Mixed-Effects Models Neural Networks for Improved Speech-Based Predictions of Cognitive Decline

Abstract:
This study explores mixed-effects machine learning approaches to predict future cognitive test scores based on speech and demographic data from the Interdisciplinary Longitudinal Study on Adult Development and Aging. Framing the task as a longitudinal prediction problem, we compare standard and mixed-effects versions of linear models, random forests (RF), and neural networks (NN), including a linear mixed-effects neural network (ME-NN). In a leave-one-sample-out evaluation, mixed-effects models outperform standard (non-mixed) models when the current cognitive score is not included, highlighting their relevance for scenarios where such information is unavailable. Under this condition, the best-performing model was a mixedeffects random forest with a mean squared error (MSE) of 0.840. When the current score is available, predictive performance improves, with the best MSE of 0.836 from a standard RF. ME-NN consistently outperforms standard NN, emphasizing the value of explicitly modeling individual variation.