Long-term cognitive training enhances fluid cognition and brain connectivity in individuals with MCI

Abstract

Amnestic mild cognitive impairment (aMCI) is a risk factor for Alzheimer’s disease (AD). Multi-domain cognitive training (CT) may slow cognitive decline and delay AD onset. However, most work involves short interventions, targeting single cognitive domains or lacking active controls. We conducted a single-blind randomized controlled trial to investigate the effect of a 6-month, multi-domain CT on Fluid Cognition, functional connectivity in memory and executive functioning networks (primary outcomes), and white matter microstructural properties (secondary outcome) in aMCI. Sixty participants were randomly assigned to either a multi-domain CT or crossword training (CW) group, and thirty-four participants completed the intervention. We found a significant group-by-time interaction in Fluid Cognition (p = 0.007, F (1,28) = 8.26, Cohen’s d = 0.38, 95% confidence interval [CI]: 2.45–14.4), with 90% of CT patients showing post-intervention improvements (p < 0.01, Cohen’s d = 0.7). The CT group also showed better post-intervention Fluid Cognition than healthy controls (HCs, N = 45, p = 0.045). Functional connectivity analyses showed a significant group-by-time interaction (Cohen’s d ≥ 0.8) in the dorsolateral prefrontal cortex (DLPFC) and inferior parietal cortex (IPC) networks. Specifically, CT displayed post-intervention increases whereas CW displayed decreases in functional connectivity. Moreover, increased connectivity strength between the left DLPFC and medial PFC was associated with improved Fluid Cognition. At a microstructural level, we observed a decline in fiber density (FD) for both groups, but the CT group declined less steeply (1.3 vs. 2%). The slower decline in FD for the CT group in several tracts, including the cingulum-hippocampus tract, was associated with better working memory. Finally, we identified regions in cognitive control and memory networks for which baseline functional connectivity and microstructural properties were associated with changes in Fluid Cognition. Long-term, multi-domain CT improves cognitive functioning and functional connectivity and delays structural brain decline in aMCI (ClinicalTrials.gov number: NCT03883308).

Introduction

Cognitive training (CT) has received increased attention due to its potential in delaying the onset of Alzheimer’s disease (AD) [1–7]. AD is the most common type of dementia and is characterized by a progressive deterioration of cognitive functioning leading to an overall lack of independence [8]. There is currently no cure for AD, but even an onset delay of 1 to 5 years can reduce formal and informal healthcare costs and improve quality of life for patients [9, 10]. The appeal of CT as it relates to AD is that it is a non-pharmacological intervention that has shown effects in cognition and across brain markers, identified by structural and functional magnetic resonance imaging (MRI) or positron emission tomography (PET) [11–13].

In CT studies, participants typically train one or more cognitive functions for a given period of time, with the goal of increasing general cognitive ability. Extant data show that CT has positive effects on cognition in individuals with Mild Cognitive Impairment (MCI) who are at a greater risk of developing AD [14–17]. Still, there is limited evidence regarding specificity and the overall effects of interventions on large-scale brain dynamics with most research focusing on healthy older adults instead [18–23]. CT reports in individuals with MCI and AD have shown increased hippocampal activity [13, 24] and default mode network (DMN) connectivity [25] as well as widespread structural and functional changes in memory- and learning-relevant regions, including the frontoparietal control network (FPCN). This indicates that, even at later stages, brain plasticity is present and can be manipulated [12, 24, 26–34]. Although these studies are valuable in understanding brain alterations concomitant with CT, they often lack an examination of whole-brain interactions and do not investigate these changes in relation to performance. Previous research has also typically conducted single-domain CT, with a focus on episodic and working memory. This is well rationalized, given that one of the major symptoms of AD is memory loss, but patients suffer from a deterioration in several other domains including attention and executive control processes.

