Data-driven detection of lucid episodes in advanced dementia: Application of latent class analysis to observational data

Abstract

Background

Lucid episodes (LEs) in advanced dementia are significant clinical events yet are challenging to investigate as they are characterized by a transient and unexpected recovery of abilities and present variably across individuals. Prospective observational studies of LEs in people living with advanced Alzheimer's disease and related dementias currently rely on expert reviewers/informants to detect and confirm LEs, in a process that is arduous and not always feasible.

Objective

We seek to examine the utility of data-driven methods for within-person LE detection and determine if such methods need to be individualized.

Methods

We fit multiple latent class analysis (LCA) models to examine longitudinal segments (n = 1283) from video-observations of participants living with advanced dementia (N = 3) who had informant validated LEs using previously established procedures. Fitted models included both individually specified classes and classes constrained to be equal across participants. Estimated classes were compared to the prior informant validations.

Results

Model fit was best with a three-class model, fit separately for each individual. Multiple model fit measures deteriorated when class definitions were constrained to be equal across participants. For each participant, there was a clear candidate lucid class containing the majority of validated LEs.

Conclusions

This report demonstrates the potential for using LCA for data-driven detection of LEs and demonstrates that detection may require participant-level modeling.