A total of 60 participants (64–78 years old) were recruited through an advertisement in the local newspaper. When applicants showed interest in participating in this study we contacted them to explain the purpose of the study as well as to verify that they were eligible to participate. Exclusion criteria included diabetes, medication known to influence dopamine, having a neurological disease, scoring below 27 on the mini-mental state examination (MMSE), having claustrophobia, or regularly performing moderately high to high-intensity exercise. A radiologist assessed structural images but no abnormality warranting exclusion was found. Participants provided written informed consent prior to the start of the study and were compensated with 1000 SEK for their participation. After baseline assessments, the participants were randomized into either an aerobic training group or an active stretching and toning control group. One female participant in the aerobic group had to discontinue training due to a foot injury, and one male participant in the control group due to a knee problem. Thus, the final sample included 29 participants in the aerobic training group and 29 participants in the active control (Table 1). The regional ethical committee in Umeå, Sweden, approved this study.

After inclusion, each participant was scheduled for baseline data collection on six separate days. On day one, participants underwent aerobic fitness testing and body composition measurements (bone densitometry, not reported here). Neuropsychological testing was performed on three occasions on three separate days. On additional 2 days, scanning of the brain was made with magnetic resonance imaging (MRI, with T1 weighted imaging reported here, but also resting-state functional MRI, diffusion tensor imaging, anterior spin labeling, C2-C3 blood flow, and T2 weighted imaging), as well as positron emission tomography to measure dopamine D2 receptors (not reported here). In addition, several questionnaires about daily activities, sleep, motivation, subjective memory, and wellbeing were filled out but will not be reported here.

Aerobic fitness was estimated from a standardized graded cycle (Monark 839E, Monark Exercise AB, Sweden) ergometer test performed by an experienced tester at the Sport Science Lab at Umeå School of Sport Sciences. Every 3 min, the resistance was incremented by 30 W, with starting values being 30 W for females, and 40 W for males. Expired air was measured through a mouthpiece and analysis of O₂ uptake (VO₂ = O₂ ml/kg ^*min) was performed in an open system (Oxycon Pro, Erich Jaeger GmbH & Co, Würzburg, Germany) every 20 s. Heart rate was registered each minute with a Polar Sport chest transmitter (Polar H1, Kempele, Finland). Self-perceived exertion was assessed on a 15-point RPE-scale (ranging from 6 to 20) (Borg, 1970) at the end of each 3 min interval. The test was terminated at volitional exhaustion or when the self-perceived exertion was rated 15 or above at baseline, and 17 at follow-up. VO₂ peak was estimated as the highest VO₂ reached before test termination.

In the aerobic training group, participants trained in order to increase their VO₂ by walking or jogging on an athletic indoor track, cycling on stationary cycles, or using cross-trainers. Their heart rate (HR) was monitored while they exercised 3 days a week for 6 months (30–60 min per session). As training continued, heart rate (HR) load of each session was increased incrementally, from 40 to 80% of their estimated maximum HR.

In the stretching and toning control condition, participants performed circuit training involving various bodily exercises aimed to improve muscle strength, flexibility, and balance without influencing VO₂ to a large extent. They made between 6 and 10 repetitions per exercise and each exercise had various levels of difficulty.

Testing was performed on three separate days and lasted between 60 and 90 min per session. On the first day, tests of reasoning and visuospatial ability were performed. Trials on those tasks were self-paced and served to accommodate participants to the testing situation and computerized testing. On all computerized tasks recording reaction times (RT), individual responses deviating by more than 2.5 standard deviations from the mean were excluded. In addition, RT on a given trial was included only for correct responses. Unless stated otherwise, the tasks were presented using the E-Prime 2.0 software (Psychology Software Tools, Pittsburgh, PA). Each task will be described under its respective cognitive domain but see Figure 1 for a flow chart of the task order.

Structural imaging was performed on a 3T General Electric scanner equipped with a 32-channel head coil. High-resolution T1-weighted structural images were collected with a 3D fast spoiled gradient echo sequence (180 slices with a 1 mm thickness, TR 8.2 ms, TE 3.2 ms, flip angle 12°, field of view 25 × 25 cm).

Freesurfer (Fischl et al., 2002), version 6-beta (20151015) longitudinal stream (Reuter et al., 2012), was used to segment the brain into known anatomical structures. In the longitudinal stream a subject specific base template is created from several time points. Errors visually observed on the base template were manually edited prior to creation of the final longitudinal segmentation. For the subcortical segmentations, the volume (mm³) of hippocampus was used as dependent measures. For the cortical segmentation, the dependent measure was cortical thicknesses (mm) from the Destrieux atlas (Destrieux et al., 2010).

