Cognitive training effects are shaped more by individual brain dynamics than age—evidence from younger and older women

Abstract

Given the well-established structural and functional changes in the aging brain, it is widely assumed that cognitive aging is primarily driven by robust group-level differences between young and older adults. However, our individual-level EEG functional connectivity analysis challenges this notion. We investigated the impact of cognitive training on functional brain connectivity using task-switching paradigms in 39 younger (18–25 years) and 40 older (60–75 years) women. Participants were randomly assigned to either a training group, which completed eight individualized 1-h cognitive training sessions, or a no-contact control group. EEG was recorded at both pre- and post-training sessions across three task-switching paradigms (trained and near-transfer versions). Unique functional connectivity of different sources of variation was examined by calculating how much variance was shared across stable traits (e.g., individual, age, and common factors) or dynamic states (e.g., task and training effects). Our results revealed that age accounted for only a modest proportion of variance, whereas self-similarity was a dominant factor—particularly in older adults. Similarly, group-level training effects were small but strongly modulated by individual neural profiles, suggesting person-specific trajectories. Participants recruited distinct neural networks across tasks, and even within the same task engaged unique, individual-specific network configurations, reflecting personalized brain adaptations to cognitive demands. Importantly, older adults displayed a shift from common to individual network patterns, consistent with increased neural specialization and compensatory mechanisms. These findings underscore the importance of moving beyond group-level contrasts toward models that capture the complexity of individual brain dynamics in cognitive aging and training responsiveness.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s11357-025-02039-0.

Introduction

Cognitive training refers to structured practice on tasks designed to enhance specific cognitive functions such as attention, memory, or executive control. These interventions aim to strengthen or maintain cognitive performance through repeated engagement with cognitively demanding activities. Identifying effective strategies to slow or prevent brain changes associated with brain aging, or to compensate for such changes, is a central challenge in the field of cognitive aging and dementia research.

Despite extensive efforts, cognitive training has yielded mixed results. A major limitation has been the insufficient understanding of the mechanisms through which such interventions exert their effects [1, 2]. While training typically leads to improved performance on the practiced tasks, broader benefits have proven more elusive [3]. Near-transfer effects—improvements on untrained tasks that rely on overlapping cognitive processes—are observed more consistently than far-transfer effects, which would reflect generalized gains in domains not directly targeted by training. However, even documented far-transfer effects tend to be modest in magnitude and of limited clinical relevance. Meta-analyses have reported inconsistent findings: some have shown benefits of cognitive training [4], while others have found no effect [5], fueling the debate over whether such gains transfer beyond the trained tasks [2, 6]. Moreover, age may moderate training outcomes. While many studies report larger gains in younger than in older adults [e.g., 7, 8], which can be interpreted as a magnification effect due to greater cognitive resources, others have found larger benefits in children or older adults, consistent with compensation effects among individuals with lower baseline performance [e.g., 9, 10].

These heterogenous findings are further compounded by methodological variability across studies, including small sample sizes, differences in control conditions and session numbers, and the use of qualitatively different training protocols—such as strategy-based vs. process-based interventions [11]. Previous work suggests that adaptive and multicomponent paradigms are generally more effective, although their implementation also varies across studies. Moreover, participants’ baseline cognitive abilities may also influence training outcomes: individuals with lower initial performance may benefit through compensatory mechanisms to bolster task-relevant skills, while higher-performing individuals may show a magnification of their existing abilities. For some tasks, there may be a threshold of performance below which participants are unlikely to benefit from training [11]. Consequently, while some individuals exhibit substantial, generalizable gains that persist over time, others show little to no improvement.

We can approach this heterogeneity by focusing on the neural mechanisms that underlie training effects. One common view considers brain function as dynamic in that it reflects ongoing cognitive processes and can reconfigure with training. Indeed, previous work suggests that for brain structure, training-related changes include increases in grey matter and cortical volume [for a review, see 4]. Functional changes are less consistent, but several studies have demonstrated decreased frontal and increased subcortical activation, suggesting that training may reduce the need for compensatory processes, especially for executive function. Precision neuroimaging (i.e., sampling sufficient data over multiple sessions to reveal meaningful individual divergence in the organization of the brain) also has shown intra-individual change, such that functional networks can vary across tasks and time [e.g., 12–14]. Alternatively, brain function can be viewed as stable in the sense that network properties reflect more stable traits such as age rather than cognitive content. For example, precision neuroimaging has shown that brain networks are relatively stable and show little variation over time and across states [e.g., 15–18]. In this latter case, brain function can be linked to long-term histories of regional co-activations. The truth likely lies along a continuum between these extremes, and it is therefore important to examine to what extent brain function reflects stable traits or dynamic states.

