Abstract:
|
Wide availability of electronic medical records nowadays allows for a more comprehensive understanding about disease progression dynamics and their associations with patients' outcomes which is critical in early detection, risk stratification, and personalized treatment. However, few computationally efficient algorithms adequately capture the underlying heterogeneity between patients. In addition, clinical findings (e.g. ICD9 diagnosis codes) from hospital visits is high dimensional and cannot be used directly to represent disease severity. We will describe a continuous-time hidden Markov model to learn the hidden states of a disease and patient-level progression patterns. To model the effect of the covariates, we assume that the log of transition rates among hidden states are related to a linear function of covariates. Our disease progression model can explicitly include demographic information and characterize treatment effects under different treatments. We will show how our model parameters can be estimated using the EM algorithm, and illustrate its usefulness through simulation studies and an application to a congestive heart failure patient cohort.
|