Similar to Vijayakumar et al. (2014), a dlPFC region-of-interest (ROI) was produced by combining the bilateral superior and middle frontal gyri, a vlPFC ROI by combining the bilateral opercular, orbital, and triangular gyri, and an ACC ROI by combining the bilateral anterior and middle-anterior part of the cingulate gyri and sulci.

Details for the CFAs are reported in Supplementary 1. In short, a five factor solution with a “Cognitive score” factor with loadings from EM, PS, UPD, and EF showed best fit, χ(50,N=118)2 = 75.199, p < 0.012, RMSEA = 0.065, CFI = 0.946, AIC = 3652.311, SRMR = 0.079. The loading from UPD to “Cognitive score” was highest, followed by EF, PS, and finally EM. Due to the small sample size we do not consider the estimated loadings to be reliable however [see e.g., (Wolf et al., 2013) who, based on simulation studies, recommend n > 150 when all (true) loadings are expected to be above.8]. Hence, instead of using a factor score, the EM, PS, UPD, and EF constructs were averaged to form a unit-weighted “Cognitive score.”

Both groups had, compared to baseline, higher aerobic fitness after the intervention (Table 2, Figure 2). The rmANCOVA revealed a significant group by time interaction favoring the aerobic group, F_{(1, 53)} = 5.215, p = 0.0264, confirming the desired effect of the intervention.

Due to presence of six multivariate outliers we decided to conduct rmANCOVAs on the “Cognitive score” and the separate cognitive constructs instead of a MANOVA on the constructs (Table 2, Figure 3). A significant group by time interaction on the “Cognitive score” favored the aerobic group, F_{(1, 53)} = 7.700, p = 0.0076. There were no group by time interactions on EM, F_{(1, 53)} = 2.707, p = 0.1058, PS, F_{(1, 53)} = 2.090, p = 0.1542, UPD, F_{(1, 53)} = 1.411, p = 0.2402, and EF, F_{(1, 53)} = 2.080, p = 0.1551, or SRS, F_{(1, 53)} = 2.439, p = 0.1243. However, a significant interaction favoring the control group was seen on the RS construct, F_{(1, 53)} = 5.219, p = 0.0264, but only after controlling for age, gender, and education. For descriptive purposes, results for all tasks and constructs in the neuropsychological test battery, including within group rmANCOVAs, are reported in Supplementary Table 1.

The MANOVA with dlPFC, vlPFC, and ACC thickness, and hippocampus volume revealed no group differences in cortical thickness or hippocampus volume, F_{(4, 108)} = 0.128, p = 0.972; Wilks' Λ = 0.995. There were no group interactions or significant effects of time in any of the regions, all p > 0.05 (Table 3).

Additional analyses were conducted to understand possible associations between aerobic capacity, cortical thickness, and cognitive performance. Results of the regressions are reported in Table 4. At baseline, higher VO₂ peak was positively associated with dlPFC thickness, but was unrelated to vlPFC and ACC thickness as well as hippocampus volume (Figure 4). The change in VO₂ peak was positively associated with the change in hippocampus volume, but unrelated to change in cortical thickness (Figure 5).

“Cognitive score” at baseline was significantly associated with larger thickness in dlPFC (p = 0.01) and vlPFC (p = 0.01), but not to ACC, or to hippocampus volume (Figure 4). Following the 6-month training period the change in “Cognitive score” was positively associated with the change in dlPFC (Figure 6). The association between changes in “Cognitive score” and dlPFC thickness was driven by the aerobic group, showing a significant effect, F_{(1, 28)} = 4.828, p = 0.037, r = 0.39, (Figure 6B) whereas the control group did not, p = 0.34.

Further analyses revealed construct specific influences in expected anatomical structures. Higher EF at baseline was specifically associated with thickness in dlPFC, F_{(1, 57)} = 8.167, p = 0.006, r = 0.35, vlPFC, F_{(1, 57)} = 5.596, p = 0.021, r = 0.30, and ACC, F_{(1, 57)} = 8.309, p = 0.006, r = 0.36, but not to hippocampus volume F_{(1, 57)} = 0.62, p = 0.44, r = 0.10. Conversely, EM showed a specific association with larger hippocampus volume at baseline, F_{(1, 57)} = 4.661, p = 0.035, r = 0.28, but not to cortical thickness. Over time, improved EF was positively associated with the change in dlPFC thickness specifically, F_{(1, 57)} = 4.761, p = 0.033, r = 0.28. EM was unrelated to changes in cortical thickness and hippocampus volume.