To do this, the present study leverages a previously published dataset that has already been analyzed from both behavioural and electrophysiological perspectives, where the aim was to examine the extent, mechanisms, and durability of training-induced improvements and transfer effects in younger and older adults. In Gaál and Czigler [19], we examined performance and event-related potentials in reference and transfer task-switching paradigms. We found that after training, older adults reached the performance level of younger participants, with P3b components emerging after both cues and targets. Behavioural gains were also evident in near-transfer tasks and to a lesser extent in far-transfer ones, particularly in alerting and orienting networks. In Nagy et al. [20], we investigated changes in resting-state EEG following training, using spectral power density and multiscale entropy. The results revealed that cognitive training modulated age-related differences in resting-state dynamics, suggesting a shift toward more controlled and internally oriented processing in the older training group.

We now use this same dataset to examine stable and dynamic aspects of oscillatory activity during task performance. Previous work has characterized brain oscillations using power spectral density and functional connectivity (measured most commonly using phase-based metrics and envelope correlations; [21, 22]) over different frequency bands. This is relevant given that distinct frequency bands have been associated with different cognitive functions. For example, in task-switching paradigms, delta oscillations are involved in inhibition and control processes [23–25]. Theta oscillations are associated with flexible cognitive switching between task rules, memory retrieval, and attentional control [26–28]. Alpha oscillations reflect proactive control in task switching, through memory processes such as retrieving information from working- and long-term memory, and inhibitory processes involved in attention and response selection [23]. Beta oscillations are involved in selective inhibition, decision-making, and memory processes [29–31]. Age effects are most consistently observed in the theta and alpha bands. Theta activity, often associated with episodic and working memory processes, tends to increase at rest in older adults [32] or remain unchanged [33, 34], but may decline during task engagement [34, 35], possibly reflecting compensatory effort or reduced neural efficiency. Similarly, alpha power—particularly in the upper alpha range, is linked to attentional control and memory manipulation—typically decreases with age [32], which may reflect slower processing speed and reduced interference resolution [36].

To better isolate age- and training-related neural effects, we focused on women—a group still underrepresented in the cognitive aging and training literature. This design choice also minimized variance associated with sex differences in brain structure, aging trajectories, and EEG signal characteristics. Building on this, we investigated the neural mechanisms underlying the observed training effects. Specifically, we examine the extent to which brain function reflects stable traits like individual, age, and common factors, or dynamic states such as task and training effects. By examining the relative contribution of these components, our aim is to move beyond group averages and toward a more nuanced understanding of person-specific brain dynamics in cognitive training.

Materials and methods

Participants

Younger (ages 18–25) and older (ages 60–75) women were assigned to either a training (younger: n = 19, age: 21.4 ± 1.7 years, education: 14.8 ± 1.6 years; older: n = 20, age: 65.3 ± 3.3 years, education: 17.7 ± 4.1 years) or a control group (younger: n = 20, age: 21.7 ± 1.6 years, education: 14.9 ± 1.4 years; older: n = 20, age: 66.1 ± 3.1 years, education: 17.7 ± 4.1 years). All participants were right-handed, reported normal or corrected-to-normal vision, and had no history of neurological or psychiatric disorders. WAIS-IV (Wechsler Adult Intelligence Scale, [39]) screening confirmed the absence of cognitive impairment in older adults (older control, IQ = 120.1 ± 15.8; older training, IQ = 117.9 ± 16.7) and indicated that younger adults were also within the normal range of cognitive functioning (younger control, IQ = 109.4 ± 14.4; younger training, IQ = 107.9 ± 12.3). Participants provided written informed consent and were compensated for participation. The study was approved by the Joint Psychological Research Ethics Committee (Hungary).

Procedure

Participants completed two sessions: pre-training (baseline) and post-training (after ∼1 month). WAIS-IV was administered at baseline and during subsequent sessions. EEG recordings were obtained in each session. The training group received eight 1-h adaptive training sessions between pre-training and post-training; controls received no intervention.

