Positive affect disrupts neurodegeneration effects on cognitive training plasticity in older adults

Abstract

Cognitive training for older adults varies in efficacy, but it is unclear why some older adults benefit more than others. Positive affective experience (PAE), referring to high positive valence and/or stable arousal states across everyday scenarios, and associated functional networks can protect plasticity mechanisms against Alzheimer’s disease neurodegeneration, which may contribute to training outcome variability. The objective of this study is to investigate whether PAE explains variability in cognitive training outcomes by disrupting the adverse effect of neurodegeneration on plasticity. The study’s design is a secondary analysis of a randomized control trial of cognitive training with concurrent real or sham brain stimulation (39 older adults with mild cognitive impairment; mean age, 71). Moderation analyses, with change in episodic memory or executive function as the outcome, PAE or baseline resting-state connectivity as the moderator and baseline neurodegeneration as the predictor are the methods used in the study. The result of the study is that PAE stability and baseline default mode network (DMN) connectivity disrupted the effect of neurodegeneration on plasticity in executive function but not episodic memory. The study concludes that PAE stability and degree of DMN integrity both explained cognitive training outcome variability, by reducing the adverse effect of neurodegeneration on cognitive plasticity. We highlight the need to account for PAE, brain aging factors and their interactions with plasticity in cognitive training.

Introduction

Cognitive training for older adults has garnered widespread interest from its potential for maintaining cognition or delaying its decline (Rebok et al., 2014; Lin et al., 2016, 2020; National Academies of Sciences, Engineering, and Medicine, Health and Medicine Division, Board on Health Sciences Policy et al., 2017). While the aging brain remains plastic, even in pre-clinical stages of Alzheimer’s disease (AD), such as mild cognitive impairment (MCI) (Burke and Barnes, 2006), evidence supporting cognitive training is inconsistent(National Academies of Sciences, Engineering, and Medicine, Health and Medicine Division, Board on Health Sciences Policy et al., 2017; Sikkes et al., 2021). The efficacy of cognitive training varies widely across individuals, with some benefitting more than others (Ball et al., 2002; Owen et al., 2010; Voss et al., 2010; Basak et al., 2020). Although older adults at risk for AD are a primary population of interest, interactions between brain aging factors (e.g. AD-related neurodegeneration) (Jack et al., 2015) and plasticity have received limited consideration in understanding variability in cognitive training outcomes. Older adults with MCI are highly varied in their AD-related neurodegeneration which refers to the atrophy in selected brain regions (e.g. entorhinal cortex) (Khan et al., 2014) particularly vulnerable to AD pathologies. Neurodegeneration may interfere with neurogenesis and synaptic remodeling, processes that are critical for brain plasticity (Haass and Selkoe, 2007; Mu and Gage, 2011; Müller-Schiffmann et al., 2016) and contribute to individual variability in cognitive training outcomes. There is therefore a pressing need to better understand factors that may protect against or disrupt the adverse effect of neurodegeneration on plasticity. Progress in this area will help identify at-risk older adults who will most likely benefit from cognitive training, as well as inform intervention design for those who do not benefit from existing paradigms.

Positive affective experience (PAE) refers to the extent to which individuals experience pleasant and active states across everyday scenarios. A shift towards positive affect (i.e. ‘positivity effect’) is a common phenomenon in aging, even in MCI and AD (Gorenc-Mahmutaj et al., 2015; Waring et al., 2017; Verissimo et al., 2022). Older adults are thought to suppress negative affect more effectively and show greater affective stability (Carstensen et al., 2000). Recent longitudinal findings show that individual differences in affective experience predict cognitive outcomes, with positive affect associated with less memory decline (Askelund et al., 2019; Choi et al., 2019; Formánek et al., 2020; Hittner et al., 2020). Further, PAE can protect cognition against risk factors, such as depression and anxiety (Donovan et al., 2014; Formánek et al., 2020) and genetic factors (e.g. APOE ɛ4) (Gharbi-Meliani et al., 2021). Yet, little to no empirical work has investigated whether and how PAE affects training gains.

Moreover, cognition and PAE are supported by common resting-state networks, including the default mode (DMN), frontoparietal control (FPCN) and ventral attention (VAN) networks (Arioli et al., 2021; Lindquist and Barrett, 2012; Lindquist et al., 2016; Mikkelsen et al., 2022). Resting-state networks are characterized by stronger functional segregation, reflected by higher within-network than between-network functional connectivity (FC) (Bullmore and Sporns, 2009). Within-network FC in the DMN, FPCN and VAN progressively weakens over time in brain aging and contributes to cognitive and affective dysregulation (Malagurski et al., 2020; Zhang et al., 2022). These networks exhibit functional aberrations in AD (Bai et al., 2008; Li et al., 2012; Badhwar et al., 2017; Ibrahim et al., 2021), but they are also consistently implicated in training gains (Lin et al., 2020).

We recently completed a randomized control trial on the effect of visual attention cognitive training with(out) concurrent transcranial direct current stimulation on left somatomotor cortex (L-SMC) on affective symptoms in older adults with MCI (Turnbull et al., 2023). We found tentative evidence for reduced negative affective symptoms via changes in L-SMC activation and L-SMC—amygdala resting-state FC, suggesting improved emotional regulation. In our secondary analyses, we observed two phenomena: (i) there were heterogeneous improvements in EM (episodic memory) and EF (executive function); and (ii) participants across groups displayed varying degrees of positive affect, indicated by more positive valence ratings above ‘neutral’ on the Self-Assessment Manikin (SAM) (Turnbull et al., 2023). Here, we investigated why some older adults benefit more from cognitive training than others by examining the degree and stability of PAE during the period of a 4-week intervention. We also explored whether preserved functional integrity of brain networks that support cognition and PAE play a role in disrupting the effect of neurodegeneration on plasticity. We hypothesized that individual variability in PAE and the integrity of networks vulnerable to brain aging will predict training gains in older adults at risk for AD.

