Proximal improvement and higher-order resting state network change after multidomain cognitive training intervention in healthy older adults

Abstract

Prior randomized control trials have shown that cognitive training interventions resulted in improved proximal task performance, improved functioning of activities of daily living, and reduced dementia risk in healthy older adults. Neural correlates implicated in cognitive training include hub brain regions of higher-order resting state networks including the default mode network, dorsal attention network, frontoparietal control network, and cingulo-opercular network. However, little is known about resting state network change after cognitive training, or the relation between functional brain changes and improvement in proximal task performance. We assessed the 1) change in proximal task performance, 2) change in higher-order resting state network connectivity via functional magnetic resonance imaging, and 3) association between these variables after a multidomain attention/speed-of-processing and working memory randomized control trial in a sample of 58 healthy older adults. Participants in the cognitive training group improved significantly on seven out of eight training tasks immediately after the training intervention with the largest magnitude of improvement in a divided attention/speed-of-processing task, the Double Decision task. Only the frontoparietal control network had significantly strengthened connectivity in the cognitive training group at the post-intervention timepoint. Lastly, higher frontoparietal control network connectivity was associated with improved Double Decision task performance after training in the cognitive training group. These findings show that the frontoparietal control network may strengthen after multidomain cognitive training interventions, and this network may underlie improvements in divided attention/speed-of-processing proximal improvement.

Introduction

Cognitive training is used to improve or maintain cognitive abilities that typically decline in aging, with the ultimate goal of altering the trajectory of cognitive decline and reducing dementia risk. Prior randomized clinical trials and reviews of the literature suggest that the most effective and long-lasting interventions target cognitive domains of speed-of-processing and working memory [1–3]. The most robust and reliable way to measure cognitive training impact is via proximal effects or improvements in performance of the task trained [2]. Results from a large randomized clinical trial [ACTIVE trial n =  ~ 2800] show that immediate proximal improvement in reasoning and speed-of-processing tasks lasted up to 10 years, with the greatest magnitude of improvement in speed-of-processing [3, 4]. Cognitive training interventions using speed-of-processing tasks have also resulted in reduced dementia risk and improved functional abilities in an older adult population [3, 5, 6]. Despite the promising effects of cognitive training, few studies have explored the functional brain changes that occur after cognitive training intervention, particularly in resting state networks known to associate with cognitive abilities that decline with age [7, 8].

The frontoparietal control [FPCN], cingulo-opercular [CON], dorsal attention [DAN], and default mode [DMN] networks are considered “higher-order” networks due to their association with complex cognitive functions that typically decline in aging [9–12]. The frontoparietal control [FPCN] and cingulo-opercular [CON] networks are associated with attention and executive functioning performance in aging populations [11, 12]. Higher connectivity of the FPCN and CON is also associated with better performance on a complex divided visual attention and speed-of-processing task [Useful Field of View] regularly used in cognitive training interventions [13]. The DAN is involved in tasks of visual and sustained attention and is important in orienting attention to relevant visual information [14, 15]. Lastly, the DMN is involved in self-referential processing and episodic memory, and connectivity of the DMN reliably changes in pathological aging, such as Alzheimer’s disease and dementia [16, 17].

Prior research suggests that the connectivity of these resting state networks may be impacted through cognitive training targeting executive functioning [e.g., attentional control, inhibition, and working memory] and speed-of-processing, although no study has explored this directly [8]. After a speed-of-processing cognitive training, Ross and colleagues [18] found higher resting state connectivity between brain regions included in CON and FPCN. In a study using multidomain [memory, reasoning, and problem-solving] training, Li and colleagues [19] reported increased connectivity in key areas of the DMN. Cao and colleagues [20] reported increased connectivity in other attention-/executive-based networks that overlap with the FPCN and CON, as well as the DMN. Chapman and colleagues [21] also found increased cerebral blood flow and resting state network connectivity of the DMN and a central executive network [similar to the FPCN] after training targeting innovative thinking. Findings from these studies suggest that cognitive training may induce functional brain changes specific to higher-order resting state networks. However, no study has directly assessed FPCN, CON, DAN, and DMN changes after speed-of-processing or working memory training, even though these training domains may be the most successful.

