Partially Observable Predictor Models for Identifying Cognitive Markers.

Repeated assessments of cognitive performance yield rich data from which we can extract markers of cognitive performance. Computational cognitive process models are often fit to repeated cognitive assessments to quantify individual differences in terms of substantively meaningful cognitive markers and link them to other person-level variables. Most studies stop at this point and do not test whether these cognitive markers have utility for predicting some meaningful outcomes. Here, we demonstrate a partially observable predictor modeling approach that can fill this gap. Using this approach, we can simultaneously extract cognitive markers from repeated assessment data and use these together with demographic covariates for predictive modeling of a clinically interesting outcome in a Bayesian multilevel modeling framework. We describe this approach by constructing a predictive process model in which features of learning are combined with demographic variables to predict mild cognitive impairment and demonstrate it using data from the Einstein Aging Study.