Each EEG session included five tasks in fixed order: (1) resting-state (eyes open and closed); (2) informatively cued task-switching with nogo trials (reference task); (3) non-informatively cued task-switching with nogo trials (near-transfer); (4) informatively cued colour-shape classification (near-transfer); and (5) Attention Network Test (far-transfer)—tasks 2, 3, and 4 were analyzed here. Instructions were delivered both orally and in writing. Stimuli were pseudorandomized and task order was consistent across sessions.

While deep phenotyping approaches benefit from the inclusion of multiple task paradigms, we restricted our analyses to three structurally similar task-switching tasks—the trained reference task and two near-transfer tasks (see Fig. 1)—in which the largest training-related effects have previously been demonstrated. This targeted selection enhances methodological consistency and allows us to better isolate the effects of cognitive training.

Fig. 1.

Schematic overview of the three task-switching paradigms. (1) Informatively cued task switching with nogo stimuli—letter classification and parity task; (2) non-informatively cued task switching with nogo stimuli—letter classification and parity task; and (3) informatively cued task switching—colour and shape classification task. Each trial began with a black screen, followed by a cue (informative in tasks 1 and 3, non-informative in task 2) displayed for 700 ms. The target then appeared (a letter-number pair with possible nogo stimuli in tasks 1 and 2; or a pair of shapes in task 3) and remained onscreen for 2000 ms or until response. In the informatively cued conditions, cue colour indicated the relevant task. In the non-informatively cued condition, task identity was conveyed by the target’s colour. Feedback was provided for 200 ms following response

Reference task: informatively cued letter–number switching

This paradigm required participants to classify either a letter (vowel/consonant) or a number (odd/even) based on a preceding colour cue (700 ms, yellow/orange for letter task; blue/green for the number task). Cues were never repeated consecutively to dissociate task switching from cue switching. Targets were grey letter-number pairs (1.8° × 1.4°), presented centrally for 2000 ms or until response. Response mapping was counterbalanced across participants. In 25% of trials, a special character was presented, requiring participants to withhold their response (nogo trials). Both incongruently and congruently mapped targets were included to discourage strategic simplification; congruent trials and their immediate successors were excluded from analysis. Immediate feedback was provided after each trial for 200 ms. The task included single-task blocks (two each for letter and number classification; 104 trials per block) and mixed-task blocks (10 blocks; 500 trials total).

Near-transfer task: non-informatively cued letter–number switching

Identical to the reference task except that cues were non-informative (grey), and task identity was conveyed solely by the colour of the target (yellow/orange for letter, blue/green for number). All other parameters matched the reference task.

Near-transfer task: informatively cued colour–shape switching

Participants classified pairs of geometric shapes (triangles, squares, hexagons; 5.5° × 2.75°) by either colour or shape based on the cue colour (purple/brown for shape; white/yellow for colour). Tasks included single-task blocks (1 per task, 48 trials/block) and mixed blocks (5 blocks, 240 trials total). No nogo trials were used in this paradigm.

Training session

Participants completed an informatively cued task-switching paradigm with nogo trials across 20 blocks of 50 trials each. Four tasks were cued by distinct colour pairs: (1) parity task (odd vs. even number), (2) letter classification task (vowel vs. consonant), (3) magnitude task (less than vs. greater than five), and (4) case judgment task (uppercase vs. lowercase letter). Difficulty level was adjusted after each block based on performance: it increased following good performance (at least 33 hits), decreased after poor performance (less than 30 hits), and remained unchanged for intermediate outcomes.

EEG recording and preprocessing

Participants were seated in an electrically and acoustically shielded room. Stimuli were presented using Presentation software at the centre of a monitor placed 125 cm from the participant. EEG was recorded at 1000 Hz using NuAmps amplifiers (bandpass, DC–70 Hz) and NeuroScan 4.4 software (Compumedics, Victoria, Australia). Thirty-five passive Ag/AgCl electrodes were placed according to the 10–20 system at Fp1, Fp2, F7, F3, Fz, F4, F8, FT9, FC5, FC1, FC2, FC6, FT10, T7, C3, Cz, C4, T8, TP9, CP5, CP1, CP2, CP6, TP10, P7, P3, Pz, P4, P8, PO9, O1, Oz, O2, and PO10. The reference was the tip of the nose, and FCz served as ground. Vertical and horizontal electrooculogram (EOG) signals were also recorded. All electrode impedances were kept below 10 kΩ.

