The present study included 78 patients who met National Institute of Aging–Alzheimer’s Association (NIA-AA) criteria (McKhann et al., 2011; Albert et al., 2011; Jack et al., 2018) and 70 cognitively-unimpaired older adults. All participants were recruited from research cohorts at UCSF Alzheimer’s Disease Research Center (UCSF-ADRC). The diagnosis of AD patients was established by consensus in a multidisciplinary team. Among 78 AD patients, 20 had autopsy-confirmed AD neuropathology, other 41 patients were positive on the Aβ-PET scans, plus another nine patients showed cerebrospinal fluid (CSF) assays of amyloid and tau levels consistent with AD diagnosis. The remaining eight patients were clinically diagnosed, based on clinical evaluations and the characteristic pattern of cortical atrophy on MRI. Control participants were recruited from an ongoing longitudinal study of healthy aging at UCSF-ADRC. The eligibility criteria for cognitively normal controls included normal cognitive performance, normal MRI, and absence of neurological, psychiatric, or other major medical diseases. Forty-seven (out of 70) controls were evaluated with Aβ-PET and 8 were read as positive (39 as negative). The remaining 23 control participants were not evaluated with Aβ-PET. All participants underwent MMSE and a structured caregiver interview to assess the clinical dementia rating scale (CDR). All control participants were identified at CDR 0, indicating cognitively-unimpaired status on the CDR scale. Patients with AD ranged from 0.5 to 2 on the CDR scale. The results of demographic, functional, and cognitive assessments are shown in Supplementary file 1. Informed consent was obtained from all participants or their assigned surrogate decision makers. The study was approved by the Institutional Review Board of UCSF (UCSF-IRB 10–02245).

Structural brain images were acquired using a unified MRI protocol on 3T Siemens MRI scanners (MAGNETOM Prisma or 3T TIM Trio) at the Neuroscience Imaging Center (NIC) at UCSF, within an average of 1.05 years (range: -6.91–0.78) and 0.29 years (range: -2.13–1.29) of the MEG evaluation for controls and patients, respectively. The acquired MRI was used to generate the head model for source reconstructions of MEG sensor data and to evaluate GM volumes. The region-based GM volumes corresponding to the 94 anatomical regions included in the Automated Anatomical Labeling 3 (AAL3) atlas (Rolls et al., 2020, Supplementary file 2) were evaluated using the Computational Anatomy Toolbox [CAT12 version 12.8.1 (1987)] (Gaser et al., 2022), which is an extension of SPM12 (Penny et al., 2011); the regional GM volumes were calculated using the morphometry pipeline implemented in CAT12 with default parameters. The total intracranial volume (TIV), the sum of all segments classified as gray and white matter, and CSF, was also calculated for each subject.

The AECs are defined as Pearson’s correlation coefficients (PCCs) between any two amplitude envelopes of regional time courses (total 94×93/2=4371 pairs). Regional time courses were first processed by a band-pass filtering, and then their envelopes were extracted by the Hilbert transform. To discount spurious correlations caused by source leakages, we orthogonalized any two band-limited time courses before computing their envelopes by employing a pairwise orthogonalization (Hipp et al., 2012; Sekihara and Nagarajan, 2015). The AEC with leakage corrections is often expressed as AEC-c and is known as a robust measure (Briels et al., 2020b). The pairwise orthogonalization provides asymmetric values between two-time courses; the value depends on which time course is taken as a seed. Therefore, the PCCs between orthogonalized envelopes for both directions were averaged, resulting in a symmetric AEC matrix. Regional AECs, that represent the connectivity strengths of each ROI, were computed by averaging over row/column components of the symmetric AEC matrix.

The spectral power of a given band, which has often been used as a metric to discriminate patients with AD from controls (Jeong, 2004; Engels et al., 2016; Wiesman et al., 2021), is defined by the ratio of a band power to total power and was calculated from regional PSDs. Regional PSDs were calculated from the 94 regional time courses using Welch’s method (50% overlap) with 0.293-Hz (=fs/2048) frequency steps.

To identify general trends in changes in long-range and local synchrony with the severity of AD, we performed group comparisons of the regional synchrony metrics between AD patients and controls. Based on the group contrasts of regional metrics observed, we introduced scalar synchrony metrics by calculating the averages within several regions where large region-level group contrasts were identified. The scalar MEG metrics were used in the SAC-EBMs.