The prefrontal cortex (PFC), which projects to the medial temporal lobe, is affected in AD, and of particular importance for most higher-order cognitive functions. Multi-modal interventions tax the PFC and appear superior in improving cognitive and brain functioning [35]. They can also culminate some of the heterogeneity in results by combining different types of cognitive abilities into the same regimen. Finally, CT can differ extensively between studies, with most having short intervention periods (i.e., 3 months or less) and involving passive control groups, leaving a scattered view of methodology and findings [6, 17, 36–38]. Moreover, discrepancy in previous research can be attributable to the use of rudimentary measurements. White-matter integrity based on conventional diffusion tensor imaging is incompetent in representing complex fiber configurations and has limited biological interpretation [39]. Fixel-based analysis (FBA) overcomes many of these limitations as it can adequately give estimates even in crossing fibers, has increased sensitivity to microstructural abnormalities in neurites (axons and dendrites), and shows excellent test-retest reliability [40]. In this context, a specific fiber population within a voxel is referred to as a “fixel”. This enables the quantification of fiber-specific metrics for assessing the properties and changes of white matter (WM). Fixels are derived from WM fiber orientation distributions (FODs) computed via Constrained Spherical Deconvolution (CSD). Compared to voxels, fixels are more directly related to WM anatomy. Fiber density (FD) is a fixel-wise metric that reflects white matter microstructure. Its value is proportional to the total intra-axonal volume and can be directly computed from FODs. Additionally, individual subject warps to a common template space can be used to measure macroscopic differences in fiber-bundle cross-section (FC). A fixel-wise analysis combining fiber density and cross-section (FDC) can also be computed. Our study is the first to utilize FBA to link measures of white-matter neurite microstructure to CT, which allows us to detect longitudinal changes from baseline to follow-up.

In this clinical trial, we addressed the aforementioned gaps by investigating the effects of a long-term (6 months) multi-domain computerized CT in a group of individuals with amnestic MCI (aMCI), making it one of the longest-duration multi-domain CTs in the literature. Individuals in the CT group received at-home computerized training involving memory and executive functioning throughout the 6-month period. We also included an active control group who were engaged in crossword puzzles training for the same length of time. Our primary outcomes were Fluid Cognition derived from National Institute of Health (NIH) Toolbox Cognition Battery [41], and functional connectivity in memory and executive functioning networks. Fluid Cognition is comprised of tests of executive functioning, attention and inhibitory control, episodic and working memory, and processing speed. These domains are needed to adequately process and integrate information and solve problems.

Overall, we hypothesized that patients in the CT group would improve on measures of Fluid Cognition and functional and microstructural connectivity, with the CT group showing increased functional connectivity in the FPCN and improved microstructural connectivity in the cingulate-temporal-frontal tracts subserving memory and executive functioning.

Materials and methods

Participants

Sixty aMCI participants (age 65–85 years, 30 CT and 30 CW) and 45 healthy controls (HCs, age 65–85 years) were recruited for the study. The aMCI participants were randomized to the cognitive training (CT) or crossword puzzle training control group (CW) using a stratified block randomization model. The factors used for stratification were age (65–70, 71–75, >75 years old), gender (Male or Female), and education (0–7, 8–15, >15 years). We monitored potential imbalance through data collection and selected participants on the imbalanced demographics until the groups were matched. The study was single-blind, and the codes were unmasked only to the PI. The HC participants underwent assessment solely at baseline, and their data served as a reference for comparison. All the study coordinators and test administrators were kept blind to group assignments. Participants were informed that they would be randomly assigned to either a structured computerized CT or nonspecific computerized training activities (i.e., CW). Due to the COVID-19 pandemic, some participants did not complete the intervention, and some could not return for the post-intervention visits (Fig. 1). Overall, thirty-four aMCI patients (age 65–85 years, 20 CT and 14 CW) successfully completed the intervention. The study is registered as a clinical trial on ClinicalTrivals.gov (identifier: NCT03883308).

Fig. 1.

CONSORT diagram.