Offline EEG data were processed in the following sequence: (1) the data were bandpass filtered using a finite impulse response (FIR) filter between 0.1 and 30 Hz (48 dB/octave), (2) segmented into epochs time-locked to the target (from −100 to 1000 ms), (3) linear trends were removed, (4) baseline correction was applied using the prestimulus interval, and (5) trials exceeding ± 80 µV on any channel were automatically rejected as artefacts. Additionally, the first trial of each block, error trials, congruently mapped target trials, and the trials immediately following them were excluded from further analysis.

Connectivity analysis

Using the Fieldtrip toolbox [40] in MATLAB version 2019b, we estimated functional connectivity by calculating imaginary coherence (IMCOH; Nolte et al., 2004) across epochs for each unique pairing of 34 electrodes (561 pairs). IMCOH is a phase-based measure of the synchronization of brain regions at the EEG sensor level and involves selecting only the imaginary component from a coherency calculation. It is robust to artefacts of volume conduction by being insensitive to zero time-lag synchrony between regions, thereby removing spurious connectivity.

Data analysis

We performed analyses on all groups together and each group separately to examine the similarity of functional connectomes matched on properties including participant, age, training, task, and session following the methodology by Gratton et al. [17] and van der Wijk et al. [16]. An overview of the analysis procedure is presented in Figs. 2, 3, and 4. First, we extracted phase information for each epoch, channel, and frequency bin (ranging from 1 to 30 Hz, bins of 1 Hz) with Fieldtrip’s fast Fourier transformation algorithm (Fig. 2a). A Hanning taper was used. Then, we estimated functional connectivity for all electrode pairs within each individual, session, and task using IMCOH with Fieldtrip’s connectivity function (Fig. 2b). Thus, for each individual, session, and task, there was a 561 (electrode pairs) × 30 (frequencies) functional connectivity matrix in which each cell represented the IMCOH value for that electrode pair and that frequency.

Fig. 2.

Overview of the analysis procedure. The example shown illustrates the analysis including all groups, but the same steps were applied in the group-specific analyses. a The EEG data for the target stimuli from the informatively cued letter–number switching (reference), non-informatively cued letter–number switching (near-transfer), and informatively cued colour–shape switching (near-transfer) tasks were preprocessed using the same pipeline and concatenated separately for each participant, task, and session. b Functional connectivity (FC) was then calculated as the correlation between the time courses of each electrode pair for each frequency from 1 to 30 Hz, yielding an electrode-by-electrode FC matrix for each participant, task, session, and frequency. These matrices were then averaged for four frequency bands: delta (1–4 Hz), theta (4–7 Hz), alpha (8–12 Hz), and beta (12–30 Hz). c The upper triangle (grey) of each FC matrix was correlated with all others (across participants, tasks, and sessions) to generate a similarity matrix. This procedure was conducted separately for each frequency band, resulting in four similarity matrices (adapted from Fig. 1 in van der Wijk et al. ([16])

Fig. 3.

Schematic overview of different sources of similarity in the data. The analysis including all groups is used as an example, but the same steps were applied in the group-specific analyses. The contribution of each source of variance was estimated by calculating average similarity over different cell configurations of the similarity matrix. The presented patterns display the configuration for the calculation in the following order: (a) Row 1 (left to right): Common effect (similarities in FC across all participants and sessions), Session effect (similar FC across participants within time points), Task effect (similarities in FC across participants within tasks), and Age effect (similarities in FC among older and younger participants); (b) Row 2 (left to right): Training effect (similarities in FC among trained participants and controls), Training × Session interaction (similarities in FC among trained participants and controls within time points), the Training × Age interaction (similarities in FC among older trained participants, younger trained participants, older controls, and younger controls), and Training × Task interaction (similarities in FC among trained participants and controls within tasks); (c) Row 3 (left to right): Session × Age interaction (similarities in FC among older and younger participants within time points), Session × Task interaction (similarities in FC across participants within time points and tasks), Task × Age interaction (similarities in FC among older and younger participants within tasks), Training × Age × Session interaction (similarities in FC among older trained participants, younger trained participants, older controls, and younger controls within time points); (d) Row 4 (left to right): Training × Session × Task interaction (similarities in FC among trained participants and controls within time points and tasks), Training × Age × Task interaction (similarities in FC among older trained participants, younger trained participants, older controls, and younger controls within tasks), Session × Age × Task interaction (similarities in FC among older and younger participants within time points and tasks), and Training × Session × Age × Task interaction (similarities in FC among older trained participants, younger trained participants, older controls, and younger controls within time points and tasks); and (e) Row 5 (left to right): Individual effect (similarities in FC within individuals), Individual × Task interaction (similarities in FC within individuals and tasks), and Individual × Session interaction (similarities in FC within individuals and time points) (adapted from Fig. 1 in van der Wijk et al. ([16])

Fig. 4.