Given the most likely sequence S¯ as a result of the exhaustive search, the probabilities that a subject j falls into a stage k are evaluated by the posterior distribution:(3)pj(k)≡p(tj=k|Zj,S¯)=p(Zj|S¯,tj=k)∑k′p(Zj|S¯,tj=k′).

These probabilities describe the contribution of a subject j to stage k, allowing us to evaluate the stage value of any metric x of i at a stage k as a weighted mean:(4)x¯i(k)=∑j=1Jpj(k)⋅xij∑j=1Jpj(k).

Then, we represented the trajectory of the metric i by a series of the stage values, x¯i(k). The standard error (SE) of the weighted mean at stage k was evaluated by(5)SE(k)=σi⋅∑j=1Jpj(k)2(∑j=1Jpj(k))2,

where σi is a standard deviation of a metric i. This definition of SE provides an usual expression of the standard error of the mean, σi/J, if all subjects contributed equally to all stages.

These formulations of trajectories were applied to several metrics. In the AC-EBM, the metrics i denote the PHG volume loss z-score and the MMSE scores. In the SAC-EBM, they denote each scalar neural synchrony metric in addition to the PHG volume loss z-score and the MMSE score. We also used Equation 4 to evaluate the progressions of the regional neural synchrony metrics and the regional GM volume loss z-scores along the estimated EBM stages. When evaluating the proportion of subjects classified into each stage, we treated (xi1,xi2,…,xiJ) as a vector in which a metric i represents a category of subjects provided by the CDR scale. For example, when evaluating the ratio of subject with CDR 0.5, xij=1 only when a subject j has CDR scale of 0.5, otherwise xij=0.

To test demographic differences between AD patients and controls, the unpaired t-test was used for age and the chi-square test for sex. The age was defined at the time of the MEG scan date. In statistical analyses, p-values below 0.05 were considered statistically significant. For group comparisons of GM volumes, MMSE scores, and neural synchrony metrics, two-sided significance tests (against a null value of zero) were performed using the general linear model (GLM). For statistical tests on GM volumes, TIV, age, and the difference between MRI and MEG dates were included as covariates. For statistical tests on MMSE scores, age and the difference between MMSE and MEG dates were included as covariates. For statistical tests of neural synchrony metrics, age was included as a covariate. The problem of multiple comparisons between 94 regions was solved by controlling the Benjamini-Hochberg false discovery rate (FDR) (Benjamini and Hochberg, 1995). The FDR-adjusted p-value (i.e., q-value) below 0.05 or 001 was considered statistically significant.

A non-parametric test was performed to statistically compare metrics between stages, i.e., to test statistical significance of the difference between stage values represented by weighted means [e.g. stage k vs k′:δx=x¯i(k)−x¯i(k′)]. For a metric i, we used bootstrap resampling (50,000 samples) of an original data set, 𝒙i=(xi⁢1,xi⁢2,…,xi⁢J), to generate new data sets, (xi1⋆,xi2⋆,…,xiJ⋆), using a random number generator, where each xij⋆ is one of the components of the original data set 𝒙i. We then calculated the weighted means x¯i⋆(k) (Equation 4) for each sample. The same procedures were performed for stage k′, obtaining weighted means x¯i⋆(k′) for each sample. We then tested the null hypothesis that a weighted mean in stage k, x¯i(k), is equal to a weighted mean in stage k′, x¯i(k′), evaluating the null distribution of differences in the weighted mean values, δx⋆=x¯i⋆(k)−x¯i⋆(k′). The problem of multiple comparisons across stages was solved by controlling the FDR. The q-value below 005 was considered statistically significant.

This study included a cohort of 78 patients with AD (50 female; 28 male) including 35 patients with AD dementia and 43 patients with MCI due to AD, and also included 70 cognitively-unimpaired older adults as controls (41 female; 29 male). The CDR scales were 0 for the cognitively-unimpaired controls, 0.5 for patients with MCI, and 1⁢(n=27) or 2⁢(n=8) for patients with AD dementia. There were no differences in sex distribution between the AD and control groups [χ2(1)=0.477;p=0.49]. The average age at the time of MEG was slightly higher in the control group than patients with AD (controls, mean ± SE: 70.5±0.99, range: 49.5–87.7; AD, mean ± SE: 63.9±1.01, range: 49.0–84.4) [unpaired t-test: t(146)=−4.708;p<0.001]. The mean MMSE in patients with AD was 22.7±0.43 (mean ± SE) while the mean MMSE in the controls 29.2±0.48. MMSE scores were adjusted for age and time differences between MMSE administration and MEG scan using a GLM (Figure 1—figure supplement 3B). MMSE-decline z-scores, zMMSE, were standardized by adjusted MMSE scores of the control group and sign-inverted (Figure 1B).