Methods

Design

We used data from a recently completed Stage 0 double-blinded randomized control trial study. The clinical trial was pre-registered as NCT04099524 at clinicaltrial.gov. Details of the design and intervention can be found in the primary report (Turnbull et al., 2023). Briefly, a sample of 40 older adults with MCI were randomly assigned to a 4-week intervention of cognitive training with or without concurrent transcranial direct current stimulation (cognitive training with vs without tDCS). The intervention was the application of tDCS; all participants completed the same cognitive training for five consecutive sessions/week for 2 weeks, followed by two sessions/week for 2 weeks. Participants and outcome assessors were blinded to group randomization. Cognitive training consisted of a multiple object tracking task (Pylyshyn and Storm, 1988) (see Supplementary, Section 2 for details). Data used in the present study included structural and functional MRI at baseline, EM and EF at baseline and within 7 days post-intervention and affective ratings across the intervention sessions. Human subjects research procedure was approved by the University of Rochester Research Subjects Review Board. All participants gave written informed consent.

Participants

Community-dwelling older adults with MCI were enrolled. MCI due to AD was defined based on 2011 NIA-AA diagnostic criteria (Jack et al., 2018), including (i) Montreal Cognitive Assessment version 2 education-adjusted total score of 18 ≤ × ≤ 26(Nasreddine et al., 2005); (ii) one s.d. below age- and/or education-corrected population norms for Rey’s Auditory Verbal Learning Test (Lists C & D) (Harris et al., 2002; Zhao et al., 2015); (iii) activities of Daily Living-Prevention Instrument self-report total score ≤30 (Galasko et al., 2006); and (iv) absence of dementia, determined from informant-reported medical history. Participants with contraindications (e.g. MRI: pacemaker; tDCS: history of seizures, repetitive motor conditions, skin condition or sensitivity) were excluded. Baseline characteristics are presented in Table 1. The study procedure used in this study in terms of the consort diagram for our primary report (Turnbull et al., 2023) is shown in Supplemental Figure S1.

Table 1.

Baseline characteristics

N	39 a
Age (years), mean (s.d.)	71.5 (6.96)
Sex
Male	16 (41.0%)
Female	23 (59.0%)
Education (years), mean (s.d.)	15.7 (2.71)
Cognition, mean (s.d.)
EMb	−0.30 (0.41)
EFc	0.13 (0.54)
ADL-PI self, mean (s.d.)	18.20 (3.11)
ADSCT (mm), mean (s.d.)d	2.63 (0.13)
Ethnicity
Hispanic or Latino	1 (2.6%)
non-Hispanic	38 (97.4%)
Marital status
Married	23 (59.0%)
Divorced	7 (17.9%)
Never married	4 (10.3%)
Widowed	3 (7.7%)
Other	2 (5.1%)
Living arrangement
Alone	13 (33.3%)
With family, no spouse/partner	3 (7.7%)
With spouse/partner & other family	6 (15.4%)
With spouse/partner only	16 (41.0%)
Other	1 (2.6%)
Retirement status
Retired	4 (10.3%)
Not retired	30 (76.9%)
Unknown	5 (12.8%)
Top three comorbidities
Major depression	10 (25.6%)
Migraines or severe headaches	9 (23.1%)
Glaucoma	4 (10.3%)

^aOne participant who withdrew after baseline assessment was excluded from analysis.

^bEM: mean z-scored composite of verbal and visuospatial memory.

^cEF: composite of cognitive control, working memory and fluency (see Methods).

^dADSCT (mm): lower values indicate worse neurodegeneration.

EM: episodic memory; EF: executive function; ADL-PI self: Activities of Daily Life—Instrument Project, self-report; ADSCT: Alzheimer’s disease-signature cortical thickness.

Neuroimaging pre-processing

Data acquisition conducted at baseline

Imaging was acquired with Siemens 3T Prisma (VE11C), equipped with a 64-channel head coil. Acquisition began with a scout image, followed by a single-shot echo-planar imaging (EPI) MPRAGE scan [repetition time/echo time (TR/TE) = 1400/2.34 ms, slice thickness = 1 mm, resolution = 1 mm isotropic, 192 slices, PE acceleration = GRAPPA, flip angle = 70°, fat suppression = OFF, orientation = sagittal, echo spacing = 7 ms, field of view (FOV) = 256 mm] to provide high-resolution structural-weighted anatomical images. A-P and P-A field maps were acquired to correct for distortions in EPI sequences. Resting-state BOLD data were collected using a gradient EPI multiband (MB) sequence (TR/TE = 1010/44 ms, slice thickness = 2 mm, resolution = 2 mm isotropic, R = 1, MB acceleration factor = 8, flip angle = 70°, fat suppression = ON, orientation = transversal, echo spacing = 0.56 ms, FOV = 256 mm, acquisition matrix = 128 x 128, 80 slices, 253 volumes). Slice acquisition order was interleaved. Participants were instructed to keep their eyes open.

Data processing

Structural MRI: Data were processed using Freesurfer v7.2 (recon-all) (https://surfer.nmr.mgh.harvard.edu/fswiki/FreeSurferWiki). Processing consisted of motion correction, normalization, skull-stripping, registration, white-matter segmentation, smoothing and parcellation. Processed images were visually inspected for artifacts or poor quality.