The aMCI participants and healthy controls underwent baseline cognitive, neuropsychological, and neuroimaging data collection. Written informed consent was obtained from all participants who fit the inclusion criteria, and the Stanford Institutional Review Board approved the protocol. Cognitive and behavioral screening assessments were administered to ensure eligibility. Inclusion criteria for aMCI participants included the following: (1) age ≥ 65 and ≤85 years old, (2) diagnosis of aMCI in the past year: cognitive concerns by subject, informant, or physician, impairment in memory domain (delayed recall of one paragraph from the Logical Memory II (LM II) subscale from the Wechsler Memory Scale-Revised with cutoff scores of ≤8, ≤4 and ≤2 for 16, 8–15, and 0–7 years of education respectively), (3) absence of dementia, (4) essentially normal functional activities (intact IADL), (5) Clinical Dementia Rating (CDR) of 0.5 (Memory Box score of at least 0.5), (6) Mini-Mental State Examination (MMSE) scores ≥24, (7) stability of permitted medications (e.g., cholinesterase inhibitors, hypertension medication, etc.) for at least two months, and (8) no Axis I disorder as assessed by the Mini International Neuropsychiatric Interview (M.I.N.I). Participants were excluded from the study if they had the presence of suicidality, current regular use of psychiatric medications, opiates, or thyroid medications, claustrophobia, MRI contraindications, present substance abuse, post-traumatic or psychotic disorders, bipolar disorder, any significant neurologic disease, including possible and probable dementia, multi-infarct dementia, Parkinson’s or Huntington’s disease, brain tumor(s), progressive supranuclear palsy, a seizure disorder, subdural hematoma, multiple sclerosis, “uncontrolled” hypertension, history of significant head trauma, history of alcohol or substance abuse or dependence within the past 2 years, or any significant systemic or unstable medical condition which could lead to difficulty in complying with the training protocol. The Behavior Rating Inventory of Executive Function (BRIEF) was used to assess real-world executive function abilities [42]. The cognitive measures, Fluid Cognition (primary outcome) and Crystallized Cognition, were measured using the NIH Toolbox Cognition battery [43] (nihtoolbox.org) (see more details in Supplementary Methods). aMCI patients were invited for a follow-up appointment to complete cognitive, neuropsychological, and neuroimaging sessions approximately six months following their initial visit. Thirty-four aMCI patients completed cognitive assessments for sessions one and two, while twenty-four participants (15 CT and 9 CW) completed both MRI and cognitive assessments for both sessions.

Multi-domain cognitive training program

An online computerized training curriculum was curated from a set of CT games available through the Lumosity platform (Lumos Labs, Inc.) and administered using the participant’s home computer. The exercises were designed to train and practice memory and executive functions. Particularly, training tasks were comprised of games of switching, visual and spatial working memory, and face and spatial memory recall. Each session consisted of 6 games. For details regarding tasks employed and cognitive domain(s) targeted, see Supplementary Methods. Exercises were adaptive to individual ability, increasing in difficulty as participants progressed. The program provided immediate feedback and reinforcement on performance. The designed training curriculum included three sessions per week—each 20–30 min in duration—for 20 weeks. Participants were required to login to their individual online account and complete five exercises in each session. Each time the participant logged in, the program automatically delivered the exercises assigned for that session. The curriculum and schedule were hard coded into the program by Lumos Labs, Inc. and were the same for each participant. Exercises began with the option to start, or view written and animated instructions. General instructions for completing the CT program were provided in writing and online to participants. Participants’ progress was tracked weekly through a log file provided by Lumos Labs and they were notified if they were on track. The authors do not have any financial relationships with Lumos Labs and have no other conflicts of interest related to the program.

The Lumosity Crossword Control Platform, which is specifically designed to serve as an active control platform for research studies, was used to provide crossword puzzle training for the control group (CW). Unlike the CT, the crossword platform was non-adaptive such that all available puzzles were the same difficulty (i.e., medium). Procedures were the same for both groups, except those in the CW group were asked to complete a variety of crossword puzzles available through this online platform with the same curriculum assigned to the treatment group (three sessions per week for 20 weeks).

Neuroimaging data acquisitions and analyses

Participants underwent a Magnetic Resonance Imaging (MRI) scan pre-and post-intervention. All MRI data were acquired on a 3T GE system (General Electric Healthcare, Milwaukee, WI, USA) equipped with a 32-channel head coil (Nova Medical, Wilmington, MA, USA) at the Center for Cognitive and Neurobiological Imaging at Stanford University (http://www.cni.stanford.edu/). T1-weighted images were collected using MPRAGE pulse sequence with 450 ms inversion time (TI), flip angle = 12°, and 1 mm slice thickness.

Multi-band task functional MRI (fMRI, 8 min) scans were collected using a multi-band acceleration factor = 6, TR = 0.710 s, TE = 35 s, flip angle = 54°, slice number = 60, and resolution = 1 mm isotropic while participants performed a delayed match-to-sample working memory task (Fig. S3). In each set of trials, a series of lower case and uppercase letters (total of six letters) were sequentially presented (encoding phase), and participants were asked to memorize the lower-case letters. After the encoding phase, a 3 s fixation cross was displayed (delay phase). After the delay, a target letter was presented in red, and participants had 3 s to respond using a button box in their right hand if the target letter had been presented during the encoding phase. For details on the in-scanner task, see Fig. S1.

