Patients and ethics

This retrospective observational study used a large clinical dataset of 327 patients (188 women and 139 men; mean age ± standard deviation, 77.9 ± 7.1 years [range 49–93 years]) who visited our outpatient department for dementia. The inclusion criteria were patients visiting the department for the first time between 5 June 2019 and 11 April 2024. The exclusion criteria were as follows: (1) lack of documentation of regular physical activity, (2) refusal to consent for data reuse, and (3) not undergoing MEG. According to the clinical records, 32 were diagnosed with healthy ageing, 60 with mild cognitive impairment, and 220 with dementia. Thirteen were diagnosed with other medical conditions such as depression or disuse syndrome. Two patients did not present on the day of diagnosis. We included all patients who met the aforementioned criteria, regardless of their diagnosis, because we would like to apply our findings to all patients who seek lifestyle intervention. This study was approved by the Ethics Committee of Kumagaya General Hospital (#34) and adhered to relevant Japanese guidelines and regulations. All data used for the present study were anonymised at the first stage. Only two of the authors (KF and YS) had the ability to identify individual patients using comparative tables, if necessary. Data were accessed for research purposes between 31 May and 5 July 2024. All participants provided written informed consent to participate in this study if they were cognitively healthy. Otherwise, their legal guardians (i.e., family members) provided informed consent on their behalf. Because all patients visited the outpatient department for dementia and any clinical problems were suspected to be cognition-related, we obtained written informed consent from both the patients and their legal guardians, except for special circumstances. The procedure was performed in accordance with the Ethical Guidelines for Medical and Health Research Involving Human Subjects, published by the Japanese Ministry of Education, Culture, Sports, Science, and Technology.

Classification of active and nonactive groups

Documentation of regular physical activity was obtained from clinical records written by the attending physician (Author YH) during medical interviews. Patients were classified as ‘Active’ if their records indicated any regular exercise routine, such as daily walking or participation in a gymnastics club, whose amount, duration, and frequency were at a reasonable level considering each patient’s background judged by the physician. Those without such a routine were classified as ‘Nonactive’. Patients whose records did not describe regular exercise routines were excluded. We intentionally did not set thresholds for the classification of ‘Active’ and ‘Nonactive’ across participants, such as types of activities (e.g., walking, swimming, playing golf, and trekking), duration (e.g., two hours per day), and frequency (e.g., every two days), because the definition of ‘Active’ varies depending on patients’ backgrounds. As described in a guideline by the World Health Organisation for preventing cognitive impairment and dementia, even small amounts of physical activity can be considered as an exercise for some [43]. Smaller than recommended amounts of physical activity can still benefit cognition [44]. Old individuals often suffer from physical disabilities such as dizziness or pain in their hip or knee. Even 2,000–3,000 steps of walking every two days represents a significant effort for them. While we may deem this a small amount of exercise for us, it is significant for them and improves their physical/mental condition and cognition. Due to the diversity in the clinical population, it was neither practical nor fair to apply uniform thresholds for the classification of ‘Active’ and ‘Nonactive’ groups. Instead, despite being subjective, the patients were classified by a physician, considering their daily activity status and their clinical background.

Neuropsychological assessments

Regarding the cognitive state, we employed three neuropsychological assessment scores that are routinely used in our department: the total scores of the Japanese versions of the MMSE (MMSE-J) [45,46], Frontal Assessment Battery (FAB-J) [13,47], and Alzheimer’s Disease Assessment Scale-Cognitive section (ADAS-J cog) [48]. MMSE-J and FAB-J scores range from 0–30 and 0–18, respectively, with lower scores indicating more severe impairment. ADAS-J cog scores range from 0–70, with higher scores indicating more severe impairment. MMSE-J scores were missing for one patient, FAB-J scores for five patients, and ADAS-J cog scores for 65 patients because of clinical limitations.

MEG data acquisition

To quantify electrophysiological brain activity, three spectral parameters—MF, IAF, and SSE—were retrieved from clinical records [30,36]. They had been computed from five minutes of resting-state brain activity (i.e., spontaneous neural oscillatory activity). The activities were recorded using a 160-channel whole-head magnetoencephalography system (RICOH160−1; RICOH, Tokyo, Japan) in a magnetically shielded room as part of clinical practice. Patients were asked to remain calm in the supine position with their eyes closed during the scan. The sampling frequency was 2000 Hz with 500 Hz low-pass filtering during the recording.