Resting-state fMRI: Data were analyzed using scripts adapted from previous research using multi-band resting-state fMRI (Risk et al., 2021), using functions from FMRI Brain Software Library (FSL) v6.0.5.1 and AFNI v21.1.07. The first four volumes were dropped to allow for signal stabilization. Pre-processing consisted of motion correction (FSL MCFLIRT), distortion correction (FSL topup), slice-timing correction (FSL slicetimer), co-registration and normalization to MNI space (FSL FLIRT and FNIRT). fMRI timeseries underwent simultaneous nuisance regression (9 point nuisance regression model: six rigid body realignment parameters, global signal, and average cerebrospinal fluid and white-matter signals (Ciric et al., 2017) and temporal filtering (bandpass 0.009−0.08 Hz), followed by spatial smoothing (full width at half maximum of the Gaussian kernel 6 mm), using AFNI 3dTproject. Timeseries extraction and FC matrix generation (including r-to-z transformation) were performed using Python 3 and nilearn v0.9.1. Brain parcellation was performed using the Yeo 7 surface-based cortical atlas (Thomas Yeo et al., 2011). Co-registration and normalization, as well as FC matrices, were visually inspected for artifacts or poor quality. One participant was excluded for structural and functional motion exceeding 1 mm mean root mean square. Two participants were excluded for poor functional co-registration.

Measures

Positive affective experience

Affective dimensions of valence and arousal were measured with Self-Assessment Manikin (SAM), a non-verbal, pictographic Likert scale (Bradley and Lang, 1994). Immediately before and after each intervention session, participants rated their momentary pleasantness (valence) and activation (arousal), by pointing to one of five pictures for each dimension. The pictorial ratings for both dimensions were coded according to a 5-point Likert scale, with 1 = very little and 5 = very much. Although SAM was not developed to measure positive affect exclusively, valence ratings for all participants were ≥3 (i.e. ‘neutral’), suggesting that, in this sample, SAM measured varying degrees of positive affect. We therefore used SAM as a measure of positive affective experience. Distributions of valence and arousal ratings across groups are displayed in Supplementary Figure S3. There were no significant changes in either dimension between pre- and post-training session ratings. We therefore used pre-intervention ratings to quantify level and stability of PAE. To operationalize PAE, we derived two scores from each participant’s 14 pre-training session ratings: level, calculated by averaging valence or arousal across all sessions, with higher values indicating more positive valence or higher arousal, respectively; and stability, calculated as the standard deviation (s.d.) in valence or arousal across sessions, with lower values indicating higher stability.

Cognitive plasticity

Cognitive plasticity was quantified as the change in EM or EF from baseline to post-intervention. EM was measured as a composite mean score of Rey’s Auditory Verbal Learning Task (RAVLT) long-term percent retention (LTPR) and Brief Visuospatial Memory Test-Revised (BVMT-R) delayed recall (Benedict et al., 1996). RAVLT consists of five learning trials in which the participant learns a list of unrelated words, an interference trial, immediate recall and delayed recall. RAVLT LTPR is a standard summary score calculated as (learning trial 5/delayed recall) × 100 and represents the percentage of words retained after a 30 min delay (Zhao et al., n.d.) BVMT-R consists of three learning trials in which the participant learns the shape and location of six 2D geometric figures and draws as many as they can recall and a 25 min delayed recall. RAVLT Lists C and D and BVMT-R Form 1 were administered at baseline. RAVLT Lists A and B and BVMT-R Form 2 were administered at post-intervention. RAVLT LTPR and BVMT-R raw delayed recall scores were z-scored before calculating the composite.

EF was measured with a subset of five tasks (flanker, set-shifting, dot-counting, category fluency, 1-back) from the NIH EXAMINER v3.6 battery (Kramer et al., 2014). Detailed descriptions of the tasks are provided in the Supplementary. EXAMINER generates an EF composite by applying item response theory empirical Bayes to the raw continuous task scores. The composite is presented in the original metric used in the item response theory algorithm; the score is not a norm-referenced measure and is not adjusted for age. Further information on the validity and scoring is described in the User Manual (https://memory.ucsf.edu/research-trials/professional/examiner).

Functional network integrity

We focused on three cortical resting-state networks: default mode, frontoparietal and ventral attention networks. Mean within-network connectivity was calculated by averaging raw (i.e. positive and negative) FC for each network using in-house R scripts. Networks were defined according to the Yeo 7 atlas (Thomas Yeo et al., 2011).

Neurodegeneration

Neurodegeneration was measured as AD-signature cortical thickness (ADSCT), an empirically established index of atrophy in regions vulnerable to early AD-related neurodegeneration (Jack et al., 2015; Schwarz et al., 2016). Although the hippocampus is a hallmark region of AD, the hippocampus is excluded because ADSCT is meant to reflect regions that are prone to early atrophy in the disease progression, which starts in the entorhinal and inferior temporal cortices (Khan et al., 2014). Previous work comparing AD biomarkers has shown that ADSCT is superior to hippocampal volume and cerebrospinal fluid markers in predicting short-term (i.e. within 1 year) MCI-AD conversion (Dickerson and Wolk, 2013). Cortical thickness was extracted from Freesurfer output (aparcstats2table) for each region from T1 structural data that was parcellated with the Desikan-Killiany-Tourville atlas (Desikan et al., 2006). ADSCT was calculated as mean cortical thickness of six bilateral regions: entorhinal cortex, middle and inferior temporal gyri, inferior parietal, fusiform gyrus and precuneus (Schwarz et al., 2016).

Statistical analysis

Primary analysis

Generalized Estimating Equation model with AR (1) covariance matrix and identity link was used to analyze the within-group intervention effect: y = Time + error. Group comparison of PAE was conducted using an independent t-test. Pearson’s r correlations (partial correlations when covariates were considered) were calculated between the main variables. Correlations with PAE indices and change in EM or EF were adjusted for randomized group. Correlations with baseline network FC were adjusted for randomized group and head motion. Two sets of moderation analyses were conducted using Generalized Linear Model: change in EM or EF = baseline neurodegeneration + moderator + neurodegeneration x moderator + covariate(s). The behavioral analysis (N = 39) included change in EM or EF as the dependent variable, baseline neurodegeneration as the independent variable, PAE (mean or s.d. of valence or arousal) as the moderator and randomized group as a covariate. The brain network analysis (N = 36) included change in EM or EF as the dependent variable, baseline neurodegeneration as the independent variable, baseline within-network FC as the moderator and randomized group and head motion covariates. Predictors and moderators were standardized. Across analyses, each moderator was tested separately for each dependent variable, yielding four behavior models and three brain models. Multiple comparisons were adjusted with False Discovery Rate (FDR). Statistical significance for the behavior and brain models was adjusted for four and three comparisons, respectively. Statistical analysis was performed with R v4.2.1. R packages are reported in the Supplementary.