Multi-shell diffusion MRI (dMRI) were acquired using a multiband echo-planar imaging (EPI) acquisition scheme (multiband factor of 3) with isotropic 2.0 mm³ spatial resolution in 80 diffusion directions with diffusion gradient strength set to b = 2855 s/mm² and 30 diffusion directions with diffusion gradient strength b = 710 s/mm². Each dMRI scan also contained nine images without diffusion weighting (b = 0 s/mm²). An additional scan was acquired in the opposite phase encoding direction consisting of 6 diffusion directions (b = 2855 s/mm²) and two non-diffusion-weighted images for EPI distortion correction. Other dMRI parameters were TR/TE = 2800/78 ms, matrix size = 112 × 112, and 63 axial slices.

The fMRI and dMRI data were preprocessed using a combination of well-known preprocessing software tools and entered for further processing (for detailed procedures, see Supplementary Methods) [44–46].

Statistical analyses

Primary outcomes

Changes in Fluid scores over time were assessed using linear mixed effects models (LME; lme4 R package), including group (CT and CW) and time (pre- and post-intervention) as fixed factors and sex as a covariate. All participants’ data were included in the mixed effect models following intent-to-treat principle which includes total 54 subjects (32 CT and 22 CW) with Fluid Cognition. Within-group changes in the outcomes were examined using paired samples t-tests. Further, whole-brain functional connectivity analyses were carried out in the CONN toolbox using a two-way mixed ANOVA between-group (CT and CW) and time (pre-and post-intervention) (alpha = 0.05, corrected for multiple comparisons using false discovery rate (FDR)) adjusting for age and sex.

Secondary outcomes

The FBA analysis was performed using the CFE (part of MRtrix3 software) method that provides a permutation-based, family-wise-error (FWE) corrected p-value for every fixel (specific fiber population within a voxel) in the template image. A longitudinal design matrix was used to test the change in FBA metrics between groups over time, including age and sex as covariates.

Exploratory data analyses

Crystalized Cognition and BRIEF scores over time were assessed using linear mixed-effects models (LME; lme4 R package), including group (CT and CW) and time (pre- and post-intervention) as fixed factors and sex as a covariate. BRIEF outputs two summary index scales of everyday executive functioning abilities including Behavioral Regulation Index (BRI) and Meta-cognition Index (MI) that were entered into the analyses. Crystalized Cognition was used as a control for the primary outcome. The relationships between change in Fluid Cognition, functional connectivity, and structural connectivity based on intra-cellular volume fraction (ICVF) and orientation dispersion index (ODI) metrics were tested using multiple linear regression models. To investigate if the performance of CT individuals pos-intervention reached similar levels to that of HCs, we performed two-sample t-tests or general linear models for our main outcomes. It is important to acknowledge that the selection of statistical methods was driven by their availability within each imaging toolbox and their suitability for the specific needs of longitudinal studies.

Results

Demographics and clinical characteristics

Descriptive statistics and neurocognitive scores for baseline and post-intervention are summarized in Table 1. Baseline and post-intervention data showed that age, sex, MMSE, Fluid and Crystallized Cognition, and BRIEF scores were not significantly different between groups at either session. Logical Memory II scores were different between groups at baseline (p = 0.043, uncorrected) but not in the post-session.

Table 1.

Characteristics of the aMCI participants (n = 34) and HC (n = 45).