Evaluating functional brain networks associated with cognitive training interventions is important, as functional brain changes typically arise before structural changes in pathological cognitive decline [22]. Functional brain changes may also underlie improvements in cognition observed after cognitive training. For example, Lampit and colleagues [23] found functional connectivity changes after 3 weeks of multidomain cognitive training within a key region of the DMN that was spatially distant from gray matter thickness changes at 12 weeks post-training. Moreover, the degree of acute functional connectivity changes, but not structural changes, predicted subsequent cognitive improvement [23]. The results of this pilot study suggest two potentially distinct [functional and structural] mechanisms of neural change after cognitive training, with acute functional changes as a more sensitive indicator of future cognitive status. In another study, Mozolic and colleagues [24] found perfusion increases in areas overlapping with the CON after 8 weeks of attention-based cognitive training, but without gray matter structural changes. Together, these findings suggest that functional brain changes may be a better indicator of acute neural training effects, as a longer period of time may be needed for structural changes to arise. Functional brain changes after cognitive training could also be an indicator of cognitive training efficacy that could further inform studies focused on augmenting cognitive training gains through noninvasive brain stimulation [25, 26]. Additionally, exploring the link between functional brain changes and proximal improvement in training performance is a critical piece in understanding the underlying neural mechanism of cognitive training effects.

No study to date has explored higher-order resting state network change after cognitive training interventions in healthy older adults, or whether any change in network connectivity is related to improvements in proximal training performance. Therefore, we aimed to 1) confirm proximal improvement after attention/speed-of-processing and working memory cognitive training in a sample of healthy older adults; 2) explore within-network connectivity change in the FPCN, CON, DAN, and DMN after cognitive training; and 3) explore the association of resting state network change with proximal improvement on trained tasks.

Materials and methods

Participants

Participants were part of an ongoing National Institute on Aging funded, double-blinded randomized clinical trial [Augmenting Cognitive Training; ACT NCT028511] [26]. The primary goal of this clinical trial is to assess if noninvasive brain stimulation via transcranial direct current stimulation [tDCS] augments gains of cognitive training in healthy older adults. This trial uniquely consists of two phases. In phase 1, participants were randomized to a cognitive training group or an education control group to confirm efficacy of cognitive training and determine conditional power for phase 2. An equal number of participants in each phase 1 arm were also randomized to receive either active transcranial direct current stimulation [tDCS] or sham tDCS. The primary purpose of phase 1 is to assess if cognitive training is superior to an active control in proximal improvement of trained abilities, irrespective of tDCS group. However, recruitment for the ACT study is still ongoing, and therefore tDCS group allocation of phase 1 must remain blinded. All participants in the current study were part of phase 1. Therefore, tDCS group [active or sham] is not of interest and was included as a covariate in all analyses. For a detailed description of this clinical trial, please reference Woods et al. [26].

Healthy older adults were recruited at the University of Florida and the University of Arizona via local research registries, community outreach, community agencies, newspaper advertisements, public service announcements, mailings, and posted flyers. Eligibility requirements included right handedness, age range of 65–89, no history of neurological disorders [e.g., brain injury or dementia], no history of major psychiatric illness, and no contraindications to MRI. Please reference Woods and colleagues [26] for more details of inclusion/exclusion criteria. Cognitive status was screened via administration of the National Alzheimer’s Coordinating Center [NACC] Uniform Data Set [UDS-III] [27]. Participants were not eligible if they performed 1.5 standard deviations below the age, sex, and education corrected mean on a general cognitive screener or domains of memory, executive functioning, language, or visuospatial function [28]. All participants provided written informed consent, and the study was approved by the Institutional Review Boards at the University of Arizona and the University of Florida, and research was carried out in accordance with institutional guidelines and the Declaration of Helsinki.

Of the 87 participants recruited for phase 1, 17 participants did not complete the MRI for various reasons [e.g., time limits, claustrophobia], 4 participants were not able to complete the follow-up visit due to reasons unrelated to the study [e.g., moving away], 2 participants withdrew their participation, and 1 participant experienced discomfort during testing that was unrelated to the intervention. Five additional participants were excluded from analyses due to either outlier resting state network variables [± 3 standard deviations from the mean] or were < 80% adherent to training tasks. This resulted in a final sample of 30 participants in the cognitive training group and 28 participants in the education control group. The training groups did not significantly differ on any demographic variables [Table 1].

Table 1.