Secondary analysis

To account for possible confounds from baseline characteristics, we repeated the primary analyses, with baseline cognition and age as additional covariates in the behavioral analysis, and age as an additional covariate in the brain analysis.

Results

Heterogenous intervention effects on cognition

Both groups showed significant improvement in EM (B =0.65, SE =0.11, Wald’s ² =35.41, P <0.001) and EF (B =0.29, SE =0.08, Wald’s ² =12.84, P <0.001) for main effects on time, but no significant main group effects or group × time interaction effects. The null findings suggest that cognitive training but not brain stimulation contributed to cognitive improvement. Individual trajectories of cognition by group before and after intervention are displayed in Supplementary Figure S2.

Characterization of PAE

We first examined the relationships between the four indices of PAE. The intervention and control groups were similar in level and stability of valence and arousal (t values from −0.55 to 0.71, P-values >0.24). Even though all indices were significantly correlated (P < 0.018), their relationships with each other ranged widely. Degree of valence and stability of arousal had the least amount of shared variance (R² = 0.15), and degree of valence and degree of arousal had the greatest amount of shared variance (R² = 0.72). The distributions of valence and arousal by group are displayed in Supplementary Figure S3.

Moderating effect of PAE on cognitive plasticity after training

The behavior analysis sample (N = 39) had a mean (s.d.) ADSCT of 2.63 (0.13) mm. Correlational analyses on the relationships between PAE, baseline neurodegeneration and change in EM or EF showed a positive relationship between ADSCT and baseline EM (r = 0.41, uncorrected P = 0.01). We did not find significant relationships between ADSCT and baseline EF, change in EM or EF or PAE, or between PAE and baseline or change in EM or EF (Table 2).

Table 2.

Correlations between PAE, baseline neurodegeneration and baseline or change in EM and EF

Parameter 2	Parameter 1	Pearson’s r	t	Unadjusted (FDR-adjusted) P
ADSCT	Baseline EM	0.40	2.65	0.012 (0.023)
ADSCT	Baseline EF	0.17	1.06	0.295
ADSCT	Change in EM	−0.02	−0.13	0.895
ADSCT	Change in EF	0.03	0.20	0.842
s.d. arousal	Baseline EM	0.00	−0.01	0.996
s.d. arousal	Baseline EF	−0.07	−0.44	0.659
s.d. arousal	Change in EM	−0.20	−1.24	0.223
s.d. arousal	Change in EF	0.08	0.46	0.649
s.d. valence	Baseline EM	−0.16	−0.96	0.341
s.d. valence	Baseline EF	−0.14	−0.88	0.386
s.d. valence	Change in EM	−0.29	−1.81	0.079
s.d. valence	Change in EF	0.14	0.86	0.394
Mean arousal	Baseline EM	−0.01	−0.08	0.938
Mean arousal	Baseline EF	0.07	0.42	0.680
Mean arousal	Change in EM	0.21	1.31	0.199
Mean arousal	Change in EF	−0.15	−0.90	0.376
Mean valence	Baseline EM	0.20	1.25	0.221
Mean valence	Baseline EF	0.14	0.83	0.414
Mean valence	Change in EM	0.15	0.90	0.377
Mean valence	Change in EF	−0.01	−0.06	0.951

Significant P-values are denoted in bold. Baseline correlations were not adjusted for covariates. Correlations with positive affective experience or change in EM or EF were adjusted for randomized group. P-values between baseline (change in) EM and EF and Parameter 2 variables were FDR-corrected for two multiple comparisons. ADSCT: Alzheimer’s disease-signature cortical thickness (mm; lower indicates worse neurodegeneration); EF: executive function; EM: episodic memory; PAE: positive affective experience; s.d: standard deviation.

To examine whether PAE is related to training-induced plasticity, we performed a moderation analysis using generalized linear models. In our primary analysis, we included change in EM or EF as the outcome, baseline ADSCT as the independent variable, degree or stability of PAE as the moderator and randomized group as a covariate. Degree and stability of PAE were operationalized as mean and s.d., respectively, where higher mean values represent more positive PAE and lower s.d. values represent greater stability. For change in EF, we observed a main effect of arousal stability [F(34) = -14.19, P = 0.0005, corrected P = 0.002]; and an ADSCT × arousal stability interaction [F(34) = 5.51, P = 0.0004, corrected P = 0.0016]; a main effect of valence stability [F(34) = 9.63, P = 0.005, corrected P = 0.01] and an ADSCT × valence stability interaction [F(34) = 3.77, P = 0.004, corrected P = 0.008)]; and a main effect of arousal degree [F(34) = 4.63, P = 0.02, corrected P = 0.027] and an ADSCT × arousal degree interaction [F(34) = −1.92, P = 0.018, corrected P = 0.024]. There were no significant main or moderating effects of valence degree. The interaction effects indicate that the adverse effect of baseline neurodegeneration on change in EF was diminished in older adults with high arousal stability (Fig. 1A), valence stability (Fig. 1B) or arousal degree (Fig. 1C). There were no significant interaction effects on change in EM. Full models for change in EF are displayed in Table 4 and change in EM in Table 5.

Fig. 1.

Moderating effects of PAE on change in EF after cognitive training in older adults with MCI. Moderation effects were plotted above and below the median split for visualization, but all predictors were treated as continuous variables in the analyses. ADSCT: Alzheimer’s disease-signature cortical thickness (lower values indicate worse neurodegeneration); EF: executive function; s.d: standard deviation.

