Interplay of physical and cognitive performance using hierarchical continuous-time dynamic modelling and a dual-task training regime in Alzheimer’s patients

2.1 Sample and experimental design

We recruited older adults aged 60-80 years with diagnosed mild to moderate Alzheimer’s disease (F00.1; Mini Mental Status Examination (MMSE): 18-26) from the memory clinic of the German Center for Neurodegenerative Diseases (DZNE), Magdeburg (see Supplemental Material: Methods). The total sample size included N = 17 older adults (age: M = 73.33 ± 3.43; MMSE: M = 23.50 ± 3.45; female = 8) who were free of depressive symptoms (Geriatric Depression Scale: M = 1.88 ± 0.93), pulmonary and cardiovascular diseases. The study contained a 24-week (72 sessions, 15-min each) dual-task regime (for details see also (19)). The training support per week was provided by the relatives, a team member and a senior volunteer. Each participant received personalized training with an individualized cadence (40- 80 rotations per minute) and exercise intensity of 65-75% of the maximum heart rate (HR, ∼ 90-115 bpm) using the Karvonen method (21). Simultaneously, each training session included a cognitive task with a memorization of 30 pictures, such as animals or landscapes. The picture recognition memory performance was assessed immediately after training and included 30 subsequent test screens (original vs. lure picture). All picture sets were matched regarding difficulty level using LaMem score (22). The MMSE and 12-Item Short Form Survey (SF-12) physical health questionnaire (23) were assessed before and immediately after the regime.

2.2 Dynamical Bayesian modelling and statistical analysis

We included up to 72 training sessions per participant in the analysis if completed (90% in sample). The training observations were z-standardized on the group level. Physical performance (PP) was measured as the ratio of power output (unit: Watt) and HR of each training session. Higher values reflect higher PP. Cognitive performance (COG) was operationalized as reaction times corrected for the number of errors. The reaction time (RT, in seconds) within a valid range of ≤ 13 seconds was used and corrected for the proportion of error (PE) using the linear integrated speed-accuracy score (LISAS) (24): with mean reaction time and proportion of errors at the measurement occasion j, and respective standard deviations. Lower scores do reflect higher cognitive performance i.e. faster reaction time corrected for error.

All statistical analyses were conducted in R version 4.0.2 using RStudio version 1.3.1056. The modelling approach was previously established and implemented in the R package ctsem (20) relying on the Stan software (25). Results were visualized using ctsem and ggplot2 from tidyverse (26). Model fit was compared using Chi-square tests. The alpha level for additional frequentist statistical tests was defined as p < .05.

Hierarchical Bayesian continuous-time dynamic modelling was used to simultaneously analyse the temporal dynamics of PP and COG reflected in a 2-dimensional state variable η(t) =[PP(t), COG(t)]^T at time t (20). At the core is a subject-level latent dynamic model using a stochastic differential equation (or state equation): with time-varying latent process η(t) and its temporal derivative dη(t) encoding the systems current state and its change over time, respectively. The DRIFT matrix A contains free parameters and defines the temporal dynamics of the process with auto-effects (connectivity: main-diagonal) and cross-effects encoding interplay or interactions (coupling: off-diagonal) Eq.(2) contains the continuous time intercept (CINT) b and the effect M of time dependent predictors χ on η(t).

The continuous-time parameters of A contain changes of η over a small time interval in the differential equation and can be transformed better interpretable discrete-time equivalents (A*) for any given time interval length (Δt) where A^*_Δtu includes the associated auto and cross-regression effect for the effect of η at the measurement occasion u-1 on η at measurement occasion u (20). The used approach also includes a linear measurement model relating latent states η(t) to observables y(t) using factor loadings Λ, the manifest intercepts τ and residuals ε with covariance matrix Θ (see Supplemental Material).

The dynamic model was specified with two fully connected state variables enabling bi- directional coupling between PP and COG over 72 measurement occasions. The observable indicator Power/HR of each session were loading on PP and reaction times (RT) on COG. All other parameters (except b and M which were set to zero) of the state equation and measurement model were left free to be estimated using the data. The latent process means at t=0, τ (intercepts) and A (DRIFT-matrix) were allowed to vary free across participants, resulting in 49 parameters. Population and individual level parameters are estimated simultaneously using all data from all subjects. The hierarchical Bayesian model estimation was set to default priors and initial starting values using 4 chains and 8.000 iterations (under Stan’s optimizer for maximum a posteriori estimates).

Four models were hypothesized and further compared. First, a full 2-CR model including both auto- and cross-effect parameters in drift matrix A resulting in 13 free population mean parameters. Secondly, two 1-CR models with both auto-effects and the cross-effect PP on COG (PP◊COG) and COG on PP (COG◊PP), restricting to uni-directional interaction only. Lastly, a zero-model (0-CR) in which both cross-effects were fixed to zero (without domain interplays). The full 2-CR model was compared against both 1-CR (and 0-CR model with regard to their model fit using a Chi-Square difference test (20).