Sample demographics

	Cognitive training group [n = 30], mean, SD	Education control group [n = 28], mean, SD
Age	70.67, 3.99 [range 65–80]	71.11, 5.28 [range 65–84]
Sex M:F	14:16	15:13
Education years	16.37, 2.36 [range 12–21]	16.50, 2.41 [range 12–21]
MoCA	26.57, 2.01 [range 21–30]	26.50, 1.72 [range 22–29]

Note. SD standard deviation, M male, F female, MoCA Montreal cognitive assessment

Study design

At their first visit, all participants completed screening assessment measures and cognitive training measures to capture pre-intervention performance. Within 60 days of their first visit, they completed a second visit that included a functional magnetic resonance imaging scan [fMRI]. They were randomized into one of four study arms: cognitive training with active tDCS, cognitive training with sham tDCS, education control with active tDCS, or education control with sham tDCS. In total, cognitive training and education control participants completed sixty, 40-min training sessions over a span of 12 weeks, resulting in forty total hours of training. Active and sham tDCS groups received tDCS stimulation during 20 of their 60 training sessions using identical montages and stimulation parameters except for the duration of stimulation. At the end of their training [approximately 12 weeks or 3 months later], participants returned for a post-intervention follow-up that included a post fMRI scan and administration of cognitive training measures to assess post-training performance. A visual flow of the study design is depicted in Fig. 1. A detailed description of full study design can be found in Woods et al. [26].

Fig. 1.

Study timeline

Randomization procedures

Randomization was performed by the clinical trial statistician. Permuted block randomization was used with block sizes of 8 and 12 and with treatment site as stratification factor. Therefore, at each site, two participants were assigned to each one of the four conditions among the first eight participants in random order. Three participants were assigned to each one of the four conditions among the next twelve participants in random order.

Cognitive training procedures

All cognitive training was Web-based and completed via laptop computer. Four tasks targeted attention/speed-of-processing and four tasks targeted working memory from the Posit Science Brain HQ suite via its research portal. Training tasks are commercially available at www.positscience.com. All training tasks adapt for increasing difficulty unique to that task by increasing number of items asked to be remembered, shortening presentation time, or increasing the number of distractors. Participants were asked to complete four tasks per day, spending 10 min per task. During this time, participants could move up “levels” on each task. On average, it takes 15–20 levels to complete 40 min of training. When the 10-min limit was met, a timer built into the portal allowed the participant to move to the next task. Presentation of the tasks was counterbalanced and randomized so that participants were exposed to training tasks equally over the 12-week training period, with different tasks each day. Training performance was monitored for adherence, and interventionists were available for remediation strategies throughout the study to ensure participants reached their training dose. For this study, adherence was defined by completing greater than or equal to 80% of total expected levels. A list and description of training tasks, as well as their associated outcome measures, is in Table 2. Some outcome variables were log transformed for normality. Outcome measures of milliseconds represent the average presentation time of correct trails and lower scores reflect better performance [i.e., perceiving stimuli accurately at faster presentation times]. For linear units, higher scores represent better performance for that particular task [i.e., tracking more objects or remembering more speech units]. For a more detailed description of cognitive training protocol and tasks, please reference Woods et al. [26].

Table 2.

Cognitive training tasks

	Subtest description	Outcome measure
Attention/speed-of-processing
Hawk Eye	Participants must quickly identify a target object among distractors presented for varying amount of time	log10 transformed milliseconds
Divided Attention	Participants must quickly match colors, shapes, and/or fill patterns while ignoring distractor information	log2 transformed milliseconds
Target Tracker	Participants must accurately track several items moving around the screen amid distractor items	Linear units
Double Decision	Participants must correctly identify a target object in the center of the screen, while correctly locating a simultaneously presented target object in the periphery among distractors	log10 transformed milliseconds
Working Memory
To Do List	Participants must remember auditorily presented instructions	Linear units
Memory Grid	Participants must match spatially distributed cards quickly	Linear units
Auditory Ace	Participants are presented with auditory information about a playing card and must decide if the current card matches the card a specific number of cards back [auditory n-back]	Linear units
Card Shark	Participants are presented with a playing card and must decide if the current card matches the card a specified number of cards back [visual n-back]	Linear units

Education training procedures

Participants were asked to watch 40-min National Geographic Channel educational videos covering a range of topics [e.g., history, nature, wildlife]. Each video was unique to that day of training. To ensure active engagement, participants were asked and reminded at the end of each video to complete questions regarding the content of videos found in a binder provided to them. Questions were returned to intervention coordinators and served as a gauge of training adherence [greater than or equal to 80% of questions correct was considered adherent]. As with the cognitive intervention group, if necessary, remediation strategies were discussed to ensure participants met their training dose. For more details of education training protocol, see Woods et al. [26].