Table 4.

Moderating effect of PAE on change in EF after training

	s.d. arousal							s.d. valence
Predictors	B	SE	Std. B	Std. SE	F	df	Unadj. P	B	SE	Std. B	Std. SE	F	df	Unadj. P
Randomized group	0.17	0.12	0.21	0.14	1.45	34	0.147	0.19	0.12	0.22	0.15	1.51	34	0.132
ADSCT	−2.92	1.00	0.12	0.15	−2.92	34	0.0035	−1.33	0.73	0.15	0.15	−1.81	34	0.07
s.d. arousal	−14.19	4.07	0.23	0.15	−3.49	34	0.0005
ADSCT × s.d. arousal	5.51	1.57	0.48	0.14	3.52	34	0.0004
s.d. valence								−9.63	3.42	0.26	0.16	−2.81	34	0.005
ADSCT × s.d. valence								3.77	1.32	0.41	0.14	2.87	34	0.004
R 2	0.317							0.262
	Mean arousal							Mean valence
Predictors	B	SE	Std. B	Std. SE	F	df	Unadj. P	B	SE	Std. B	Std. SE	F	df	Unadj. P
Randomized group	0.14	0.13	0.17	0.16	1.07	34	0.283	0.17	0.14	0.21	0.16	1.27	34	0.205
ADSCT	8.12	3.39	0.13	0.16	2.40	34	0.396	8.40	4.79	0.08	0.16	1.75	34	0.08
Mean arousal	4.93	2.12	−0.18	0.15	2.32	34	0.0204
ADSCT × mean arousal	−1.92	0.81	−0.36	0.15	−2.37	34	0.018
Mean valence								4.80	2.77	−0.13	0.17	1.73	34	0.083
ADSCT × mean valence								−1.86	1.07	−0.28	0.16	−1.74	34	0.082
R 2	0.214							0.139

Significant P-values are denoted in bold. Refer to text for adjusted P-values. ADSCT: Alzheimer’s disease-signature cortical thickness (mm; lower values indicate worse neurodegeneration); EF: executive function; PAE: positive affective experience; s.d: standard deviation.

Table 5.

Moderating effect of PAE on change in EM after training

	s.d. arousal							s.d. valence
Predictors	B	SE	Std. B	Std. SE	F	df	Unadj. P	B	SE	Std. B	Std. SE	F	df	Unadj. P
Randomized group	0.14	0.24	0.100	0.17	0.61	34	0.544	0.15	0.22	0.100	0.16	0.66	34	0.509
ADSCT	−2.25	1.98	−0.04	0.17	−1.14	34	0.256	−2.17	1.33	−0.03	1.33	−0.17	34	0.104
s.d. arousal	−9.30	8.07	−0.16	0.17	−1.15	34	0.249
ADSCT × s.d. arousal	3.38	3.10	0.17	0.16	1.09	34	0.277
s.d. valence								−11.46	6.21	−0.23	0.16	−1.84	34	0.0651
ADSCT × s.d. valence								4.17	2.39	0.27	0.15	1.75	34	0.08
R 2	0.092							0.179
	Mean arousal							Mean valence
Predictors	B	SE	Std. B	Std. SE	F	df	Unadj. P	B	SE	Std. B	Std. SE	F	df	Unadj. P
Randomized group	0.20	0.24	0.14	0.17	0.81	34	0.417	0.24	0.24	0.17	0.17	0.97	34	0.331
ADSCT	−2.01	6.35	−0.06	0.17	−0.32	34	0.751	−7.87	8.62	−0.06	0.17	−0.91	34	0.361
Mean arousal	−0.82	3.98	0.22	0.17	1.33	34	0.836
ADSCT × mean arousal	0.41	1.52	0.04	0.16	0.27	34	0.786
Mean valence								−4.24	4.98	0.20	0.18	−0.85	34	0.395
ADSCT × mean valence								1.72	1.93	0.15	0.17	0.89	34	0.372
R 2	0.066							0.059

Significant P-values are denoted in bold. Refer to text for adjusted P-values. ADSCT: Alzheimer’s disease-signature cortical thickness (mm; lower values indicate worse neurodegeneration); EM: episodic memory; PAE: positive affective experience; s.d: standard deviation.

Associations between network integrity and PAE

Correlational analyses, adjusting for randomized group and head motion showed a negative association with VAN FC and s.d. valence (r = −0.58, P = 0.0004, corrected P = 0.0014) and positive associations between VAN FC and valence degree (r = 0.42, P = 0.015, corrected P = 0.029) and arousal degree (r = 0.39, P = 0.025, corrected P = 0.034) (Table 3). These results indicate that higher VAN FC is related to greater valence stability (i.e. lower s.d. values), more positive valence and higher arousal states.

Table 3.

Correlations between within-network FC, neurodegeneration, baseline or change in EF and EM, and PAE

