Initial Results in Alzheimer's Disease Progression Modeling Using Imputed Health State Profiles

摘要

This paper describes an initial step in developing a set of quasi-patient profiles, each representing a complete longitudinal medical history of Alzheimer's disease (AD) - from normal health to the clinical emergence of the disease and beyond. Quasi-patient is the term given to a unified medical record created through the optimal imputation of individual records, and the guided merger and completion of multiple patient records from the Alzheimer's Disease Neuroimaging Initiative (ADNI). In the present paper, imputation strategies and boosted ensemble decision trees are used to characterize the health states of patients in the ADNI database which consistently yield year-by-year health state predictions of 80% or greater accuracy. In addition, relative to ordinarily ignoring missing medical records in a patient's history, imputation and state estimation guided by globally-optimal decision criteria resulted in an accuracy increase from 76.1% to 81.9%.