MEG analysis

The three spectral parameters—MF, IAF, and SSE— had been computed using the same sensor-level analysis protocol used in our previous studies and those of other groups [25,29–31,34,36–38,40,49–51] during daily clinical practice at Kumagaya General Hospital. MEG data are sometimes contaminated by artefacts such as signal fluctuation caused by dental works, which can deteriorate the quality of analysis. Before starting the analyses, artefacts were manually removed using principal component analysis, if necessary, by experienced physicians or clinical laboratory technicians (Authors FK, SI, and YS). The artifact removal was conducted using RICOH MEG Analysis software (RICOH, Tokyo, Japan), an analysis software provided by the MEG manufacturer. The technique was applied to 125 of 207 MEG datasets as part of clinical practice. The number of removed components were adjusted for each dataset and limited to as few as necessary. Subsequent MEG analyses were performed offline using MATLAB (MathWorks, Natick, MA, USA). The time-series signals were band-pass (1–70 Hz) and band-stop filtered at 50 Hz to remove the power line noise. The power spectral density (PSD) was computed using the Blackman-Tukey approach [52] with non-overlapping 5-s segments of the filtered signals. In the Blackman-Tukey method, PSD is computed as a discrete Fourier transform of the autocorrelation function of the filtered time-series signals, which has better precision than other approaches [53] and is commonly used for computing MEG spectral parameters [25,30,31,34–37,50]. The original PSD was divided by the total power in the frequency range of interest (1–70 Hz) and the normalised PSD (PSDn) was obtained.

Three spectral parameters (i.e., MF, IAF, and SSE) were calculated to summarise different characteristics of the PSDn [20,23,25,29–32,34–38,40,49,50]. The MF quantifies the frequency at which the spectral power is balanced between low and high frequencies. It divides the PSDn into two equal halves between 1 and 70 Hz. The IAF represents the dominant frequency corresponding to the peak of the PSDn in the alpha band. It is defined similarly to MF, but with an adjusted frequency range between 4 and 15 Hz (i.e., extended alpha band) instead of 1–70 Hz, to obtain a robust estimator of the dominant alpha oscillations. SSE is computed as the normalised Shannon’s entropy to the PSDn, which can be calculated as a probability density function:

(1)

where N is the number of frequency bins of the PSDn. SSE represents an irregularity measure closely related to the concept of order in information theory, which quantifies the homogeneity in the distribution of the oscillatory components of the PSDn. Sensor- and epoch-wise MF, IAF, and SSE were computed. Then, they were averaged across all sensors and epochs. These parameters have been used in clinical practice at some hospitals [25,34], with lower values indicating a lower cognitive state, and are associated with neuropsychological assessment scores [20,23,25,29–38].

Changes in the three spectral parameters (i.e., MF, IAF, and SSE) reflect changes in patients’ cognitive states [34] and neurophysiological conditions in the brain, such as changes in excitation/inhibition ratios and neuromodulator release. Regarding cognitive states, IAF and SSE are associated with MMSE and FAB scores, respectively, while MF is well associated with both MMSE and FAB scores [25]. This discrepancy suggests that three MEG spectral parameters are associated with distinct cognitive functions. Regarding neurophysiological conditions, low-frequency oscillatory activity (e.g., delta and theta) is associated with ascending cholinergic input to cortices from subcortical regions, such as the nucleus basalis of Meynert [54,55]. Amyloid deposit in the cortex also enhances the low-frequency oscillatory activity [39]. Stroke also enhances low frequency oscillatory activity [56]. Clinically, pathological structural changes often influence on its amplitude. Alpha oscillatory activity, which is considered an intermediate frequency, is produced by the thalamocortical network [57] and plays a key role in cognition [58]. Although its frequency is distributed from 8 to 12 Hz, it slows down in patients with mild cognitive impairment [59] and Alzheimer’s disease [60]. Its power attenuates in patients with Alzheimer’s disease [61] and dementia due to Lewy body [62]. High-frequency oscillatory activity (e.g., beta and gamma) is associated with gamma-aminobutyric acid level [63] which is a inhibitory neurotransmitter [64]. It is related to synaptic activities [65], network activities [66], and neuroplasticity [67]. In the predictive cording theory [68], low-frequency oscillatory activity is associated with top down processing in the brain, while high-frequency oscillatory activity is associated with top down processing bottom-up processing [69]. MF represents the power balance between wide frequency range from delta to gamma, while IAF just focus on frequency balance around alpha which is sensitive to frequency change in peak alpha frequency. SSE covers the same frequency range but evaluates different aspects of PSD: entropy rather than a simple power balance between high- and low-frequency oscillatory activities.

Statistical analysis

Statistical analyses were performed using MATLAB (MathWorks, Natick, MA, USA). Given the large number of samples (N = 327), we used parametric statistical tests for all analyses. First, to examine the effect of regular physical activity on cognitive states and electrophysiological activities, we examined group-level differences in each neuropsychological score, MEG spectral parameter, and age between the two groups (‘Active’ and ‘Nonactive’) using two-sample t-tests.