A considerable strength of the present investigation was the inclusion of several tests for each cognitive construct, providing the opportunity to investigate cognition in a broad sense. The CFA displayed good fit indices despite the small sample size, disregarding TS. Previous studies have generally had fewer tasks and/or investigated fewer constructs (Burns et al., 2009; Erickson et al., 2011; Legault et al., 2011; Thomas et al., 2016). Thus, the present observation that aerobic exercise compared to stretching and toning control training was more effective in improving cognition in older adults is an important contribution to a field showing equivocal results, owing at least in part, to a lack of robust test batteries enabling use of latent constructs. Different tasks have been used across studies and although the tasks are taken to represent the same cognitive constructs, this is not always the case, e.g., improved digit-symbol task performance in aerobically fit individuals could be taken to support the visuospatial hypothesis (Stones and Kozma, 1989), although this particular task is commonly viewed as a task measuring speed of processing (Salthouse, 1996). Thus, we support a general improvement in cognitive function which is in contrast to Young et al. (2015) but according to several other meta-analyses (Colcombe and Kramer, 2003; Angevaren et al., 2008; Smith et al., 2011). Our conclusion of an effect in a broad sense was further supported by our construct specific analyses. Compared to baseline, aerobic exercise resulted in improvements in several cognitive constructs; PS, UPD, EF, and SRS, with a trend also for EM (Figure 3, and Supplementary Table 1). The largest effect size was seen for EF, assimilating conclusions from an earlier meta-analysis (Colcombe and Kramer, 2003). The control group improved PS and EF abilities, albeit to a lesser degree, and not in the other cognitive constructs. This suggests that even low-intensity exercises aiming to promote strength, balance, and flexibility may have positive effects on cognitive functioning. Motor fitness and resistance training have been related to improved executive functioning (Liu-Ambrose et al., 2008; Voelcker-Rehage et al., 2010, 2011) and processing speed (Voelcker-Rehage et al., 2010, 2011), hence, group comparisons may mask aerobic exercise-induced changes to cognition due to training-induced improvements also in the control group. There was an interaction favoring the control group on the reasoning construct not expected a priori. However, there could have been a regression to the mean effect influencing this particular analysis (Figure 3), and the interaction was due to the aerobic group's worsened performance, rather than improved performance for the control group. In sum, our observations indicate that aerobic exercise, compared to stretching and toning control training improves cognition generally, rather than being specific to any given component.

Regarding aerobic exercise, fitness, and cortical thickness, results were rather equivocal. A direct effect of the intervention on cortical thickness in dlPFC, vlPFC, or ACC was absent. However, subsequent analyses indicated that aerobic fitness at baseline had a positive relationship with cortical thickness in dlPFC, partly consistent with previous cross-sectional data showing positive effects on PFC and ACC gray matter from aerobic fitness (Colcombe et al., 2003; Weinstein et al., 2012) and physical activity (Flöel et al., 2010). As the aerobic exercise group improved their aerobic fitness considerably compared to the control group one could have expected to find an effect from aerobic fitness in dlPFC over time, explaining also the improved cognitive performance, considering that “Cognitive score” predicted dlPFC thickness both at baseline and change over time in the aerobic group. However, one explanation is that the exercise-induced improvements in cognitive performance immediately after a period of training is not mainly due to gray matter changes, but may depend on altered cardiovascular responses (Colcombe et al., 2004; Voelcker-Rehage et al., 2011), white matter changes (Voss et al., 2013; Oberlin et al., 2016), functional connectivity (Voss et al., 2010, 2016), brain-derived neurotrophic factor (Leckie et al., 2014), or other molecular processes. Wheel-running for example, has been associated with alterations in dopaminergic markers in striatum (MacRae et al., 1987; Meeusen et al., 1997; Kim et al., 2016), and considering the influence of dopamine both on prefrontal and striatal function in relation to executive control functions, (Frank and O'Reilly, 2006; Cools and D'Esposito, 2011; D'Esposito and Postle, 2015) this could be a potential pathway for exercise-induced improvements in cognition.