Parameter 1	Parameter 2	Pearson’s r	t	Unadjusted (FDR-adjusted) P
Baseline DMN	ADSCT	0.44	2.78	0.009 (0.027)
Baseline FPCN	ADSCT	0.34	2.06	0.048 (0.072)
Baseline VAN	ADSCT	0.25	1.49	0.146
Baseline DMN	s.d. arousal	−0.22	−1.30	0.203
Baseline FPCN	s.d. arousal	−0.07	−0.38	0.703
Baseline VAN	s.d. arousal	−0.28	−1.62	0.115
Baseline DMN	s.d. valence	−0.15	−0.84	0.408
Baseline FPCN	s.d. valence	−0.13	−0.74	0.467
Baseline VAN	s.d. valence	−0.58	−4.04	<0.000 (<0.000)
Baseline DMN	Baseline EF	−0.15	−0.87	0.391
Baseline FPCN	Baseline EF	−0.03	−0.20	0.846
Baseline VAN	Baseline EF	0.05	0.28	0.784
Baseline DMN	Baseline EM	0.34	2.07	0.047 (0.055)
Baseline FPCN	Baseline EM	0.34	1.99	0.055
Baseline VAN	Baseline EM	0.36	2.20	0.043 (0.055)
Baseline DMN	Change in EF	0.30	1.75	0.089
Baseline FPCN	Change in EF	−0.13	−0.72	0.474
Baseline VAN	Change in EF	0.09	0.50	0.617
Baseline DMN	Change in EM	−0.12	−0.67	0.509
Baseline FPCN	Change in EM	−0.16	−0.94	0.355
Baseline VAN	Change in EM	−0.13	−0.76	0.455
Baseline DMN	Mean arousal	0.07	0.41	0.686
Baseline FPCN	Mean arousal	−0.07	−0.37	0.713
Baseline VAN	Mean arousal	0.36	2.15	0.039 (0.117)
Baseline DMN	Mean valence	0.23	1.35	0.187
Baseline FPCN	Mean valence	−0.03	−0.17	0.869
Baseline VAN	Mean valence	0.42	2.62	0.013 (0.039)

Significant P-values are denoted in bold. Baseline correlations were adjusted for head motion. Correlations with PAE indices or change in EM and EF were adjusted for head motion and randomized intervention group. P-values between network FC and each Parameter 2 variable were FDR-corrected for three multiple comparisons. ADSCT: Alzheimer’s disease-signature cortical thickness (mm; lower values indicate worse neurodegeneration); DMN: default mode network; EF: executive function; EM: episodic memory; FC: functional connectivity; FPCN: frontoparietal control network; PAE: positive affective experience; VAN: ventral attention network.

Moderating effect of network integrity on baseline neurodegeneration and cognitive change after training

The network analysis sample (N = 36) had a mean (s.d.) ADSCT of 2.64 (0.13) mm. Correlational analyses on the relationships between baseline network connectivity, baseline neurodegeneration and baseline or change in EM or EF showed a positive association between ADSCT and DMN (r = 0.47, P = 0.005, corrected P = 0.016) but not FPCN or VAN FC (Table 3). Positive associations between baseline EM and DMN (r = 0.37, P = 0.033) and VAN (r = 0.36, P = 0.035) FC did not survive FDR-correction (corrected P’s = 0.052). None of the networks was associated with baseline EF, change in EM or change in EF.

Next, to gain insight into the influence of baseline network integrity on plasticity after training, we performed a second moderation analysis with change in EF or EM as the outcome, baseline neurodegeneration as the independent variable, baseline network integrity as the moderator and head motion and randomized group as covariates. We observed a main effect of DMN integrity [F(30) = 117.06, P = 0.0017, corrected P = 0.005] and an ADSCT × DMN interaction [F(30) = -42.96, P = 0.0023, corrected P = 0.007] on change in EF. The interaction effect indicates that the adverse effect of baseline neurodegeneration on change in EF was diminished in older adults with high baseline DMN integrity (Fig. 1D). There were no significant moderating effects of baseine VAN or FPCN integrity on change in EF. None of the three networks moderated change in EM. Full models for change in EF are displayed in Table 6 and change in EM in Table 7.

Table 6.

Moderating effect of baseline within-network FC integrity on change in EF after training

	DMN							VAN							FPCN
Predictors	B	SE	Std. B	Std. SE	F	df	Unadj. P	B	SE	Std. B	Std. SE	F	df	Unadj. P	B	SE	Std. B	Std. SE	F	df	Unadj. P
Randomized group	−1.19	0.86	−0.24	0.17	−1.39	30	0.164	−0.24	1.04	−0.05	0.21	−0.23	30	0.817	−0.20	1.10	−0.04	0.22	−0.18	30	0.856
Head motion	0.30	0.14	0.35	0.16	2.17	30	0.030	0.20	0.16	0.23	0.19	1.23	30	0.217	0.18	0.16	0.21	0.18	1.15	30	0.249
ADSCT	6.32	2.42	−0.37	0.19	2.62	30	0.009	1.40	2.15	−0.05	0.20	−0.65	30	0.514	−0.46	1.91	0.02	0.21	−0.24	30	0.810
DMN	117.06	37.35	0.42	0.17	3.13	30	0.002
ADSCT × DMN	−42.96	14.12	−0.64	0.21	−3.04	30	0.002
VAN								15.84	19.31	0.10	0.19	0.82	30	0.412
ADSCT × VAN								−5.81	7.33	−0.14	0.18	−0.79	30	0.428
FPCN															−8.12	25.40	−0.14	0.19	−0.32	30	0.749
ADSCT × FPCN															2.74	9.56	0.06	0.20	0.29	30	0.774
R 2	0.369							0.087							0.075

Significant P-values are denoted bold. Refer to text for adjusted P-values. ADSCT: Alzheimer’s disease-signature cortical thickness (mm; lower values indicates worse neurodegeneration); DMN: default mode network; EF: executive function; FC: functional connectivity; FPCN: frontoparietal control network; VAN: ventral attention network.

Table 7.