Image acquisition, analysis, and resting state network extraction

Participants underwent a 60-min MRI scan including structural [including T1 image] and echo-planar functional imaging at pre- and post-intervention timepoints. Resting state functional magnetic resonance imaging data were collected with a 3-Tesla Siemens Magnetom Prisma scanner with a 64-channel head coil at the Center for Cognitive Aging and Memory at the University of Florida and a 3-Tesla Siemens Magnetom Skyra scanner with a 32-channel head coil at the University of Arizona. Both study sites used identical scanning procedures and sequences, including reducing head motion through the use of foam padding, and the use of earplugs to reduce scanner noise. During the 6-min resting state functional image acquisition, participants were asked to rest with their eyes open while looking at a fixation cross. Sequence protocol: number of volumes = 120, repetition time = 3000 ms, echo time = 30 ms, flip angle = 70°, voxel size = 3.0 × 3.0 × 3.0 mm3, 44 slices, field of view = 240 × 240 mm. For normalization, a 3-min T1-weighted 3D magnetization-prepared rapid gradient-echo image was collected. Sequence protocol: TR = 1800 ms, TE = 2.26 ms, flip angle = 8°, voxel size = 1.0 × 1.0 × 1.0 mm3, 176 slices; FOV = 256 × 256 mm.

Structural and functional images were preprocessed and analyzed through MATLAB R2019b functional connectivity toolbox “CONN toolbox,” version 18b [www.nitrc.org/projects/conn, RRID:SCR_009550] and Statistical Parametric Mapping [SPM] 12 [29, 30]. The default CONN pre-processing pipeline was utilized. Outlier scans were identified through global blood oxygen–dependent [BOLD] signal and subject motion via the Artifact Rejection Toolbox [ART]. In addition to the core pre-processing steps, the CONN Toolbox applies a default denoising pipeline. Noise factors were estimated from BOLD signal and removed from each voxel using ordinary least squares regression. This pipeline implements an anatomical component-based correction procedure [aCompCor] that identifies five different noise components from white matter and cerebrospinal spaces [31]. A temporal band pass filter was applied to remove frequencies below 0.008 Hz or above 0.09 Hz to isolate slow-frequency fluctuations and reduce frequencies due to psychological factors, head motion, or random noise. Quality control plots for each individual were manually assessed. Scrubbing and outlier identification through the ART tool were included as covariates in 1st level analysis, and motion parameters were included as covariates in 2nd level analysis.

Publicly available network parcellations of the FPCN, DAN, DMN, and CON were utilized [32]. Average within-network connectivity was derived by calculating the Fisher z-transformed bivariate correlations between each region of interest BOLD time-series within each resting state network. Then, average within-network connectivity was calculated as the mean pairwise correlation of every possible region of interest combination within each resting state network. This results in one correlation variable taken as a proxy of mean within-network connectivity for a given resting state network. Figure 2, created by Hausman and colleagues [12], displays the centroid point of each region of interest included in the resting state networks.

Fig. 2.

Visualization of centroid point and connections of each resting state network. A Anterior, B superior, and C right hemisphere view [12]

Statistical analysis

The primary aim of phase 1 of the ACT study is to assess if cognitive training is superior to education training in proximal transfer, irrespective of tDCS [26]. To control for tDCS group, a blinded binary covariate of tDCS group was included in all statistical models. tDCS group was also counterbalanced across cognitive training and education control groups. Therefore, we did not anticipate that tDCS group would play a significant role in our potential findings. Study site, age, sex, education, and mood change were all included as covariates in statistical models, as these variable have been shown to affect cognitive training or resting state network variables [33–37]. Mood change was measured as the difference of pre- minus post-Beck Depression Inventory-II total scores.

All statistical analyses were completed via SPSS version 27. Mixed repeated measures analysis of covariance [ANCOVA] were performed to detect group [cognitive training versus education control] by time [pre-post-intervention] interactions of average within-network connectivity of the FPCN, CON, DAN, and DMN. Mixed repeated measures ANCOVAs were also performed to detect group [cognitive training versus education control] by time [pre-post intervention] interactions for each cognitive training measure: 4 speed-of-processing and 4 working memory tasks. Post hoc Bonferroni corrected pairwise comparisons were conducted on all ANCOVAs to explore direction of group differences.

To assess whether improvement in cognitive training measures [proximal improvement] was associated with change in resting state network connectivity, difference scores were calculated by subtracting post-intervention variables from pre-intervention variables. To avoid type 1 error due to multiple comparisons, difference scores were calculated only for resting state networks that strengthened after training and only cognitive training measures that had a high magnitude interaction, as determined by a partial eta squared larger than 1.6. While the typical cut-off for what is considered a large magnitude of effect according to partial eta squared ranges from 1.3 to 1.5, the partial eta squared term could be inflated due to partialling out error variance from the inclusion of covariates in the ANCOVA models, particularly in smaller sample sizes [38]. Thus, only considering variables at or above partial eta squared of 1.6 is a more conservative approach that avoids overinterpretation of the magnitude of interaction. Then, multiple linear regressions predicted cognitive training difference from RSN difference, controlling for covariates in the cognitive training group only.

