Personalized Long-Term Prediction of Cognitive Function: Using Sequential Assessments to Improve Model Performance

Abstract Prediction of onset and progression of cognitive decline and dementia is important both for understanding the underlying disease processes and for planning health care for populations at risk. Predictors identified in research studies are typically accessed at one point in time. In this manuscript, we argue that an accurate model for predicting cognitive status over relatively long periods requires inclusion of time-varying components that are sequentially assessed at multiple time points (e.g., in multiple follow-up visits). We developed a pilot model to test the feasibility of using either estimated or observed risk factors to predict cognitive status. We developed two models, the first using a sequential estimation of risk factors originally obtained from 8 years prior, then improved by optimization. This model can predict how cognition will change over relatively long time periods. The second model uses observed rather than estimated time-varying risk factors and, as expected, results in better prediction. This model can predict when newly observed data are acquired in a follow-up visit. Performances of both models that are evaluated in10-fold cross-validation and various patient subgroups show supporting evidence for these pilot models. Each model consists of multiple base prediction units (BPUs), which were trained using the same set of data. The difference in usage and function between the two models is the source of input data: either estimated or observed data. In the next step of model refinement, we plan to integrate the two types of data together to flexibly predict dementia status and changes over time, when some time-varying predictors are measured only once and others are measured repeatedly. Computationally, both data provide upper and lower bounds for predictive performance.