Group comparisons of GM volumes for each of the anatomical regions included in the AAL3 atlas showed that GM volumes in the temporal regions are significantly smaller in AD patients than in controls (Figure 1—figure supplement 4; Supplementary file 4). Among temporal GM volumes, we focused on a volume of PHG as a key indicator of neurodegeneration in AD progression. The PHG includes the perirhinal and entorhinal cortices of the medial temporal lobe (MTL), and MRI-based studies have reported that MTL volume decreases, especially in the perirhinal and entorhinal cortices, in the early stages of typical AD (Teipel et al., 2006; Echávarri et al., 2011; Matsuda, 2016). In this study, PHG volume was defined as a sum of the volumes of left- and right-hemisphere PHGs. The average volume of PHG in AD patients (7.99⁢ml±0.09) was significantly lower than in controls (9.28ml±0.11) [unpaired t-test: t(143) = -9.508; ***p < 0.001] (Figure 1—figure supplement 3A); the PHG volumes were adjusted for TIV, age, and the difference between MRI and MEG dates by including them in a GLM as covariates. PHG volume loss z-scores, zPHG, were standardized by the adjusted PHG volumes of the control group and sign-inverted (Figure 1A).

We performed group comparisons of MEG metrics. Three canonical frequency bands were considered: 2−7Hz (delta-theta), 8−12Hz (alpha), and 15−29Hz (beta) bands (Figure 2—figure supplement 1). For regional long-range synchrony (AEC), increases in delta-theta-band synchrony in patients with AD were identified in frontal regions, and reductions in alpha- and beta-band synchrony were identified in the whole brain (Figure 2—figure supplement 2C, E; Supplementary file 5). These regional contrasts were similar to those observed between AD dementia and subjective cognitive decline (SCD) in MEG/EEG studies (Schoonhoven et al., 2022; Briels et al., 2020a). For regional local synchrony (spectral power), increases in delta-theta-band power in patients with AD were identified in the whole brain, and reductions in alpha- and beta-band power were identified in temporal regions and the whole brain, respectively (Figure 2—figure supplement 2D, F; Supplementary file 6). These regional contrasts were similar to those observed between MCI and controls in a multicenter study of MEG (Hughes et al., 2019).

Based on the group contrasts of regional metrics observed, we introduced six scalar metrics to quantify long-range and local synchrony: [i] frontal delta-theta-band AEC, [ii] whole-brain alpha-band AEC, [iii] whole-brain beta-band AEC, [iv] whole-brain delta-theta-band spectral power, [v] temporal alpha-band spectral power, and [vi] whole-brain beta-band spectral power. We computed the average within several regions where large group contrasts were identified at the region level (the temporal and frontal regions of interest (ROI) are illustrated in Figure 2—figure supplement 3). Consistent with regional group comparisons, the long-range and local synchrony scalar metrics in delta-theta band increased in AD patients compared to controls, and the long-range and local synchrony scalar metrics in alpha and beta bands were reduced in AD patients compared to controls (Figure 2—figure supplement 2A, B). We also calculated the z-scores, zMEG, of each scalar metric that was used in the SAC-EBMs.

An AC-EBM analysis with the two metrics, PHG volume loss, zPHG, and MMSE decline, zMMSE, was performed for six events (N=6; three events for each metric). Robust event thresholds were determined by the exhaustive search of multiple event thresholds (z-score thresholds) and choosing the set of event thresholds that maximize the data likelihood (Equation 1). The AC-EBM provided seven stages, each located between consecutive event occurrence times. The resulting posterior probabilities, pj(k), that a subject j belongs to a stage k are shown in Figure 1C. Based on the probabilities, the ratio of subjects classified into each stage was calculated as the probability-based weighted mean (Figure 1D). The ratio of subjects with CDR 05 was highest in stage 4, and the ratio of controls with CDR 0 in stage 4 was small compared to those in less severe stages of 1–3, indicating that stage 4 corresponds best to clinical MCI due to AD.