Assumptions of normality were assessed and pre-intervention To Do List, and pre- and post-intervention Auditory Ace scores were Blom transformed for normality. Assumptions of linear regressions [i.e., normality of residuals] were confirmed.

Results

Figures 3–5 were created in RStudio via a RainCloud plot package [39].

Fig. 3.

Attention/speed-of-processing individual change and distribution of cognitive training and education control groups at pre-intervention (baseline) and post-intervention (3-month) timepoints (n = 58 for all panels). Data are unstandardized residuals controlling for tDCS group, study site, age, sex, education, and mood change. * = high magnitude of change (partial η² > 1.6). A Hawk Eye, B Divided Attention, C Target Tracker, D Double Decision. Tasks with an outcome variable of milliseconds are identified in the title of the y-axis [A, B, and D]. For interpretation, lower scores reflect better performance for these tasks

Fig. 5.

Resting state network individual change and distribution of cognitive training and education control groups at pre-intervention [baseline] and post-intervention [3-month] timepoints [n = 58 for all panels]. Data are unstandardized residuals controlling for tDCS group, study site, age, sex, education, and mood change. A Frontoparietal control network [FPCN], B cingulo-opercular network [CON], C default mode network [DMN], D dorsal attention network [DAN]

ANCOVAs of cognitive training variables

There were significant group by time interactions for all cognitive training variables, with the exception of Target Tracker, which approached significance [p = 0.054]. Additionally, pairwise comparisons confirmed that all interactions were driven by improved performance in the cognitive training group at the post-training timepoint for all variables [p < 0.001]. Five cognitive training variables had a large effect size [partial η² ≥ 1.6]: To Do List [p = 0.002, partial η² = 0.180], Memory Grid [p < 0.001, partial η² = 0.316], Auditory Ace [p < 0.001, partial η² = 0.327], Card Shark [p < 0.001, partial η² = 0.366], and Double Decision [p < 0.001, partial η² = 0.655]. These five cognitive variables were therefore carried into the regression analysis. Reference Table 3 for comprehensive output, Fig. 3 for a visual representation of speed-of-processing tasks, and Fig. 4 for a visual representation of working memory tasks.

Table 3.

Group by time interactions for cognitive training variables

	F [df = 1,50]	p	partial η2
Attention/speed-of-processing
Hawk Eye	5.353	0.025	0.097
Divided Attention	9.22	0.004	0.156
Target Tracker	3.906	0.054	0.072
Double Decision	94.757	< 0.001	0.655*
Working memory
To Do List	10.970	0.002	0.180*
Memory Grid	23.006	< 0.001	0.316*
Auditory Ace	24.283	< 0.001	0.327*
Card Shark	28.884	< 0.001	0.366*

Note. * = large effect size, df degrees of freedom

Fig. 4.

Working memory individual change and distribution of cognitive training and education control groups at pre-intervention [baseline] and post-intervention [3-month] timepoints [n = 58 for all panels]. Data are unstandardized residuals controlling for tDCS group, study site, age, sex, education, and mood change. * = high magnitude of change [partial η² > 1.6]. A To Do List, B Memory Grid, C Auditory Ace, D Card Shark

ANCOVAs of resting state networks

There were no significant group [cognitive training versus education control] by time [pre-post intervention] interactions of within-network connectivity in the CON [p = 0.300], DMN [p = 0.728], or DAN [p = 0.289], and all had small effect sizes [partial η² = 0.002-0.021]. Bonferroni corrected pairwise comparisons for the CON, DMN, and DAN did not detect any significant group differences in connectivity at the post-training timepoint. The group by time interaction for FPCN was also not significant [p = 0.102], although the effect size for this interaction was small-medium [partial η² = 0.053]. Bonferroni corrected pairwise comparisons did reveal significantly higher FPCN connectivity in the cognitive training group at the post-training timepoint compared to pre-training timepoint [p = 0.027]. Therefore, only FPCN connectivity was included in the regression analyses. Reference Table 4 for comprehensive output and Fig. 5 for a visual representation.

Table 4.