	HC (N = 45)	CT (N = 20)		CW (N = 14)		Statistics (CT vs. CW)
		Baseline	Post	Baseline	Post
Age, mean (SD)	73.2 (5.8)	73.7 (5.71)	74.4 (5.66)	75.8 (6.06)	76.4 (5.8)	$p_{\mathit{base}}=0.28$
Sex, female	32	10		9		$p=0.21$
MMSE (SD)	28.97 (1.05)	27.75 (1.6)	28.25 (1.3)	27.78 (1.9)	28.25 (1.44)	$p_{\mathit{base}}=0.57$
Logical Memory II (SD)	11.4 (3.05)	6.97 (2.58)	9.87 (2.91)	5.53 (2.41)	8.71 (3.6)	$p_{\mathit{base}}=0.043$
Fluid Cognition (SD)	101.5 (14.7)	97.35 (14.8)	107.94 (14.3)	99.07 (19.5)	102.6 (22.6)	$p_{\mathit{base}}=0.91$
Crystallized Cognition (SD)	_	115.2 (11.1)	114.7 (13.2)	116.9 (6.88)	116.2 (8.83)	$p_{\mathit{base}}=0.74$
BRIEF (BRI) (SD)	_	53.95 (10.4)	49.61 (8.9)	53.85 (9.7)	53.3 (8.7)	$p_{\mathit{base}}=0.8$
BRIEF (MI) (SD)	_	58.65 (15.0)	54.27 (10.4)	56.42 (10.3)	57.92 (9.9)	$p_{\mathit{base}}=0.83$

Primary outcomes

Changes in neurocognitive functioning

The neuropsychological test scores at baseline and post-intervention are reported in Table 1. Using LME models, we observed a significant group-by-time interaction in age-corrected standard Fluid Cognition (p = 0.007, F (1,28) = 8.26, Cohen's d = 0.38) (Fig. 2a). Paired t-test analysis revealed significant improvement in age-corrected standard Fluid Cognition in the CT group (10.9% improvement from baseline to post-intervention, mean (SD) at baseline = 97.35 (14.8), and mean (SD) at post = 107.94 (14.3) with p < 0.01, t = 5.94, df = 17, Cohen's d = 0.70), with 18 (out of 20) participants showing improvement. One participant remained stable, and one showed decline. Also, the two-sample t-test revealed that after intervention, the CT group showed slightly better performance than HCs in age-corrected Fluid Cognition (p = 0.045). The CW group also showed slight improvement in Fluid Cognition (1.64% change post-session vs. baseline, mean (SD) at baseline = 99.07 (19.5), and mean (SD) at post = 102.6 (22.6)) but it was not significant (p = 0.24, t = 1.22, df = 12), with 9 (out of 14 participants) improving and 5 showing decline (Fig. 2a).

Fig. 2. Changes in cognitive functioning and functional connectivity in response to intervention.

a Cognitive Fluid Intelligence for each group and time point. The error bars represent the mean value and standard error. b Functional brain connections between the left DLPFC (Area 46), right DLPFC (Area 8BL) and medial PFC (Area s6-8), and between the left intraparietal sulcus (Area IP0) and right anterior PFC (Area 9-46v) showing a significant group-by-time interaction effect (${\mathrm{p}}_{\mathrm{FDR}}<0.05$). c Association between change in Fluid Cognition standard scores and connectivity strength between the left DLPFC (Area 46) and right medial PFC (Area s6-8) across the groups. d Box plots of functional connectivity strengths between the left DLPFC (Area 46), right DLPFC (Area 8BL), and right medial PFC (Area s6-8), and between the left intraparietal sulcus (Area IP0) and right anterior PFC (Area 9-46v) for each group across timepoints. The line between dark green and light green states the median value of connectivity strength in each group.

Functional connectivity changes in the executive function network and associations to fluid cognition

Using repeated measures ANOVA, we analyzed seed-based functional connectivity during the encoding phase of the in-scanner working memory task, which revealed significant group-by-time interaction effects for the connectivity between the DLPFC (Area 46) and medial PFC (Area s6-8) and right DLPFC (Lateral Area 8B), and between the left intraparietal sulcus (Area IP0) and right anterior PFC (Area 9-46v) (p_FDR<0.05, Cohen’s d > 0.86) (Fig. 2b). Particularly, the CT group showed increased connectivity while the CW group showed decreased connectivity between the left DLPFC (Area 46) and medial PFC (Area s6-8) (mean connectivity weights (baseline, post) = –0.023, 0.025) and the right DLPFC (Area 8BL) (mean connectivity weights (baseline, post) = –0.046, 0.029), and between the left intraparietal sulcus (Area IP0) and right anterior PFC (Area 9-46v) (mean connectivity weights (baseline, post) = 0.003, 0.038) (Fig. 2d). Further, increased connectivity strength between the left DLPFC (Area 46) and right medial PFC (Area S6-8) from baseline to post-intervention was significantly correlated with improvement in Fluid Cognition scores ($p_{uncor}=0.03,t=2.2$; Fig. 2c) across groups. It should be noted that functional connectivity in these networks was not significantly different between CT and CW at baseline. These findings indicate that CT resulted in increased functional connectivity in parts of the FCN which further correlated with higher Fluid Cognition.