The trajectory of PHG volume loss preceded that of MMSE decline (Figure 1F), consistent with the relationship between brain atrophy and cognitive decline described in a hypothetical model of biomarker trajectories (Jack et al., 2010; Sperling et al., 2011). Figure 1E visualizes the distribution of the seven stages in the PHG volume loss versus MMSE score. At stage 4, the value of zPHG of 1.33±0.258 was in the range of 1–2. This z-score range of PHG volume loss corresponds to a mild-atrophy range representing approximately the MCI stage, e.g., in the voxel-based specific regional analysis system for AD (VSRAD) software (Hirata et al., 2005; Matsuda et al., 2012). Furthermore, the MMSE score of 26.3±0.82 at stage 4 was in the range of 23–27. This range of MMSE scores is considered typical for MCI due to AD (Tsoi et al., 2015). Stage 4, therefore, corresponds to MCI stage, whereas stages 3 and 5 correspond to preclinical-AD and mild AD-dementia stages, respectively.

The GM volume z-scores as a function of the seven stages showed that prominent atrophy with z>1 is observed in the temporal regions starting at stage 4 (Figure 1G). This trajectory of GM volume approximated the evolution of brain atrophy in the typical progression of AD reported in MRI-based studies; GM volume loss in AD starts in the MTL in the MCI stage, spreads to the lateral temporal and parietal lobes in the mild AD-dementia stage, and spreads further to the frontal lobe in moderate AD-dementia (Scahill et al., 2002; Tondelli et al., 2012; Jack et al., 2013).

These results of the AC-EBM indicate that the PHG volume loss precedes the MMSE decline, and their metric changes track the stages of AD from preclinical AD to moderate AD-dementia. The order of events for GM volume loss and cognitive decline was consistent with the observation that cognitive decline in the early stage of AD progression reflects neuronal loss in the medial temporal regions (Jack et al., 2018; DeTure and Dickson, 2019).

For the seven stages determined by the AC-EBM (Figure 1E–G), long-range and local neural synchrony profiles were estimated (Figure 2). Along the EBM stages, the delta-theta-band synchrony was consistently increased and the alpha and beta-band synchrony was consistently decreased. Neural synchrony showed prominent changes around stage 4 (clinical stage of MCI due to AD). The long-range synchrony in the alpha and beta bands decreased steadily in stages 1–3 and then decreased further in stage 4 (Figure 2A). Local synchrony in the beta band also decreased by half from 1 to 4 (Figure 2B). On the contrary, there were little changes in delta-theta-band long-range synchrony and delta-theta- and alpha-band local synchrony from stage 1 to 3 but these changes became prominent after stage 3.

Regional patterns of long-range and local synchrony as a function of the seven stages indicated that prominent changes manifest themselves at stage 4 (Figure 2E and F; Supplementary file 7 and Supplementary file 8). The regions with prominent deviations overlapped with the regions where a significant increase and decrease in neural synchrony was observed in the group comparisons (Figure 2—figure supplement 2E, F).

The changes in neural synchrony metrics with AD progression indicate that neural synchrony is a sensitive indicator of functional change along AD progression. To further investigate the temporal association of functional deficits with neurodegeneration and cognitive decline, we included neural synchrony in addition to the PHG volume loss and MMSE decline in the EBM frameworks, performing SAC-EBMs.

SAC-EBMs that include PHG volume loss, zPHG, MMSE decline, zMMSE, and long-range synchrony metric z-scores, zMEG, were performed setting a total of nine events (N=9). SAC-EBMs separately included long-range neural synchrony metrics in the delta-theta, alpha, and beta bands. Each EBM determined the order of nine events, thus defining ten stages (Figure 3—figure supplement 1; for the corresponding positional variance diagrams of the optimal set of z-score events in the SAC-EBMs, see Figure 3—figure supplement 2 and Figure 3—figure supplement 3). The resulting posterior probabilities, pj(k), that a subject j belongs to a stage k are shown in Figure 3—figure supplement 4.

For all frequency bands, around stages 5 and 6, the weighted means of PHG volume loss z-scores were in the range of 1–2 and the MMSE scores were in the range of 23–27 (Figure 3B, F and J). Furthermore, the ratio of subjects with CDR 05 was high around stage 5 (Figure 3A, E1). These indicated that stage 5 best represents the onset of clinical MCI stage, and stages 1-4, where MMSE scores remain almost constant at or near 30, correspond to the preclinical stages of AD. Changes in long-range synchrony during the preclinical stages are shown as statistical bars, and the region-level changes are shown in Figure 3C–D, G–H, and K–L.