Moderating effect of baseline within-network FC integrity on change in EM after training

	DMN							VAN							FPCN
Predictors	B	SE	Std. B	Std. SE	F	df	Unadj. P	B	SE	Std. B	Std. SE	F	df	Unadj. P	B	SE	Std. B	Std. SE	F	df	Unadj. P
Randomized group	−1.30	1.75	−0.15	0.21	−0.75	30	0.455	−1.21	1.81	−0.14	0.21	−0.67	30	0.503	−0.42	1.87	−0.05	0.22	−0.22	30	0.823
Head motion	0.07	0.28	0.05	0.19	0.25	30	0.80	0.03	0.28	0.05	0.19	0.26	30	0.794	0.10	0.26	0.07	0.18	0.39	30	0.697
ADSCT	6.16	4.92	−0.07	0.23	1.25	30	0.211	−0.64	3.72	−0.02	0.21	−0.17	30	0.863	2.62	3.24	−0.01	0.21	0.81	30	0.419
DMN	96.41	76.08	−0.12	0.21	1.27	30	0.205
ADSCT × DMN	−37.15	28.76	−0.33	0.25	−1.29	30	0.196
VAN								−6.07	33.52	−0.13	0.19	−0.18	30	0.856
ADSCT × VAN								1.86	12.71	0.03	0.18	0.15	30	0.883
FPCN															35.45	43.09	−0.15	0.20	0.82	30	0.411
ADSCT × FPCN															−14.04	16.22	−0.17	0.20	−0.87	30	0.387
R 2	0.084							0.038							0.069

Significant P-values are denoted in bold. Refer to text for adjusted P-values. ADSCT: Alzheimer’s disease-signature cortical thickness (mm; lower values indicates worse neurodegeneration); DMN: default mode network; EM: episodic memory; FC: functional connectivity; FPCN: frontoparietal control network; VAN: ventral attention network.

Secondary analysis with baseline covariates

Whether and how to adjust for differences in baseline cognition is a topic of debate in the cognitive training literature: some argue that sufficient capacity is required to comprehend and benefit from training, while others argue that lower capacity may allow for more ‘room for improvement’ (Shaw and Hosseini, 2021; Traut et al., 2021). Adjusting for age when a model includes neurodegeneration as a predictor also lacks consensus. The exact mechanisms through which aging is associated with neurodegeneration remain poorly understood, making it difficult to conclude whether neurodegeneration reflects both chronological and biological age or only biological age. In small sample sizes, however, adjustment for baseline covariates can potentially inflate type I error rate and reduce power (Kahan and Morris, 2013). Because our sample size is small, we chose to use the simpler model and only adjust for randomized group in the primary analysis. To address potential concerns about the exclusion of baseline covariates, we performed a secondary analysis in which we repeated the primary analysis, adjusting for baseline cognition and age. Results remained significant for the moderating effects of valence and arousal stability on change in EF but not for the moderating effect of arousal degree on change in EF (Supplementary Table S1). Results remained non-significant for the moderating effects of PAE on change in EM (Supplementary Table S2). Results also remained significant for the moderating effect of baseline DMN integrity on change in EF (Supplementary Table S3) and non-significant for change in EM (Supplementary Table S4).

Discussion

We investigated whether PAE disrupts the effect of neurodegeneration on training-induced plasticity in older adults with MCI. We collected cognitive and fMRI data at baseline and post-intervention, as well as repeated valence and arousal ratings across a 1-month, 14-session cognitive intervention. Valence ratings were positively skewed across the entire sample. We used this as an opportunity to examine the role of PAE in explaining variability in cognitive training gains. We performed two separate analyses to test whether two factors moderate plasticity after training: (i) degree and stability of PAE throughout training sessions, and (ii) baseline resting FC. Greater PAE stability and higher baseline DMN FC both reduced the adverse effect of baseline neurodegeneration on plasticity of EF after training.

PAE disrupts the adverse effect of neurodegeneration on gains in EF after training

Older adults are more emotionally stable and better at regulating their affective states compared to young adults (Burr et al., 2021). Conversely, older adults at risk for AD are vulnerable to affective dysfunction that manifests as neuropsychiatric symptoms (e.g. depression, anxiety, irritability). Affective instability emerges in MCI and compounds the effects of AD by accelerating neurodegeneration and cognitive decline, consequently increasing the likelihood of conversion to AD (Lee et al., 2012; Donovan et al., 2014). Our results suggest that affective stability disrupts the adverse effect of baseline neurodegeneration on cognitive plasticity in EF over time. However, whether improving affective stability would enhance cognitive plasticity in those with neurodegeneration, irrespective of an intervention, needs future experimental manipulations or interventions. For example, mindfulness or positive emotion-based meditation reduces emotional reactivity, as well as the intensity and frequency of negative affect (Vasquez-Rosati et al., 2017; Valim et al., 2019). Empirical studies have shown that mindfulness training enhances self-regulation via at least three components that promote EF: enhanced attentional control (Tang et al., 2007; Sperduti et al., 2016), improved emotional regulation (Tang et al., 2007; Hill and Updegraff, 2012) and changes in self- and body awareness. Implementing a positive induction or mindfulness meditation before cognitive training to amplify the training effect is an avenue that warrants further investigation.

Literature on the ‘broaden-and-build’ theory suggest that positive affect promotes top-down regulation of attentional processes that facilitate EF performance (Huppert et al., 2004). One study that examined the relationship between positive affect and attentional scope found that, while positive affect impairs selective attention on a target, it increases the spatial encoding of distant distractors, thus facilitating tasks that require more global, divided attention (Rowe et al., 2007). PAE also promotes adaptation: evidence suggests that arousal modulates attentional filtering of distractions, whereas valence modulates sustained attention on task-relevant goals, suggesting that adaptation to environmental challenges can increase the availability and flexibility of neurophysiological resources to support EF (Lin et al., 2016, 2020; Chen et al., 2020). Relevant to our findings, PAE may help enhance the intervention effect on EF via two pathways that modulate interference of neurodegeneration on plasticity: (i) a top-down pathway creating alternative, specific cognitive strategies for task that require attentional resources and adaptation, and (ii) a bottom-up pathway providing general neurophysiological resources.