Next, to explore the relationships within/between cognitive states and electrophysiological activities, Pearson’s correlation coefficients were calculated to investigate the intra- and intercorrelations within/between the age, neuropsychological scores, and MEG spectral parameters for each group. The correlation coefficients were tested with the null-hypothesis of no-correlation against the alternative hypothesis of a nonzero correlation. For the t-tests and correlation analyses, p-values were adjusted for false discovery rate (FDR) using the Benjamini–Hochberg method [70], with significance set at less than 0.05. Next, we assessed the effect of regular physical activity on MEG spectral parameters while accounting for cognitive state using analysis of covariance (ANCOVA). Each MEG spectral parameter (MF, IAF, and SSE) was considered as a dependent variable, while regular physical activities (‘Active’ and ‘Nonactive’) was used as the independent variable and age or one of the neuropsychological assessment scores (MMSE-J, FAB-J, or ADAS-J cog) was used as a covariate.

Finally, the impact of regular physical activity on MEG spectral parameters was quantified in the unit of neuropsychological assessment scores using size of contribution (SOC) computed with the following method. We estimated the intercept of the group (‘Active’ and ‘Nonactive’) and slope of the neuropsychological assessment scores for subjecting the MEG spectral parameters using the linear regression models which were equivalent to the ANOCOVA model;

(2)

where y is one of the MEG spectral parameters (i.e., MF, IAF, or SSE; dependent variable), β₀/ β₁ represent the global intercept/ slope across all groups, and β_0i/ β_1i represent the intercept/ slope for group i (1 = ‘Active’ or −1 = ‘Nonactive’; independent variable). x represents the covariance in the ANCOVA model, which was age or one of the neuropsychological assessment scores (i.e., MMSE-J, FAB-J, or ADAS-J cog). ɛ represents a residual of the measurement. For each pair of MEG spectral parameter and covariate (i.e., age, MMSE-J, FAB-J, or ADAS-J cog), the parameters were estimated using the least square method. To enhance the comprehensibility of the results, each covariate was centred (i.e., the mean was subtracted) before model estimation, so that the intercepts represent the estimated y (i.e., MEG spectral parameters; dependent variable) at an average covariate. We defined the SOC as the size of group intercept (β_0i) relative to the global slope (β₁); therefore, the SOC was computed by dividing (β_{0 Active} - β_{0 Nonactive}) by β₁ for each model, indicating the difference in MEG spectral parameters between groups (‘Active’ and ‘Nonactive’) expressed by the unit of neuropsychological assessment scores. The SOC was computed for the pairs of MEG spectral parameters and neuropsychological assessment scores for which the corresponding ANCOVA showed a significant main effect of group (i.e., ‘Active’ and ‘Nonactive’) and covariate (i.e., MMSE-J, FAB-J, or ADAS-J cog). A schematic description of our statistical model and SOC is shown in Fig 2. Of note, we did not evaluate nor compare the performances of the regression models. They were designed equivalently to the ANCOVA models in which the terms for group and neuropsychological assessment scores were significant, and the coefficients were solely estimated to supplement the results with SOCs. Therefore, we did not aim to explore the best models or predictors in the present study.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Schematic Description of the Statistical Model and SOC.

The red dashed and blue solid lines represent the least square lines for the Active and Nonactive groups, respectively. The black vertical line indicates the mean of the neuropsychological assessment score (i.e., MMSE-J) across all patients, whose intersections with the least square lines corresponded to the group intercepts (β₀ + β_{0 Active} and β₀ + β_{0 Nonactive}). The intercepts are converted to the unit of neuropsychological assessment score by horizontally projecting them on the global least square line with a global slope of β₁, whose gaps between groups are considered as the SOC, representing the effect of group on MF in the unit of MMSE-J at their mean. MF: Median Frequency, MMSE-J, Japanese version of Mini-Mental State Examination; SOC, Size of contribution.