Group by time interaction for resting state networks

	F [df = 1,50]	p	partial η2
FPCN	2.733	0.102	0.053*
CON	1.098	0.3	0.021
DMN	0.123	0.728	0.002
DAN	1.149	0.289	0.022

Note. * = significant pairwise comparison of p < .05, df degrees of freedom, FPCN frontoparietal control network, CON cingulo-opercular network, DMN default mode network, DAN dorsal attention network

Association of resting state network and cognitive training variables

Change in Double Decision performance [calculated as post minus pre performance] was significantly associated with change in within-network connectivity of the FPCN [β =  − 0.444, p = 0.037]. As the output of the Double Decision task is in milliseconds, smaller values reflect better performance and negative difference values reflect an improvement in performance. Therefore, the direction of this association suggests that improved Double Decision performance after training was associated with higher FPCN connectivity after training. FPCN connectivity difference did not significantly associate with To Do list, Memory Grid, Auditory Ace, or Card Shark differences. Reference Table 5 for comprehensive output and Fig. 6 for a visual representation.

Table 5.

Association of frontoparietal control network difference with cognitive training difference

	β	t[df = 29]	p
Double Decision	− 0.444	− 2.217	0.037*
To Do List	− 0.176	− 0.945	0.355
Memory Grid	0.056	0.292	0.773
Auditory Ace	− 0.208	− 1.015	0.321
Card Shark	− 0.089	− 0.437	0.666

Note. * = p < .05, df degrees of freedom, β standardized beta

Fig. 6.

The association of frontoparietal control network difference [FPCN] and cognitive training task difference with tasks that had a high magnitude of improvement after training with 95% confidence intervals in cognitive training group only [n = 30 for all panels]. A Double Decision, B To Do List, C Memory Grid, D Auditory Ace, E Card Shark. r² reflects variance explained from the correlation between FPCN difference and cognitive training difference variables controlling for tDCS group, study site, age, sex, education, and mood change. β = standardized beta. * = significant association of p < .05

Discussion

Attention/speed-of-processing and working memory cognitive training interventions show promise in altering the trajectory of cognitive decline in older adults [2, 3, 5]. The functional brain changes that accompany improved proximal performance after training has not been thoroughly studied, although prior research suggests that there could be strengthened connectivity in the frontoparietal control, cingulo-opercular, dorsal attention, and default mode networks [8, 18–20, 40]. Therefore, this study assessed proximal improvement and resting state network changes after 12 weeks of attention/speed-of-processing and working memory computerized cognitive training, and the association between proximal improvement and resting state network change in a sample of healthy older adults.

As expected, we found that the cognitive training group significantly improved in nearly all trained tasks [proximal improvement] compared to the education control group at the post-intervention timepoint. This finding confirms and supports prior research demonstrating reliable proximal improvement after cognitive training in healthy older adults [1–4, 41]. Our findings also expand upon this body of literature. First, prior research generally reports small-to-medium effect sizes in proximal improvement, while our results show a range of medium to large effect sizes, with five out of eight tasks resulting in a large magnitude of improvement [1, 2, 41]. This suggests that the dosage, paradigm, and method of cognitive training utilized in this study is particularly effective in this sample at 3-month follow-up. Second, four of the five tasks with the greatest magnitude of improvement were working memory tasks, with the exception of the Double Decision task. Two meta-analyses of the cognitive training literature in healthy older adults found a difference in domain efficacy, in that speed-of-processing and working memory training resulted in significant improvement while simple attention training did not [1, 2]. While attention abilities decline with age, complex attentional abilities decline at a steeper rate than simple attention abilities, which remain relatively stable with age [42–45]. Therefore, our cognitively healthy and highly educated sample may not be showing large improvements in cognitive training tasks that have a large simple attention component [i.e., Hawk Eye, Divided Attention, Target Tracker] due to a ceiling effect.

The task with by far the largest magnitude of improvement was the Double Decision task, a commercialized version of the Useful Field of View [UFOV] task, that targets multiple cognitive processes at once including complex divided attention, speed-of-processing, and overall executive functioning [46]. Our results confirm and expand upon findings from the ACTIVE trial by demonstrating that the UFOV task, in a commercialized Double Decision format, again resulted in the largest magnitude of immediate proximal improvement in a multidomain training format [3]. The efficacy of the UFOV/Double Decision task is promising, as cognitive training trials utilizing this task in particular resulted in improved or maintained speed-of-processing ability, health-related quality of life, activities of daily living, and safer driving performance [3, 6, 47–49]. Importantly, individuals who underwent this training had a 29% reduced risk of dementia at a 10-year follow-up [5]. Our findings have shown that a large magnitude of improvement in this task is possible with 12 weeks of training; therefore, the neural functions related to cognitive trainings using this task will be important to explore.