Calculation of effect magnitudes. The analysis including all groups is used as an example; but the same steps were applied in the group-specific analyses. a From the average similarity for each effect, we calculated the normalized relative effect magnitude by first subtracting the baseline for each effect (indicated by the dashed red lines) and then dividing by the total explained similarity. If an effect was smaller than its baseline, its magnitude was set to zero. b We calculated the average similarity for each effect per individual by taking the average of the patterns for the six rows representing each participant (example shown for the Common effect, where each row represents one participant’s similarity to all other participants across tasks and session). We then used dependent samples t-tests to test whether these participant-specific similarities contributed significantly to the overall similarity structure of the data (adapted from Fig. 1 in van der Wijk et al. ([16])

Next, the data were separated according to four frequency bands (delta, 1–4 Hz; theta, 4–7 Hz; alpha, 8–12 Hz; beta, 12–30 Hz), and the IMCOH index was averaged for each frequency band. This reduced the functional connectivity matrices for each individual, session, and task to a 561 (electrode pairs) × 4 (frequency bands) matrix in which each cell represented the averaged IMCOH value across the frequencies included in the frequency band for each electrode pair. The data were then transposed and concatenated into a connectivity matrix, separately for each frequency band, with the averaged IMCOH values for each electrode pair, session, task, and individual (Fig. 2b). In these matrices, each cell contained a connectivity (IMCOH) value, each column an electrode pair, and each row a level of each condition, e.g. the first row was the first individual’s first task and first session, the second row was the first individual’s first task and second session, and so on. For the separate group analysis, we simply further separated these matrices by group.

Next, for each connectivity matrix we correlated the upper half of each session, task, and individual’s whole-brain connectivity matrix across tasks with that of each other session (including sessions from the same individual), task, and each individual’s whole-brain connectivity pattern to construct a similarity matrix (Fig. 2c). In other words, we correlated each column of the matrix with each of the columns of the matrix. For example, the correlation of the first and the second column represents the correlation in functional connectivity between the first and the second session for the first task and the first individual, i.e. how similar the functional connectivity is between the two sessions within task and within individual. The higher the value of the correlation, the more similar the sessions are in terms of functional connectivity within the task and within the individual. In the resulting similarity matrix, each row and column represented a specific individual, task, and session, and each cell of the matrix was a correlation value. Fisher’s z-transformation was applied to these similarity matrices to normalize the data.

To quantify the magnitude of different sources of similarity, we calculated the average over different parts of the similarity matrix (as illustrated in Fig. 3). These included all main effects: Common (all cells), Session (cells belonging to the same session), Task (cells belonging to the same task), Age (cells belonging to the same age group), and Training (cells belonging to the same training group). For example, in the case of Session, this included all cells from the matrix that represented correlations between two first sessions or two second sessions, independent of individual and task, but excluded all cells that represented correlations between a first and a second session. The average across these cells gave us the similarity of functional connectivity within a session independent of individual and task. We also calculated average similarity for the intersections which represent all possible interactions between the conditions, e.g. Session × Task, Age × Task, and so on. Intersection here means a conjunction: only cells belonging to both conditions were included in the average. To estimate the within-individual similarity, the cells belonging to the same individual were also averaged (Individual). Additionally, we calculated averages for the intersections Individual × Session and Individual × Task to analyze whether any differences in Session or Task are modified by individual characteristics. The main diagonal of the matrix (all ones by definition) was not included in the averages. In the separate-group analyses, Training, Age, and their interactions were not included.

To compare the similarities in functional connectivity between the effects, averaged values were calculated for each participant. For this calculation, we extracted the rows belonging to each participant and then averaged across the cells corresponding to each effect as described above (see also Figs. 2c and 3). Thus, we had a similarity value for each effect and for each individual. The values were then submitted to dependent sample t-tests in MATLAB with permutation testing (1000 permutations) to estimate significance. Specifically, for each comparison, the data were randomly shuffled between the two effects within an individual for each permutation. While the data were not fully independent, they were exchangeable, which is what is required for permutation testing [41]. We compared each effect with its baseline (Fig. 4b), i.e. each main effect to the Common effect, each interaction to the largest main effect included in it, and each higher-order interaction to the largest lower-order interaction (Fig. 4a). The Individual effect was compared to the largest main effect or interaction, while the Individual × Session and Individual × Task interactions were compared to the Individual effect. In other words, we tested whether the similarity in functional connectivity for each condition was significantly larger than its baseline. We used the false discovery rate (FDR) method to correct for multiple comparisons [42].

