several domains; however, an ideal representation requires a near perfect mapping between the process being modeled and the DBN. Furthermore, DBNs assume a full set of observations collected at a fixed frequency. Bayesian model selection has arisen to address biased inference and underlying assumptions about the data (e.g., distribution, representativeness) to choose a model that best fits the given observations. Per patient case, a Bayesian model is generated to maximize specificity, and the collective set of models is averaged to fit all examples. This paper demonstrates the advantages of patient-specific modeling over a DBN-driven approach. Results evaluating this approach are presented based on models for two longitudinal clinical datasets (neuro-oncology, knee osteoarthritis). Largely, the patient-specific models show improved performance in prediction relative to the DBNs.