While we found no statistical group by time interactions for the within-network connectivity of FPCN, CON, DMN, or DAN, pairwise comparisons did reveal significantly strengthened FPCN connectivity in the cognitive training group at the post-intervention timepoint. This is perhaps the first study to directly assess higher-order resting state network changes after 12 weeks of multidomain cognitive training in healthy older adults. Our finding of strengthened FPCN connectivity after training is supported by prior studies demonstrating increased connectivity in hub regions within the FPCN [dorsolateral prefrontal cortex [DLPFC], supplemental motor area] after executive functioning/speed-of-processing cognitive training [18, 20, 21, 50]. Additionally, one recent study demonstrated that increased brain activity in the DLPFC during brain training was related to improved speed-of-processing scores after training [51]. Our findings imply that connectivity of the FPCN can be strengthened after cognitive training, which is highly impactful since higher FPCN connectivity is observed in cognitively healthy adults over the age of 80, and increased activity in the bilateral DLPFC is related to better episodic memory in aging [52–54]. Sims and colleagues also found in a cognitively healthy oldest old [age 85 +] sample that the segregation of the FPCN contributed most to executive functioning performance, demonstrating that FPCN integrity is crucial in maintaining cognitive functions into healthy old age [55].

Interestingly, we did not detect significantly strengthened connectivity in the CON, DMN, or DAN after cognitive training despite their involvement in attention/speed-of-processing and reduced dementia risk. One explanation could be a lack of power, as a power analysis suggests a larger sample size is needed to detect at least a medium interaction [56]. Additionally, the current study’s post-intervention timepoint is relatively acute [~ 3 months], and our sample likely has high cognitive reserve due to a high level of education. It is possible that a longer delay in follow-up is needed to allow for the typical age-related decline in connectivity of these networks and subsequent divergence in network connectivity between groups. For example, the large ACTIVE clinical trial did not observe changes in secondary outcomes of activities of daily living until the 5- and 10-year follow-up timepoint [3, 4].

To explore the underlying functional brain changes that accompany proximal improvement, regressions were conducted predicting change in cognitive training tasks with a high magnitude of improvement from change in FPCN connectivity. Only the Double Decision task associated with FPCN connectivity, in that improved Double Decision performance was related to higher connectivity of the FPCN in the cognitive training group. While the frontoparietal connection is important in working memory functioning, transcranial magnetic stimulation studies have also demonstrated the importance of DLPFC function in divided attention and speed-of-processing, particularly in aging [57–61]. Prior work from our group also showed an association between FPCN connectivity with pre-intervention Double Decision task performance, and brain regions within the FPCN are recruited when performing an in scanner version of the Double Decision task [13, 62]. Therefore, this finding suggests that the Double Decision task targets the FPCN in a training format, and strengthened FPCN connectivity may underlie improved Double Decision performance after training. These results also broadly imply that improved proximal performance strengthens functional brain networks already targeted when performing the task, rather than recruiting additional functional networks to improve performance in cognitively healthy older adults. It is noteworthy that FPCN connectivity change did not associate with the working memory change measures when considering FPCN’s association with working memory tasks [59]. It may be the case that resting state functional brain changes observed after training underlie improvement on a specific training task rather than multiple tasks targeting a similar domain of cognition. Nevertheless, given the UFOV/Double Decision’s impact on improved cognitive functioning and reduced dementia risk, it stands to reason that the FPCN could also play a role in resilience of cognitive decline [5].

Limitations

The findings from this study should be interpreted within the context of a few limitations. First, this sample is predominately non-Hispanic White and highly educated. Therefore, conclusions cannot be generalized to systemically marginalized racial/ethnic groups. This is a particularly concerning gap in human aging research at large, as systemically marginalized racial/ethnic groups may be at a higher risk of pathological cognitive decline in the USA. However, disparities in access to culturally appropriate medical care, loss of trust in the healthcare system, social determinants of health, and racial biases in cognitive screening tools also contribute to pathological risk [63–67]. There is also evidence suggesting experiencing acute and prolonged racial discrimination alters patterns of functional connectivity across the lifespan [68–70]. Future studies need to include individuals that are representative of the US population and are at increased risk for subsequent cognitive decline and dementia by striving toward more equitable recruitment and research practices.

As evidenced by visual depictions of our findings [Figs. 3–5], there is a considerable amount of individual variability in performance. In particular, resting state functional networks show poorer test re-test reliability over time in older adults compared to younger adults [71]. A larger sample size is needed to not only boost power but possibly increase the robustness of our findings by reducing error due to variability.

Lastly, due to this sample being drawn from an ongoing clinical trial, a portion of individuals in both the cognitive training and education control group had received active tDCS. While active tDCS group assignment was counterbalanced across cognitive training and education control groups, and was included as a covariate, we are unable to definitively conclude that this had no impact on our current findings.