To show the proportion of explained similarity for each effect, we calculated the normalized relative effect magnitude as an indication of effect size. First, from each effect, we subtracted its baseline (Fig. 4a) and then divided by the sum of all baseline-corrected magnitudes. As such, normalized relative effect magnitudes reflect the proportion of each effect that was unique, i.e. not explained by its baseline, relative to the combined unique magnitude of all effects. If an effect had a lower magnitude than its baseline (the difference was negative), the baseline-corrected magnitude was set to zero. The sum of all normalized relative effects is 1.

Results

Sources of similarity

We examined the contributions of different sources of similarity relative to their baselines across four frequency bands, both across all four groups combined and within each group separately.

Combined analysis across for all four groups and three tasks

In this analysis, we focused solely on the effects of Age; other effects are addressed in separate group analyses. Age accounted for a significant portion of similarity across all frequency bands (1.7–5.2%). Additionally, the Training × Age (0.8–1.5%) and Task × Age (0.5–2.1%) interactions contributed significantly in all bands. Session × Age (0.5%) and the three-way Training × Age × Session (0.6%) interactions were significant only in the delta band, while the Training × Age × Task interaction was significant in both the delta and theta bands (1.4 and 0.7%). Nevertheless, the Age effects contributed a substantially smaller proportion to the similarity compared to the Individual FC (29.2–55.8%) and its interactions (1.0–19.3%), or the Common FC (22.6–38.5%).

Figure 5 shows the similarity matrices for each frequency as well as mean similarities for each effect. Figure 6 shows the normalized relative effect magnitude for each effect across frequency bands (i.e. how much of the similarity is explained by each significant effect); it also provides a summary of the statistical significance of each effect. Supplementary Information Table S1 summarizes the statistical results.

Fig. 5.

a Similarity matrices for each of the four frequencies: delta (1–4 Hz), theta (4–7 Hz), alpha (8–12 Hz), and beta (12–30 Hz). The matrices include all participants and are structured as seen on Fig. 2c. Darker colour indicates higher similarity. The effects are calculated as sums of the cells shown in darker grey on Fig. 3. b Averaged similarities for each effect as well as significant differences when the effects were compared to their baselines as seen on Fig. 4a for each of the four frequencies: delta (1–4 Hz), theta (4–7 Hz), alpha (8–12 Hz), and beta (12–30 Hz). The bars indicate the averaged similarity, the light grey portions indicate each effect’s baseline, and the dark grey portions indicate the part of the effect above its baseline. The red lines show which effects are significantly larger than their baselines. Error bars show standard error

Fig. 6.

Normalized relative effect magnitude for each effect across all four groups and all frequency bands as seen on Fig. 4a. Only significant effects are shown. Those explaining more than 1% of the similarity are displayed individually, while significant effects contributing less than 1% are grouped under ‘Other’. The percentage of explained similarity for each effect is shown below the bars. Stars indicate statistical significance. Although significant across all frequency bands, the contribution of Age to similarity was small compared to Individual and Common FC

Group-specific analyses across the three tasks

Session-specific FC contributed significantly to similarity only in the younger training group (delta, theta, and alpha bands, 0.4–3.5%), while the effect of training was reflected only in the Individual × Session interaction (theta, alpha, and beta bands, 3.7–8.6%) in the older training group. This interaction was also significant in the younger training group (delta, alpha, and beta bands, 9.3–11.4%) and in both control groups in the beta band, likely indicating practice effects in the latter.

Task-specific FC had a small (0.52–3.00%) but significant contribution in all groups in the delta and theta bands. Importantly, in the control groups, the Individual × Task interaction was substantial in the delta, theta, and alpha bands (12.44–29.63%), indicating that individuals engaged distinct functional networks when performing different tasks. This interaction was smaller in the training groups—significant only in the theta band for younger adults and in the delta and theta bands for older adults—presumably because the Individual × Session interaction accounted for a larger proportion of similarity. Overall, Individual FC explained the largest proportion of similarity across the four groups (24.8–60.5%), with particularly high contributions in the older training group (37.8–59.2%).