Finally, we explored how the mean functional connectivity between the left DLPFC (Area 46), the right DLPFC (Area 8BL), and medial PFC (Area s6-8), and between the left intraparietal sulcus (Area IP0) and right anterior PFC (Area 9-46v) differed in HCs compared to the CT post-session participants. This was done to assess if the intervention resulted in normalization of brain activity in the CT group. While the CT group showed lower mean connectivity strength than the HC group in both sessions (p < 0.01) (Fig. S2), the difference in mean connectivity became smaller at post-intervention, suggesting a potential rehabilitation of functional connectivity in the CT group.

Secondary outcomes

Changes in crystallized cognition and BRIEF

As expected, we did not observe improvement in Crystallized Cognition within CT or CW (p = 0.69). The BRIEF BRI scores were significantly decreased post-intervention in the CT group (p = 0.04, $Cohen’sd=0.4$), while the CW group showed no change (p = 0.72). Given that lower BRIEF scores are indicative of higher executive functioning in daily activities, these results indicate improved real-world abilities in behavioral regulation. BRIEF MI scores were not significantly different between pre- and post-intervention (p = 0.08).

Changes in white-matter microstructure

Whole-brain fixel-based analysis, associations with cognition and tract-based analyses

The two-way analysis of variance revealed a significant group-by-time interaction in white-matter FD (p_FWE<0.05; largest Cohen’s d = 0.36) for the left and right cingulum hippocampus, anterior thalamic radiation, and optic radiation (Fig. 3a). Over the six-month period, FD in these tracts decreased for both CT and CW, but the reduction was significantly stronger in the CT group (p < 0.05). Specifically, the CT group showed a 1.3% decrease in FD while the CW group had a 2% decrease in the significant fixels post-intervention (Fig. 3b). Notably, mean FD values at baseline were not significantly different between CT and CW and main effects of group and time were also not significant. The main and interaction effects were not significant for FC or combined FD × FC measures. While the association between change in FD and Fluid Cognition was also not significant within and across groups, exploratory analysis revealed that increased mean FD in left and right cingulum hippocampus, anterior thalamic radiation, and optic radiation (as shown in Fig. 3a) was associated with improved performance in NIH Toolbox List Sorting Working Memory (LSWM) standard scores ($p_{uncor}=0.004,t=4.3$) for both the CT group and across groups ($p_{\mathit{uncor}}=0.01,\mathit{t}=2.7$; Fig. 2c).

Fig. 3. Changes in white matter microstructure in response to intervention.

a Fixels showing a significant group-by-time interaction effect projected onto the group white matter FOD template. Significant regions are represented as fixels color-coded by FWE-corrected p-value. b Mean FD changes for each group and time point. The line between dark green and light green states the median value of the mean FD changes in each group. c Association between changes in FD and changes in List Sorting Working Memory (LSWM) standard scores across the groups.

Finally, we also performed a tract-of-interest analysis to investigate the potential effect of the multi-domain computerized CT on hypothesized white-matter pathways and found a significant group-by-time interaction in the rate of change in FD in the left cingulum frontal parahippocampal and parietal tracts (p_FDR < 0.05) (for details see Tract-based analysis, Supplementary Results).

Associations between baseline functional and microstructural connectivity and changes in Fluid Cognition

To explore whether specific brain connectivity patterns at baseline could predict changes in cognition, we examined correlations between baseline whole-brain functional and microstructural connectivity with changes in Fluid Cognition composite scores within the CT group (N = 20). We found distributed functional and structural networks that showed a significant positive correlation with longitudinal change in Fluid Cognition (Fig. S3) (see Supplementary Results).

Discussion

Our randomized controlled trial showed that, in aMCI individuals, the proposed 6-months multi-domain CT resulted in significant improvement in Fluid Cognition, increases in functional connectivity in parts of the FPCN, and reduced decline in white matter microstructural properties in memory and executive function networks. Effect sizes ranged from medium to large for Fluid Cognition and functional connectivity and small to medium for microstructure. Most importantly, the observed changes in functional and microstructural brain properties were associated with enhanced cognitive functioning within the CT group. Conversely, we investigated if there were differences in Crystallized Cognition as a sanity check and, as hypothesized, this domain remained stable post-intervention which further supports our findings [47, 48]. We were also able to identify key baseline functional and structural subnetworks that predicted change in Fluid Cognition for those engaged in CT.