https://doi.org/10.1371/journal.pone.0326163.g002

Group-level differences and correlations

One hundred forty-seven (79 women and 68 men) and 180 patients (109 women and 71) belonged to the Active and Nonactive groups, respectively. Descriptive statistics and the results of group-level comparison are shown in Table 1. The t-tests did not reveal any significant differences in age [p (FDR) = 0.465]. However, the MMSE-J and FAB-J scores were significantly lower [p (FDR) < 0.001], and the ADAS-J cog score was significantly higher [p (FDR) < 0.001] in the Nonactive group than in the Active group. All MEG spectral parameters (MF, IAF, and SSE) were significantly lower [p (FDR) <= 0.001] in the Nonactive group than in the Active group. The results of exploratory correlation analyses are shown in S1 and S2 Tables. Age was negatively correlated with MMSE-J and FAB-J scores in both groups, whereas it was only correlated negatively with ADAS-J cog scores in the Nonactive group and not correlated with MEG spectral parameters in either group. All pairs within neuropsychological assessment scores (i.e., MMSE-J × FAB-J, MMSE-J × ADAS-J cog, and FAB-J × ADAS-J cog) and MEG spectral parameters (i.e., MF × IAF, MF × SSE, and IAF × SSE) were significantly correlated. All pairs between neuropsychological assessment scores and MEG spectral parameters (i.e., MMSE-J × MF, MMSE-J × IAF, MMSE-J × SSE, FAB-J × MF, FAB-J × IAF, FAB-J × SSE, ADAS-J cog × MF, ADAS-J cog × IAF, and ADAS-J cog × SSE) were significantly correlated in both groups, except for the pairs including SSE (i.e., MMSE-J × SSE, FAB-J × SSE, and ADAS-J cog × SSE) for the Active group.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Profile of the dataset and results of group-level comparisons.

https://doi.org/10.1371/journal.pone.0326163.t001

Effect of groups on MEG spectral parameters

Table 2 summarises the results of the ANCOVA, in which the effect of groups on MEG spectral parameters was investigated while considering the effects of neuropsychological assessment scores. When age was considered as a covariate, the main effect of group (i.e., Active vs. Nonactive) was significant but the covariate and their interaction was not significant for all MEG spectral parameters [Table 2 (A)]. When each neuropsychological assessment (MMSE-J, FAB-J, or ADAS-J cog) was used as a covariate, both the main effects of the group (i.e., Active vs. Nonactive) and covariate were significant for all MEG spectral parameters, while their interactions were not [Table 2 (B)-(D)]. The impact of regular physical activity (i.e., effect of group) on the MEG spectral parameters was quantified as SOC. The SOCs were computed for the MMSE-J, FAB-J, and ADAS-J cog but not for age because the ANCOVA results indicated that age did not contribute to the regression model for any MEG spectral parameters. In the unit of MMSE-J, the SOCs of the groups were 4.14, 3.25, and 11.14 on MF, IAF, and SSE, respectively, indicating that regular physical activity affected the MEG spectral parameters by 3.25–11.14 in terms of MMSE-J scores. Similarly, the SOCs on MF, IAF, and SSE were 3.67, 2.99, and 6.85 in terms of FAB-J, and −10.16, −8.88, and −14.05 in terms of ADAS-J cog scores. The relationships between MEG spectral parameters, neuropsychological assessment scores, and groups are visualised in Fig 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Assessing the effects of regular physical activity on MEG spectral parameters.

https://doi.org/10.1371/journal.pone.0326163.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Scatterplots of the Relationships between MEG Spectral Parameters and Age/ neuropsychological Assessment Scores.

Open dots and filled dots represent Active and Nonactive groups, respectively. Dashed and solid lines represent least square lines for Active and Nonactive groups, respectively. MF, median frequency; IAF, Individual alpha frequency; SSE, Shannon’s spectral entropy; MMSE-J, Japanese version of Mini-Mental State Examination; FAB-J, Japanese version of Frontal Assessment Battery; ADAS-J cog, Japanese version of Alzheimer’s Disease Assessment Scale-Cognitive.

https://doi.org/10.1371/journal.pone.0326163.g003

Limitations

The current study has some limitations. First, this study was conducted in a single hospital with an MEG centre. To generalise the present findings, they should be replicated in other hospitals with different patients. We are currently preparing similar studies with various collaborating hospitals. Second, we cannot rule out the possibility of reverse causality. Although we assumed that regular physical activity enhanced the MEG spectral parameters, the causality might in fact be opposite; patients with high MEG spectral parameters could prefer a healthy lifestyle including regular physical activity. There could also be no causal relationship; for example, both could be consequences of other factors, such as different lifestyle-associated, genetic, environmental, and/or unknown factors. We cannot exclude these possibilities because this was a retrospective observational study, not an intervention study. However, the present findings are still valuable for the clinical practice. There is an association between regular physical activity and MEG spectral parameters which allows physicians to infer patients’ lifestyles. Patients with lower MEG spectral parameters than other patients with comparable current cognitive states will likely not have healthy lifestyles and should be encouraged to change them. We have previously reported the efficacy of this strategy in a case series [34] which was an intervention study. Future longitudinal studies are required to clarify the causality. Third, some potential confounding factors were not considered in the present study, both lifestyle-associated (e.g., education and environment) [86,87] and other factors (e.g., genetic factors) [88]. We have examined the comprehensive effects of regular physical activity, which includes its interactions with other lifestyle-associated factors. However, some lifestyle-associated and/or other factors could influence MEG spectral parameters independently from regular physical activity, which should have been controlled for. However, because this was a retrospective observational study based on existing clinical records, only limited information was available. We are planning further studies to investigate the influence of other factors.