Furthermore, we ran a second-level model with the MMSE and the SF12 baseline score as time- independent covariates for validation with the temporal dynamics. A higher score in both covariates is associated with higher cognitive and physical health status respectively. All other model specifications were the same as the above full 2-CR model.

To explore potential changes of the domain interplay over the course of the training, the strength of PP on COG (PP◊COG) was compared between the baseline (day 1-84) and second half (day 85-168) of the training (both free parameters to estimate). By specifying a slightly extended full 2-CR model, we included a time dependent predictor χ (named ‘secondhalf’), which is zero expect on day 84, when it is 1. Additional, an extra latent process (named ‘step2ndhalf’) was included, with all parameters and covariances fixed to zero, except the element on the time dependent predictor were set to 1 (step function), i.e. the extra latent process shifts on day 84 to 1 and stays there. The cross-effect in question was defined as a function of PP◊COG on baseline added by PP◊COG on the second half of the training multiplied by the extra latent process. A negative value reflects a stronger effect of PP◊COG (i.e. when physical performance increases the reaction time decreases) in the second half of the training and vice versa.

3.1 Demographic data

Table 1 provides the demographic data and associated neurological characteristics of the participants.

3.2 Analysis of cross-domain interplay using dynamic modelling

The range of valid measurement occasions of all 17 participants was between 59 – 70 sessions (M = 64.71 ± 4.26) resulting in 1100 manifest observations per latent state (or domain) with 25 (∼2.3%) missing values for physical performance. The time interval between successive measurement occasions ranged from 1 to 14 days (M = 2.54 ± 1.04). The longitudinal data was analysed using a 2-state dynamical model as illustrated in Figure 1 (for details see methods).

• Download figure
• Open in new tab

Fig. 1.

Schematic illustration of the 2-state-model with the first three timepoints (t0, t1, t2) reflecting successive training sessions. The graphical model contains the observed variables (manifest indicators) power of the bicycle ergometer and HR (as ratio: Power/HR) and reaction time corrected for the proportion of error (RT) loading on the latent variables (ellipsoids) PP (physical performance) respectively COG (cognitive performance). The main effect of interest is the cross-effect of PP on COG (further denoted as drift PP◊COG). 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.

A Chi-Square difference test for model comparison revealed a significant difference between the full 2-CR model and the zero-model (χ²(17) = 235.1, p < .001). In addition, a significant difference was also observed between the 2-CR model and the PP◊COG 1-CR model (χ²(9) = 149.8, p < .001) and COG◊PP 1-CR model (χ²(9) = 294.6, p < .001). Correspondingly, the full 2-CR model (Table 2) containing both cross-effects that enable a bi-directional interplay between COG and PP fitted the data best and is further reported (see Table S1 for estimated population parameters for the other models).

Changes in PP predict later changes in COG in the opposite direction as found by a more substantial negative cross-effect drift_PP◊COG (-1.335, SD = 0.201, 95 BCI [-1.725, -0.954]). As such, when physical performance levels are above baseline (suggesting higher PP) COG levels are likely to go downwards (faster reaction times). The cross-effect PP◊COG also varies between participants which suggests that some subjects benefit more from the exercise training resulting in cognitive improvements than others (between-person variability in drift_PP◊COG: 2.44, SD = 0.29, 95% BCI[1.86, 3.02]). The high between-person differences in the cross-effect PP◊COG comes along with its higher population mean (T0_MEANS, Table 2), i.e. this effect seems more relevant in the sample but its importance also differs more between participants. In contrast, COG do partially predict later changes of PP in the same direction as indicated by a small and positive cross-effect (drift_COG◊PP). When subject’s reaction time increases (i.e. reduced COG) the PP levels did also slightly increase over following time. Furthermore, the temporal changes of PP last longer than the temporal dynamic of COG (higher auto-effect of PP (drift_PP)).

Previous states of PP do impact COG negatively (-0.368, SD = 0.05, 95 BCI [-0.479, -0.266]) as shown via discrete-time parameters (one day time interval). The expected effect PP◊COG peaks around one and lasts for up to around four days, after which the random-state fluctuations dominate (Figure 2). The cross-lagged effect COG◊PP was observed to be close to zero (0.086, SD = 0.024, 95 BCI [0.041, 0.136]) and accordingly there is practically no substantial effect in this direction. Furthermore, the small autoregressive (self-connection) effect of COG (0.145, SD = 0.039, 95 BCI [0.072, 0.226] suggested low stability of the construct over time. PP on one day had a small effect on PP on another day (0.363, SD = 0.05, 95 BCI [0.264, 0.472].