Fluid Cognition, our primary outcome, improved by ~10% in the CT group, with an effect size of 0.7. Specifically, 90% (18 out of 20) of CT participants showed improvements, 5% remained stable, and 5% showed decline. This increase was also larger in CT compared with CW – our active control group. Noteworthy, the performance of the CT group post-intervention surpassed that observed in the reference HC group. In addition, BRIEF scores indicated that CT resulted in improved behavioral regulation in everyday life post-intervention, whereas such an effect was not observed for CW. This finding suggests that the proposed multi-domain CT might be transferable to improved skills in general daily activities, including set shifting, inhibition, self-monitoring, and emotional control.

Functional connections that showed significant group-by-time interactions were mainly located in the DLPFC and intraparietal sulcus, regions known to tax working memory and executive functioning [49–52]. The DLPFC is also critical for fluid intelligence and the fact that functional connectivity following CT correlated with improved Fluid Cognition further corroborates this idea [51]. Additionally, for most of these connections, coupling during task performance increased from pre- to post-intervention in the CT group but decreased in the CW group. Both the DLPFC and intraparietal sulcus belong to the FPCN [53], which has been established as playing an intermediary role and easily updating its connectivity to other regions in a task-dependent manner [54–56]. Connectivity in this network is affected in aging and MCI [57–59] and can signal compensation or dedifferentiation depending on whether it is linked to better or worse cognition [60, 61]. The proposed CT engaged a variety of higher-order cognitive processes including task shifting and inhibition, memory, processing speed, and ability to flexibly update short-term memory. Functional connectivity within and between the FPCN, which regulates these tasks, was adjusted and optimized throughout the CT. Even though the CT group showed lower connectivity compared to the reference HC group, the difference in mean connectivity between groups decreased at post-intervention. This suggest that there is a potential rehabilitatory effect of intervention on functional brain networks in CT, resulting in normalization of mean connectivity strength in the indicated networks. Our data suggest that multi-domain CT improved how the FPCN was able to realign and shift according to task demands. In addition, parietal-frontal circuits facilitate working memory by creating short-term representations that allow information manipulation over a short time. These circuits were likely trained during the 6-months intervention and contributed to improvements in cognitive control.

Microstructural white-matter properties also showed significant group-by-time interaction effects. Fixel-based analyses revealed that FD in the right cingulum hippocampus, anterior thalamic radiation, and optic radiation decreased in both groups. This decrease was significantly smaller in CT compared to CW, suggesting that the cognitive intervention helped slow down FD decline. Recent data from our group and others suggest that a decline in white-matter integrity in these tracts is associated with aMCI [62, 63]. Particularly, the cingulum hippocampus and anterior thalamic radiation connect frontal regions (namely the PFC) to the hippocampus and thalamus, respectively, and contribute to a variety of cognitive functions including executive functioning and memory [64–66]. We found that participants in the CT group with the least decline in FD showed the largest improvement in working memory. This is consistent with our tract-of-interest analysis findings, where FD along the left cingulum frontal parahippocampal and parietal tracts had downward trends in both groups but this was, again, less pronounced for CT. Older adults with low working memory capacity seem to have reduced white-matter integrity in frontal regions [67, 68]. This is in line with work showing that the cingulum frontal parahippocampal tract is involved in working memory processes [69–71] and that the cingulum - which involves limbic white-matter bundles with connections to and from the hippocampus - is among the earliest pathways affected in AD [72]. Unlike functional connectivity, FD did not improve; this can be explained by the fact that all subjects have aMCI and, therefore, could be in a more advanced stage of neurodegeneration. Additionally, functional brain patterns rely on BOLD dynamics which are more flexible and amenable to change compared to structure. The slower microstructure decline in the CT group is itself of importance as it suggests long-term, multi-domain CT may delay neurodegeneration in aMCI.