Moreover, training duration may also be an important factor, where effects of exercise may be absent at 6-months, only to be revealed after 12 months (Voss et al., 2010). Duration could potentially explain why baseline levels, and not 6-month change, predicted cortical thickness, assuming that baseline fitness levels can be taken to represent an average fitness level, stretching further back in time than 6 months. It should also be noted that recent studies have shown that resistance training may also influence the brain, e.g., by altering cardiovascular responses in PFC and ACC (Voelcker-Rehage et al., 2011). The active control group likely improved both their motor fitness and muscle strength; hence, effects may be influenced also by factors we did not measure, and mask potential group comparisons. Future studies could include additional tests to control for changes in motor and muscle fitness. In sum, our analyses of cortical thickness are only partly consistent with the view that being aerobically fit at an older age preserves brain health in executive control regions (Kramer and Erickson, 2007; Prakash et al., 2015).

As opposed to the pattern observed in dlPFC, hippocampus volume did not show an association to aerobic fitness at baseline, as was presented by Erickson et al. (2009). Conversely, the change in aerobic fitness was positively associated with the change in hippocampus volume, replicating previous studies (Erickson et al., 2011; Thomas et al., 2016). This effect could be explained by increased neurogenesis (van Praag et al., 1999; van Praag, 2005) or angiogenesis (Pereira et al., 2007; Maass et al., 2014), as a function of aerobic fitness improvement. It should also be noted that resistance training has been linked to increased hippocampus volume (Niemann et al., 2014). Functionally, hippocampus volume was related to EM at baseline, but the change in hippocampus volume could not predict changes in EM. One explanation as to why hippocampus volume may not have been predictive of verbal EM over time is that new neurons may not have been incorporated in functional circuitry, considering that a large proportion of newborn hippocampal cells do not survive the first few weeks without environmental challenges (van Praag et al., 1999). An alternative explanation is that we used verbal memory tasks, and previous exercise studies finding a relation between EM and hippocampus used spatial memory tasks (Erickson et al., 2011; Maass et al., 2014), also in rodents (van Praag, 2005; Yau et al., 2012). Others, using verbal EM measures have also failed to find an association (Pereira et al., 2007; Thomas et al., 2016). Hence, spatial memory tasks could be more sensitive to structural changes in hippocampus (Bonner-Jackson et al., 2015).

Despite using the gold standard for estimating VO₂ peak, we are limited in making firm claims concerning the absolute changes in VO₂ peak. We speculate that this may be due mainly to two factors. Firstly, different termination thresholds were used at baseline and at follow-up. Secondly, leg strength has an influence on estimated VO₂ peak when using a cycle ergometer test, (Diesel et al., 1990), and both training protocols were likely to influence leg strength. Including additional measurements on leg strength could have been informative and should be considered. In addition, including accelerometers or activity bracelets during the intervention period for monitoring everyday physical activity should be an important source of information for future studies.

Limitations of this study include only measuring two time points, 6 months apart. For instance, it appears as if hippocampus is sensitive to the timing of measurements, which was recently showed in a study observing that aerobic exercise may increase hippocampus volume within 6 weeks, while returning back to baseline values also within 6 weeks (Thomas et al., 2016). The time between aerobic fitness measurements and MRI acquisition in the present study was between 1 and 2 weeks, potentially introducing a source of noise. The issue of timing is also related to the lack of a clear understanding of the dose-response relationship between aerobic fitness and various brain indices (Prakash et al., 2015). For instance, Voss et al. (2010) found effects on functional connectivity in the default mode network and a frontal executive network only after 12 months, not after 6, suggesting that longer training or time between scans may be required to detect cortical changes.

Moreover, practice effects from repeated testing are a potential limitation affecting the cognitive results, but unlikely explain the improvements entirely despite the lack of interactions on specific constructs; due to the long test-retest duration, comparably large effect sizes (Makizako et al., 2013; de Oliveira et al., 2014; Goldberg et al., 2015), and an active control group that may also have exhibited training-induced cognitive improvements (Liu-Ambrose et al., 2008; Voelcker-Rehage et al., 2010). Finally, although white matter has also been investigated in relation to physical activity (Voss et al., 2013; Hayes et al., 2015; Oberlin et al., 2016) we restricted this study to mainly gray matter thickness in frontal regions and hippocampus volume. On that note, Freesurfer may also be less specific than voxel-based morphometry, or diffusion weighted imaging, as Freesurfer assigns voxels to either gray or white matter, thereby limiting our ability to be specific about the exact changes captured in our ROIs. Nevertheless, cortical thickness from Freesurfer has shown to be predicitive of cognitive functioning in older adults (Burzynska et al., 2012), and concords with voxel-based morphometry when assessing atrophy (Lehmann et al., 2011).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.