Figure 7 shows the similarity matrices for each frequency-separated group. Figure 8 shows mean similarities for each effect for each frequency separated by group. Figure 9 shows the normalized relative effect magnitude for each effect across frequency bands separated by group (i.e. how much of the similarity is explained by each significant effect); it also provides a summary of the statistical significance of each effect. Supplementary Information Table 2 summarizes the statistical results.

Fig. 7.

Similarity matrices for each of the four frequencies: delta (1–4 Hz), theta (4–7 Hz), alpha (8–12 Hz), and beta (12–30 Hz). The matrices are separated by group (Age × Training) and follow the structure seen on Fig. 2c for Task × Session. Darker colour indicates higher similarity. The effects are calculated as sums of the cells shown in darker grey on Fig. 3 within each group

Fig. 8.

Averaged similarities for each effect as well as significant differences when the effects were compared to their baselines as seen on Fig. 4a for each of the four frequencies: delta (1–4 Hz), theta (4–7 Hz), alpha (8–12 Hz), and beta (12–30 Hz). The bars indicate the averaged similarity, the light grey portions indicate each effect’s baseline, and the dark grey portions indicate the part of the effect above its baseline. The red lines show which effects are significantly larger than their baselines. Error bars show standard error

Fig. 9.

Normalized relative effect magnitude for each effect across frequency bands, computed separately for each of the four groups as seen on Fig. 4a. Only significant effects are included in the bar graphs. The percentage of explained similarity by each effect is shown below the bars. Stars indicate statistical significance. The largest proportion of similarity was explained by Individual FC in the theta, alpha, and beta bands of older adults, and in the alpha and beta bands of younger adults. The second largest contribution came from the Common FC, which was most prominent in the delta band of older adults and in the delta and theta bands of younger adults. Together, these two factors accounted for 85.5–89.28% of similarity (normalized relative effect magnitude) in the beta band, and even in the delta band—where their contribution was the lowest—they still explained 67.54–77.7%, indicating that Task and Training effects and their interactions were substantially smaller in comparison

Sensitivity power analysis

As our study builds on previously collected data, an a priori statistical power analysis could not be performed. In such cases, sensitivity power analysis can be conducted to show the minimal effect size that can be detected with the present sample [43]. In the combined analysis, we reached 95% power for effect sizes (Cohen’s d) above 0.411, i.e. we should have been able to detect reliably medium and large effects. In the group-specific analyses, we reached 80% power for effect sizes above 0.66, i.e. we should have been able to detect reliably only large effects. Details are provided in Supplementary Information Table S3 and Figure S1.

Discussion

Building on our earlier work [19], we set out to identify the factors that shape cognitive training-induced changes in functional connectivity, with the broader goal of understanding how individual variability may inform the development of personalized interventions. Using a deep phenotyping approach [16, 17], we characterized the unique contributions of age, cognitive training, and individual-level factors through EEG-based functional connectivity measures during task-switching paradigms in younger and older adults, who completed an eight-session adaptive training protocol designed to improve cognitive functions.

One of our most unexpected and intriguing findings was that age per se accounted for only a small proportion of the variance (1.75–5.18%) in EEG functional connectivity, despite the widespread assumption in the cognitive aging literature that young and older adults differ substantially. This suggests that while chronological age contributes to variability in neural responses, individual differences may play a more substantial role, at least in our sample of women aged 18–25 and 60–75 examined in a cross-sectional design. Aging may reflect a shift toward increased interindividual variability rather than uniform decline.

In line with this age-associated shift, the main effect of Session (i.e. pre–post change) was not significant in the older training group—similarly to what was expected and observed in both control groups—but the Individual × Session interaction accounted for substantial similarity. This indicates that training responses in older adults are highly individualized, and that group-level effects may underestimate the functional plasticity retained in later life. In contrast, younger adults showed a significant main effect of Session in all frequency bands except beta, yet even here, the Individual × Session interaction explained a much larger share of the similarity (except in the theta band), again underscoring the dominance of idiosyncratic training trajectories. Importantly, in both age groups, and especially in the training conditions, the Individual × Session interaction was substantially larger than the main effect of Session, indicating that the effectiveness of cognitive training cannot be fully captured by group-average changes. Instead, training appears to interact strongly with individual neural architectures or strategies, potentially explaining the high variability across cognitive training studies in the literature. These findings are consistent with the revised Scaffolding Theory of Aging and Cognition (STAC-r, [37]), suggesting that functional reorganization and compensatory mechanisms—such as functional redundancy—may underlie training responsiveness in older age.