• Download figure
• Open in new tab

Fig 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 PP (drift PP) and COG (drift COG) peak around approximately one 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 COG (PP◊COG) peaks around one day and seem to improve predictions of COG for up to around four days. This can be understood as rather short-term benefits from physical training on cognitive performance. The cross- lagged effect COG on PP (COG◊PP) is very close to zero, suggesting that changes in cognitive performance do not improve predictions of physical performance.

The temporal dynamics of PP showed more inter-individual differences compared to COG. Some subjects showed relatively persistent PP levels over the entire training time, while the PP fluctuated more in other participants (between-person variability in drift_PP: 9.89, SD = 0.52, 95% BCI[8.89, 10.90]). The measurement error (MANIFEST_VAR) of the manifest indicator Power/HR (0.208) was found to be higher compared to RT (0.046), i.e. measurement limitations and short-term situational influences (e.g. subjective stress or sunny days) are more present in the PP indicator (Table 2)., COG showed higher session-to-session fluctuations within-person (0.188, SD = 0.014, 95 BCI [0.163, 0.217]) compared to PP (0.004, SD = 0.002, 95 BCI [0.001, 0.012]). The model prediction of COG and PP over training time are illustrated in Figure 3 (see also Supplemental Material Fig. S1 for five randomly selected participants).

• Download figure
• Open in new tab

Fig. 3.

Individual estimates. Individual level analyses for all participants of the sample (n = 17) over time interval in days (x-axis). The solid lines presents the model prediction of the smoothed estimates of participant’s individual latent states COG (upper panel) and PP (lower panel) within a 95% BCI. Each coloured solid line presents the individual model prediction for one subject each. The temporal dynamics of PP show more individual differences compared to COG.

3.3 Drift coefficients as a function of clinical baseline scores

Participants with higher MMSE baseline scores show lower persistence in the cross-effect COG◊PP (drift_COG◊PP; -0.215, SD = 0.022, 95% BCI [-0.259, -0.172], Fig 4A). Higher MMSE baseline scores were associated with lower persistence in COG (drift_COG; -0.538, SD = 0.174, 95% BCI [-0.887, -0.209]). The effect of MMSE on the auto-effect PP and the cross-effect PP◊COG is close to zero (-0.02, SD = 0.003, 95% BCI [-0.03, -0.01]. Participants with a lower health score show a stronger cross-effect of COG on PP (drift_COG◊PP; -0.231, SD = 0.035, 95% BCI [-0.303, -0.165], Fig 4B). Likewise, higher physical health scores seem to be associated with lower persistence in their physical performance (drift_PP; -0.036, SD = 0.008, 95% BCI [- 0.052, -0.021]). The effect on the other auto-effect and cross-effect are close to zero.

• Download figure
• Open in new tab

Fig. 4.

Estimated effect of subject-level covariate predictors on dynamical parameters. We show (A) MMSE and (B) SF12 physical health baseline score effects on drift parameters (auto-effects and cross- effects) within a 95% BCI. drift_PP, auto-effect PP (red solid line); drift_PP◊COG, cross-effect PP on COG (green solid line); drift_COG◊PP, cross-effect COG on PP (blue solid line); drift_COG, auto-effect COG (purple solid line).

3.4 Changes of dynamics over the course of training

An extended model with the drift coefficient PP◊COG as a function of time was specified. The model estimated if the strength of the cross-effect PP◊COG changed between the baseline (day 1-84) and the second half (day 85-168) of training. Results suggested that the cross-effect PP◊COG becomes positive in the second half of the training (2.08), i.e. the strength and associated effect of physical performance on cognitive performance was found to be reduced in the second half of training. This change was estimated as a function of PP◊COG on baseline (-0.11, SD = 0.63, 95% BCI [-1.34, 1.14]) added PP◊COG on the second half of the training (-1.50, SD = 0.08, 95% BCI [-1.66, -1.34]) multiplied by the extra latent step process (-1.45, SD = 0.38, 95% BCI [-2.21, -0.72]).

Ethics approval and consent to participate

The study was approved by the ethics committee of the Otto-von-Guericke University, Magdeburg, Germany (approval number: 68/17). All participants and their relatives as representatives signed a written informed consent form for participation.

Availability of data and materials

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Competing interests

The authors declare that they have no competing interests.

Funding

The project (ID: ZS/2016/05/78611) belongs to the Research Association Autonomy in old Age (AiA) funded by the European Union (ERDF-European Regional Development Fund) and the State of Saxony-Anhalt, Germany.

Author’s contributions

GZ and SS contributed to the conceptualization and the data curation. SS, GZ, and MV performed the formal analysis. ED contributed to the funding acquisition, resources, and the supervision. The investigation was conducted by NB, WG, AB, and SS. ED, MV, GZ, AB, and SS contributed to the methodology and AB and NB to the project administration. The software and validation was conducted by AB. SS contributed to the visualization and the writing of the original draft preparation. SS, GZ, MV, WG, AB, NB, and ED contributed to the writing of the review and editing. All authors contributed to the article and read and approved the submitted version.