Summary and future directions

To summarize, this study is the first to explore higher-order resting state network changes after attention/speed-of-processing and working memory cognitive training in healthy older adults and to explore the association of brain network change and proximal improvement. Findings from this study highlight the efficacy of cognitive training to produce a large magnitude of proximal improvement in the Double Decision task over 12 weeks of intervention. Our results also suggest that increased within-network resting state connectivity of the FPCN is possible after cognitive training and that improved FPCN connectivity may underlie improvement in the Double Decision task. These findings advance our understanding of the functional brain changes that underlie cognitive training and add to the growing body of literature focused on improving cognitive training efficacy.

The impactful findings from this study give rise to multiple directions for future research. First, future research should explore the longitudinal trajectory of proximal improvement and functional brain network change after cognitive training ends. Through assessing multiple timepoints over sequential years, the impact of cognitive training and FPCN connectivity on cognitive decline and dementia risk and prevalence can be tracked.

Additionally, emerging research suggests that between-network connectivity is an important component in cognitive performance and pathological cognitive decline, particularly with the DMN and other higher-order resting state networks [72–74]. Prior research has also demonstrated the role of the FPCN in internetwork modulation, which then in turn impacts cognitive functioning [50, 75, 76]. To further understand resting state network change after cognitive training, subsequent studies might consider exploring between-network connectivity of these four higher-order resting state networks. A key brain region included in the FPCN is the dorsolateral prefrontal cortex. This region is also a common brain area to target when applying noninvasive brain stimulation [tDCS] to boost cognitive training gains [25]. Therefore, it would be important to also consider how the application of noninvasive brain stimulation may also boost connectivity of the FPCN.

Lastly, it is also important to consider other neural changes that could underlie proximal improvement but were not measured in this study. The hippocampus and the prefrontal cortex are two brain structures that are most vulnerable to structural and functional changes in aging, and both structures show alterations in synaptic plasticity or long-term potentiation [LTP] that underlie declining memory and cognitive functioning in aging rodents [77, 78]. Moreover, LTP-induced network reorganization enhances the long-range connections between the hippocampus and the prefrontal cortex, and cognitive training in older adults may alter the connectivity of the hippocampus to other brain areas, including the prefrontal cortex [79, 80]. The prefrontal cortex is also a key structure involved in episodic memory encoding and retrieval in aging [81]. The relationship between synaptic plasticity and the functional connectivity of neural networks, like the FPCN, is not well understood in aging humans, although prior research suggests that long-term potentiation could underlie strengthened connectivity of brain networks [82]. Future research could study the impact that cognitive training has on LTP in the aging brain and specifically within the hippocampal/prefrontal cortex connection.

Author contribution

CH wrote the first draft of the manuscript under the mentorship of Adam J. Woods, and all the authors commented on and approved the previous versions. Material preparation and data collection were performed by JK, AA, HH, NE, EB, AO, EVE, PB, HS, and SS. CH, JK, and HH contributed to data analysis and processing. CH, AW, EP, SD, GH, SW, MM, RC, and GE contributed to study conception and design. KL contributed to manuscript editing. All the authors read and approved the final manuscript.

Funding

This work was supported by the National Institute on Aging [NIA R01AG054077, NIA K01AG050707, NIA P30AG019610, T32AG020499], the State of Arizona and Arizona Department of Health Services [ADHS], the University of Florida Center for Cognitive Aging and Memory Clinical Translational Research, the McKnight Brain Research Foundation, and National Heart, Lung, and Blood Institute [T32HL134621].

Data availability

The data analyzed in this study is subject to the following licenses/restrictions: data are managed under the data sharing agreement established with NIA and the parent R01 clinical trial Data Safety and Monitoring Board in the context of an ongoing phase III clinical trial [ACT study, R01AG054077]. All trial data will be made publicly available 2 years after completion of the parent clinical trial, per NIA and DSMB agreement. Requests for baseline data can be submitted to the ACT Publication and Presentation [P&P] Committee and will require submission of a data use, authorship, and analytic plan for review by the P&P committee [ajwoods@phhp.ufl.edu]. Requests to access these datasets should be directed to ajwoods@ufl.edu.

Declarations

Ethics approval

The studies involving human participants were reviewed and approved by the University of Florida Institutional Review Board and University of Arizona Institutional Review Board. The patients/participants provided their written informed consent to participate in this study.

Consent to participate

Freely given, informed consent to participate in the study was obtained from all study participants.

Consent for publication

The participants provided informed consent regarding publishing their data.

Conflict of interest

The authors declare no competing interests.