BOLD findings regarding training studies have been inconsistent and varied depending on type of intervention [73]. Still, increases following CT are well reported in the literature [74], whereas its effects on the control group are less consensual. The transition from MCI to AD can be accompanied by multifaceted alterations in cerebral blood flow, sometimes displayed by a loss of functional connectivity in both MCI and AD [75]. As such, it appears plausible that an increase of connectivity in the CT occurred as a consequence of training, while the decrease in the CW group followed the natural trajectory that might be expected in individuals who are on the course to develop AD dementia.

Baseline brain connectivity patterns were good predictors of change in Fluid Cognition. These factors might help differentiate who is better suited or more likely to benefit from CT [76]. Regions with the strongest correlations displayed overlap across functional and microstructural (FD and fiber complexity) networks. Of note are networks connecting middle and visual areas, anterior cingulate and medial PFC, and inferior parietal cortex, for which links to Fluid Cognition were analogous regardless of brain measurement. Functional connectivity within the inferior frontal cortex, anterior cingulate and medial PFC and inferior parietal cortex (part of the DMN and FPCN) [53, 77, 78] as well as visual regions, were the nodes with highest associations with change in Fluid Cognition [79]. For FD, networks related to memory, cognitive control, and visuospatial processing - which were also among those reported for functional connectivity - showed the strongest links. Finally, for fiber complexity, the inferior parietal cortex and visual regions - which overlap with networks reported for functional connectivity - and anterior and medial PFC, and temporoparietal occipital junction - which overlap with FD - were the nodes with the strongest associations. As such, most of the regions showing significant patterns are part of the DMN, FPCN, and visual networks. The DMN is reportedly the most highly impacted network in early stages of AD [80, 81] and where amyloid first begins to accumulate [82], whereas the FPCN is essential for domains directly trained during CT, such as executive functioning, problem-solving, and working memory [83].

Few studies have investigated white matter changes due to CT in aMCI. Most research is also derived from tensor models [84, 85] and oversimplifies the underlying neuroanatomy [39]. These have known methodological problems, including an inability to classify multiple fibers within the same voxel which can represent between 60% and 90% of all voxels in the brain [86–89]. Recent evidence suggests that, compared to DTI, FBA results in larger effect sizes and increased sensitivity [39, 90–92], making FBA better suited to identify longitudinal changes after CT and possible associations to cognition. We cannot make claims about the CW group given the lack of significant correlations but find it encouraging that the effect sizes in the CT group were larger and that, compared with active controls, CT still performed better across cognitive and brain outcomes.

It is important to note that our study has some limitations. Unfortunately, some of the participants did not complete post-intervention neuroimaging visits due to the COVID-19 pandemic. The small sample size, particularly in the context of the neuroimaging analysis, suggests that these findings should be considered preliminary. While our results provide early evidence that warrants further investigation, the neuroimaging results should be reproduced through larger, more robustly powered studies. By addressing these limitations and pursuing the proposed future research directions, we can potentially make a significant impact on understanding the intervention’s effectiveness and refine it for broader application.

In summary, the proposed long-term, multi-domain cognitive intervention appears promising in enhancing cognitive and brain functioning and delaying microstructural decline. Improved Fluid Cognition was associated with enhanced functional connectivity in parts of the FPCN and slower microstructural decline in the cingulum bundle, anterior thalamic radiation, and optic radiation. Also, specific brain patterns at baseline—mainly involving the DMN and FPCN - were predictive of changes in Fluid Cognition and might serve as guidelines to investigate who can benefit from CT.

Author contributions

EG contributed to MRI data collection, carried out image processing and statistical analysis, and drafted the manuscript. BAP contributed to MRI data collection, statistical analysis, and drafted the manuscript. HF and LD contributed to recruitment, study coordination, MRI data collection, and behavioral data collection and interpretation. BJ, RO, and LW contributed to study design and drafting the manuscript. SMH conceived of the study, participated in its design and coordination, and helped with drafting the manuscript. All authors read and approved the final manuscript.

Data availability

The data that support the findings of this study are not publicly available due to privacy restrictions but can be available from the corresponding author on reasonable request. Open-source neuroimaging software including FSL, MRtrix3, and SPM were used to analyze the data. R (version 4.1.3) was used to apply the LME models, and the code can be shared upon request.

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

The study was approved by the Stanford University Institutional Review Board (IRB #40335). All methods were conducted in accordance with relevant guidelines and regulations to ensure compliance with established standards. Informed consent was obtained from all participants.

Supplementary information

The online version contains supplementary material available at 10.1038/s41398-024-03153-x.