Towards Early Detection of Cognitive Decline: Disentangling Retest Effects with a Bayesian Ornstein-Uhlenbeck Process

Google map

The preclinical phase of Alzheimer's disease and related dementias begins decades before diagnosis, making early detection a priority for secondary intervention. Smartphone-based cognitive assessments now allow us to measure cognitive abilities repeatedly and in an ecologically valid manner, but repeated testing comes with a catch: people get better with practice. Unaccounted for, these retest effects could mask genuine decline, and individual differences in how much and how quickly someone learns may themselves carry information about cognitive health.

In this talk, I present a Bayesian multilevel Ornstein-Uhlenbeck (OU) process for intensive longitudinal cognitive data. The continuous-time OU model handles irregularly spaced assessments, and we extend it with a time-varying mean that follows the exponential law of practice. This decomposes each person's response-time series into cognitive features (asymptotic ability, attractor force, volatility) and retest features (retest gain, learning rate), estimated in one step alongside effects related to person-level predictors.

We apply the model to the first measurement burst of the MindTrack study, in which 200 community-dwelling adults aged 40-70 completed a color-shapes feature-binding task five times a day for 14 days. Older age predicted slower asymptotic response times and larger retest gains: older adults had both a lower performance ceiling and more room to improve with practice. Plasma pTau217, a Core 1 Alzheimer's biomarker, was also included as a predictor but showed no credible associations in this healthy mid-life sample.

I will discuss limitations and next steps, including extending the model across bursts to separate practice from true cognitive change and incorporating additional blood biomarkers.