Interplay of physical and recognition performance using hierarchical continuous-time dynamic modeling and a dual-task training regime in Alzheimer's patients

Abstract

Training studies typically investigate the cumulative rather than the analytically challenging immediate effect of exercise on cognitive outcomes. We investigated the dynamic interplay between single-session exercise intensity and time-locked recognition speed-accuracy scores in older adults with Alzheimer's dementia (N = 17) undergoing a 24-week dual-task regime. We specified a state-of-the-art hierarchical Bayesian continuous-time dynamic model with fully connected state variables to analyze the bi-directional effects between physical and recognition scores over time. Higher physical performance was dynamically linked to improved recognition (-1.335, SD = 0.201, 95% Bayesian credible interval [BCI] [-1.725, -0.954]). The effect was short-term, lasting up to 5 days (-0.368, SD = 0.05, 95% BCI [-0.479, -0.266]). Clinical scores supported the validity of the model and observed temporal dynamics. Higher physical performance predicted improved recognition speed accuracy in a day-by-day manner, providing a proof-of-concept for the feasibility of linking exercise training and recognition in patients with Alzheimer's dementia.

Highlights: Hierarchical Bayesian continuous-time dynamic modeling approachA total of 72 repeated physical exercise (PP) and integrated recognition speed-accuracy (IRSA) measurementsPP is dynamically linked to session-to-session variability of IRSAHigher PP improved IRSA in subsequent sessions in subjects with Alzheimer's dementiaShort-term effect: lasting up to 4 days after training session.

Keywords: Alzheimer's disease; Bayesian; cardiovascular training; cognitive performance; continuous‐time modeling; dynamic modeling; hierarchical; intervention; longitudinal analysis.

FIGURE 1

Schematic illustration of the two‐state model with the first three timepoints (t0, t1, t2) reflecting successive training sessions. The graphical model contains the observed variables (manifest indicators [MANIFESVAR]), power of the bicycle ergometer and heart rate (HR; as the ratio: Power/HR) and reaction time corrected for the proportion of error (linear integrated speed‐accuracy score [LISAS]), loading on the latent variables (ellipsoids) of physical performance (PP) and integrated recognition speed‐accuracy performance (IRSA), respectively. The main effect of interest is the cross‐effect of PP on IRSA (further denoted as cross‐effect drift_PP→IRSA). The model also contains latent error terms (w) and the continuous‐time intercept (triangle). The model shows regression paths (red lines) and variance and covariance (orange lines). Manifest intercepts are not shown.

FIGURE 2

Auto‐ and cross‐regression over time. Temporal autoregressive effects (upper panel) and cross‐lagged effects (lower panel) over time (x‐axis, time interval in days), median and 95% quantiles for a change of 1 at time zero. The expected autoregressive effect (or self‐connection) of physical performance (PP; drift PP) and integrated recognition speed‐accuracy (IRSA; drift IRSA) peak around approximately 1 day and decrease with increasing time interval length. This suggests that the more time passes the less predictive is the performance for consecutive performance levels. The expected cross‐lagged effect (or interplay) of PP on IRSA peaks around 1 day and seem to improve predictions of IRSA for up to around 4 days. This can be understood as rather short‐term benefits from physical training on cognitive performance. The cross‐lagged effect of IRSA on PP is very close to zero, suggesting that changes in cognitive performance do not improve predictions of physical performance.

FIGURE 3

Individual estimates of integrated recognition speed‐accuracy (IRSA) and physical performance (PP), showing individual‐level analyses for all participants of the sample (n = 17) over the time interval in days (x‐axis). The solid lines present the model prediction of the smoothed estimates of participant's individual latent states IRSA (upper panel) and PP (lower panel) within a 95% Bayesian credible interval (BCI). Each colored solid line presents the individual model prediction for one subject. The temporal dynamics of PP show more individual differences compared to IRSA.

FIGURE 4

Estimated effect of subject‐level covariate predictors on dynamic parameters. We show (A) MMSE and (B) SF12 physical health baseline score effects on drift parameters (auto‐effects and cross‐effects) within a 95% Bayesian credible interval (BCI). drift_PP, auto‐effect PP (red solid line); drift_PP→IRSA, cross‐effect PP on IRSA (green solid line); drift_IRSA→PP, cross‐effect IRSA on PP (blue solid line); drift_IRSA, auto‐effect IRSA (purple solid line); IRSA, integrated recognition speed‐accuracy; MMSE, Mini‐Mental State Examination; PP, physical performance; SF12, 12‐Item Short Form Survey.