P1‐230: Personalized Predictive Modeling ror Alzheimer’s Disease Patients

Background:The rate of progression of Alzheimer’s disease (AD) varies from patient to patient, hindering attempts to generate accurate estimates of the course of the disease or time until specific disease endpoints. We previously published a longitudinal Grade of Membership (L-GoM) model that accurately predicts time to multiple endpoints in patients with AD. Here we report a new L-GoM model, based on a different patient cohort. Model predictions can be incorporated into Sullivan’s life table method. This approach allows straightforward analysis of disabled and healthy/disabilityfree life expectancy for individuals or groups. Methods: The LGoMmodel with four AD subtypes (numbered I–IV) characterized the natural history of AD in the Predictors 2 cohort, consisting of 229 participants followed over 21 biannual examinations (1997– 2011). The model included six fixed and 74 time-varying variables spanning 11 symptom domains including cognition, functioning, dependence, and behavior. GoM predictions of survival and specific disability endpoints were then incorporated into an extension of the Sullivan method. Results:Among the four AD subtypes, subtype IV had high patient dependence, poor cognition, functioning, and behavior, and high CDR ratings. Individual patient characteristics were well described, initially using weighted combinations of parameters for subtypes I–III, with longitudinal changes manifested through increasing weights on subtype IV. AD progression was fastest for subtype II. Psychiatric and behavioral symptoms were most prominent for subtype III. Subtype I had the lowest initial severity, slowest progression, and best prognoses. The L-GoM model accurately predicted survival in selected patient subgroups. Using the Sullivan method, we created estimates of disabled and healthy/disability-free life expectancy for endpoints including nursing home care and specific MMSE or CDR scores. Conclusions: AD patients were heterogeneous both at initial intake and with respect to their subsequent rates of progression; adequate characterization required a multivariate approach with informative variables relating to the four AD subtypes. The Sullivan life table extension using L-GoM can be used as the core component of endpoint and resource utilization/cost calculations for individual AD patients. This approach provides a powerful new tool ready for immediate use in personalized predictive modeling in clinical, research, and public health applications.