In addition, we observed a robust Individual × Task interaction, more pronounced in the control group, indicating that participants—including those in the control condition—engaged distinct neural networks to solve different tasks, even within the same experimental condition. Thus, we found no evidence in favour of a single, unified neural response pattern, but rather identified multiple connectivity patterns that vary across individuals. Interestingly, training appeared to reduce this heterogeneity, possibly by promoting more efficient or convergent network recruitment. A significant Task × Session interaction further supports the notion that training altered functional connectomes, albeit secondarily to the more nuanced individual-level patterns.

Finally, we observed that the proportion of explained similarity attributable to Common vs. Individual components differed across age groups: in younger adults, Common variance dominated, whereas in older adults, Individual variance was more prominent. This pattern suggests that network specialization increases with age, possibly reflecting the ways that an individual’s unique experiences interact with brain aging processes.

Most of our significant effects were observed across all frequency bands, indicating that training and task-related changes broadly influence neural networks. Exceptions to this pattern were the Task and Task × Individual effects, which were most prominent in the delta and theta bands. This may reflect the task-specific demands of the paradigm including inhibitory and control processes, which have been linked to delta band activity [23–25], and cognitive switching and attentional control, which have been linked to theta activity [26, 28].

Limitations and Conclusions

There are some limitations to this study. First, by focusing on women, our study offers new insights into age and training-related connectivity effects. However, Hungary ranks very high on the Human Development Index (HDI; a composite statistic based on life expectancy at birth, achieved and expected years of schooling, and gross national income per capita; https://ourworldindata.org/human-development-index), and therefore, our sample fits with a WEIRD (Western, Educated, Industrialized, Rich, and Democratic, [44]) characterization. Even within this context, our participants reported high levels of education, which has been shown to confer neurocognitive benefits, suggesting that our sample is skewed toward the most able, and may not be representative of the broader adult population in Hungary and elsewhere. Therefore, future work involving larger samples and more diverse, mixed-gendered participants (and researchers) from populations across a range of HDI should be conducted to extend our findings. Second, we used three structurally similar task-switching tasks, including the trained and two near-transfer task-switching paradigms. Although this approach provided a consistent framework for assessing within- and across-task neural similarity, future work should extend this approach to a wider range of cognitive domains. Third, we used a passive and not active control group, which may affect the interpretation of the training effects. The addition of active control groups would reduce bias by accounting for non-specific factors such as participant effort and attention, allowing for a clearer assessment of the specific effects of cognitive training. Building on the present results, future studies should also be specifically designed to examine behavioural-neural linkages to characterize unique neural signatures associated with the training response.

Taken together, our findings highlight that chronological age is not the primary determinant of training-related functional changes. Rather, they underscore the central role of individual-level variance in shaping how cognitive training effects unfold in the brain. We emphasize that, in most of the cognitive training literature, references to individual differences usually pertain to demographic characteristics, socio-economic status, motivation, personality traits, or baseline cognitive performance [38]. In our study, however, we focus on a different level of individuality—specifically, the uniqueness and stability of each participant’s functional brain connectome. This neural individuality appears to strongly influence both the magnitude and the trajectory of training-related brain responses, suggesting that truly effective and transferable interventions will need to move beyond group averages and adopt personalized approaches. In this context, understanding the functional architecture of individual brains may offer a promising route toward more targeted, efficient, and impactful cognitive training programmes across the lifespan.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contribution

Zs.A.G. and I.C. designed research; Zs.A.G. performed research; A.B.P., P.K., and Zs.A.G. developed analysis plan; P.K. and B.N. analyzed data; G.W. contributed analytic tools; Zs.A.G., P.K., and A.B.P. wrote the paper; Zs.A.G., P.K., B.N., G.W., I.C., and A.B.P. revised the paper

Funding

The research was supported by the Hungarian Research Fund (OTKA PD 101175 and OTKA K 145940) and the János Bolyai Research Scholarship of the Hungarian Academy of Sciences to ZAG, and by a Natural Sciences and Engineering Research Council of Canada Discovery Grant (05299–2020) to ABP.

Data availability

All data and code supporting the findings of this study are openly available in the Open Science Framework (OSF) at https://osf.io/kc9br/.

Declarations

Ethics approval and consent to participate

All participants of the study provided written informed consent, and the study was approved by the Joint Psychological Research Ethics Committee, Hungary (2011/18).

Competing interests

The authors declare no competing interests.