VAN functional integrity is related to PAE

We identified relationships between baseline VAN resting FC and PAE. Higher VAN FC was positively associated with degree of valence and arousal (i.e. higher mean values) and with valence stability (i.e. lower s.d. values). Our findings suggest that older adults with MCI and preserved VAN integrity experience more positive valence and higher arousal states, and these states do not fluctuate as much on a daily basis compared to older adults with lower VAN integrity. These results align with existing literature linking aberrant VAN FC with poorer affective regulation and stability (Kaiser et al., 2015). Weaker VAN-driven regulation allows for more salient experiences of negative scenarios, making it difficult for individuals with lower VAN integrity to disengage their attention from negative stimuli (Gotlib et al., 2004). For example, reduced within-network FC in the VAN and greater between-network FC from the VAN to other networks are associated with greater apathy in MCI and AD (Tumati et al., 2020). The disruptions consequently allow lower-order systems (e.g. limbic, subcortical) that generate negative affect to ‘highjack’ higher-order networks and weaken and/or override cortical downregulation (Tumati et al., 2020; Chen et al., 2021).

DMN functional integrity disrupts the adverse effect of neurodegeneration on plasticity in EF after training

Cognitive training predominantly focuses on the intervention effect on plasticity. Cognitive training can enhance resting within-network connectivity in older adults, and these changes have been linked to larger cognitive gains, particularly in EF (Voss et al., 2010; Chapman et al., 2015; Cao et al., 2016). Training-induced plasticity is often centered in networks vulnerable to brain aging, including the DMN (De Marco et al., 2018; Duda and Sweet, 2020; Simon et al., 2020; Bremer et al., 2022; Rahrig et al., 2022; Wu et al., 2023). However, the role of baseline functional integrity in training outcomes has received less attention. We demonstrate that baseline DMN integrity at rest disrupts the adverse effect of baseline neurodegeneration on plasticity in EF, with higher DMN integrity reducing the effect more strongly after training. These results add to a recent report that baseline resting integrity in dorsolateral pre-frontal cortex predicts working memory training gains in healthy older adults (Faraza et al., 2021). Several studies have also demonstrated that baseline resting modularity, a measure of network segregation that accounts for within-network FC, predicts training outcomes, spanning from cognitive training in healthy older adults (Gallen et al., 2016) and in adults with acquired brain injury (Arnemann et al., 2015) to aerobic exercise in healthy older adults (Baniqued et al., 2019). Our results extend a growing body of literature highlighting the importance of baseline network integrity in maintaining the capacity for adaptation and plasticity in older adults with MCI. The DMN is a focal point of structural disruptions in MCI (Khan et al., 2014; Pegueroles et al., 2017); however, preservation of select networks that are vulnerable to brain aging appears to help older adults with MCI retain the capacity for plasticity after training.

Several non-significant findings warrant further discussion. Mean arousal showed a moderating effect on baseline ADSCT and change in EF, but the effect did not survive FDR correction. Interestingly, prior work suggests that low arousal, positive affect (e.g. calm, relaxed) predicts variance above and beyond high arousal, positive affect in negative affective symptoms and well-being (McManus et al., 2019). The sample of the study had a mean age of 36 ± 10.6 years (ranging from 20 to 70 years), suggesting a predominantly younger sample without neurodegeneration. In older adults, however, high arousal PAE may disrupt the adverse effect of neurodegeneration on training-induced plasticity by improving attentional control and reducing interference of distractors during task performance (He et al., 2021), Additionally, we did not find associations between PAE and baseline cognition. Our study differs methodologically from previous cross-sectional and longitudinal work in older adults that have reported positive relationships between PAE and better recall and sustained attention (Hill et al., 2005; Carriere et al., 2010). Our measure of PAE probed the dimensions of valence and arousal rather than positive and negative affect directly. We also repeatedly probed affect over 4 weeks rather than once every few years or decades; larger changes in cognition presumably take place over longer timescales, which likely contributes to diverging results.

Our work also comes with limitations. First, practice effects may inflate cognitive performance in MCI (Machulda et al., 2013). Practice effects may also differ according to imaging biomarker status (e.g. neurodegeneration with(out) amyloidosis) (Machulda et al., 2017). Although we used different versions of the cognitive measures for each assessment point to reduce practice effects, the assessments were administered approximately 1 month apart, so some degree of cognitive plasticity may be attributed to practice effects. Next, the intervention design constrained the extent of our analyses and interpretations. We did not examine specific AD pathophysiology, as this was outside of the study design scope, making it infeasible to examine whether AD pathophysiology differentially interacts with PAE and how this might impact training gains. Additionally, mean cortical connectivity does not offer insight into topological dynamics nor does it encompass subcortical contributions. We constrained the brain analysis to within-network integrity due to the small sample size. Finally, we were unable to make inferences on neural plasticity because we focused on baseline neurodegeneration and network integrity rather than change in structural or functional connectivity. Nonetheless, we consider the network findings to be a preliminary steppingstone towards understanding how preserved baseline functional network integrity interacts with brain aging factors and their downstream effects on training-induced plasticity.

In conclusion, we provide evidence that individual differences in PAE explain why some older adults benefit more from cognitive training. Our findings indicate that PAE and preserved baseline DMN integrity reduce the adverse effect of baseline neurodegeneration on cognitive plasticity after training. These findings identify a novel source of variability in cognitive training efficacy and highlight the need to account for broader brain aging-relevant factors and their interactions with plasticity in future cognitive training development.

Supplementary data

Supplementary data is available at SCAN online.

Data availability

Imaging pre-processing scripts are available at https://github.com/adamgeorgeturnbull/BEEM. Behavioral data and analysis scripts are available at https://github.com/mmantho/projects/positive affect_cogtrain.

Author contributions

Mia Anthony (Formal analysis, Data curation, Methodology, Software, Visualization, Writing—Original draft preparation, Writing—Review & Editing), Adam Turnbull (Writing—Review & Editing), Duje Tadin (Conceptualization, Writing—Review & Editing, Supervision), F. Vankee Lin (Conceptualization, Methodology, Validation, Resources, Writing—Review & Editing, Supervision, Funding acquisition).

Funding

This work was supported by the National Institutes of Health [grant numbers MH120734-01 and U24 AG072701-02S1].

Conflict of interest

The authors declared that they had no conflict of interest with respect to their authorship or the publication of this article.