Long-range synchrony in the alpha and beta bands decreased markedly during the preclinical stages of AD, preceding both PHG volume loss and MMSE decline. Specifically, between stages 1 and 4, the alpha- and beta-band long-range synchrony decreased by more than 80% of the total drop seen from stage 1 to 10. The whole brain, but especially the temporal area, was involved in these prominent preclinical changes (Figure 3H and L). In contrast, the trajectory of delta-theta-band long-range synchrony (Figure 3B and C) was almost identical to the evolution of the PHG volume loss throughout the stages, but a large variation occurred around the MCI stages (stages 5 and 6) as was found in the AC-EBM (Figure 2A). There were no significant increases in region-level synchrony in delta-theta band during the preclinical stages (Figure 3D), consistent with an observation seen in the AC-EBM (Figure 2F).

The trajectory shapes of the PHG volume loss (almost linear) and MMSE scores (half parabola) were similar to those obtained in the AC-EBM (Figure 1G). This indicates that prominent changes in alpha- and beta-band long-range synchrony during preclinical stages can be utilized to stratify the preclinical stages determined only by neurodegeneration and cognitive deficits.

SAC-EBMs including PHG volume loss, MMSE decline, and local synchrony metric z-scores were performed, separately considering delta-theta-, alpha-, and beta-band local synchrony metrics. When considering delta-theta and alpha bands, around stages 6 and 7, the PHG volume loss z-scores were in the range of 1–2 and the MMSE scores were in the range of 23–27 (Figure 4B and F), indicating that stage 6 best represents the onset of the MCI stage. Furthermore, the ratios of subjects with CDR 0.5 were high in stages 6 and 7 (Figure 4A and E). For the beta band, based on similar observations, stage 6 best represented the MCI stage (Figure 4I and J). For all frequency bands, stages 1-5, where MMSE scores remain almost constant at or near 30, corresponded to the preclinical stages of AD. The changes in local synchrony during the preclinical stages are shown as statistical bars, and the corresponding region-level changes are shown in Figures 4C–D, G–H and K–L.

Local synchrony in the alpha and beta bands decreased during the preclinical stages of AD, preceding both PHG volume loss and MMSE decline (Figure 4F and G and Figure 4J and K). On the contrary, the local synchrony in the delta-theta band increased, lagging the evolution of PHG volume loss (Figure 4B and C). Specifically, the alpha-band local synchrony decreased considerably by the onset of the MCI stage, showing significant reductions in the temporal regions (Figure 4H) during the preclinical stages (stages 6 vs 1). It is noted that these trends were inconsistent with those found in the AC-EBM (Figure 2B), especially within the preclinical stages, where there was little change found in the alpha-band local synchrony. This can be interpreted as evidence that early stages in AD progression may be better characterized by including neurophysiological markers as AD indicators. Beta-band local synchrony also decreased during the preclinical stages, preceding PHG volume loss and MMSE decline; by stage 5, the beta-band power decreased by approximately 55% of the total drop seen throughout the stages, and the reductions were observed in the whole brain (Figure 4L). Unlike the local synchrony trajectories in the alpha and beta bands, the local synchrony in the delta-theta band increased. The hyper-synchrony lagged the evolution of the loss of PHG volume in the preclinical stages and made a large jump around the stages 6 and 7 (Figure 4D).

As shown in the previous section, large alpha- and beta-band hypo-synchrony during the preclinical stages was also observed in long-range synchrony (Figure 3F and J). Notably, the decreases in the long-range metrics were much greater than those in the local metrics, especially in the early stages during the phase of preclinical AD (stages 1–3).

A key finding of the current study is that functional deficits as depicted by reduced neural synchrony precede structural volume loss and cognitive deficits. The EBMs on cross-sectional data clearly demonstrated that alpha- and beta-band synchrony within the inferior temporal and posterior parieto-occipital regions show significant deficits in the early disease stages–stages where volumetric and clinical deficits are still not significantly deviated from their baseline trajectory. This is consistent with the finding that functional changes occur earlier in the time course than structural changes in AD (Jack et al., 2010; Sperling et al., 2011).

Previous functional MRI studies have demonstrated disrupted connectivity especially between the hippocampus and several areas of the cortical default mode network (DMN) in subjects with amyloid deposition but without cognitive impairment (Sperling et al., 2014). This disruption in DMN has also been observed in clinically normal older individuals without prominent brain atrophy in MTL that preserves hippocampal activity (Miller et al., 2008; Hedden et al., 2009), indicating altered functional connectivity during the preclinical period of AD. In contrast to such fMRI data reflecting the cascade of neural, metabolic, hemodynamic events in AD, our findings from MEG, which captures the synaptic physiology as the collective oscillatory spectra, demonstrate direct observations of AD-related altered neuronal activity.

We demonstrated that oscillatory deficits and their temporal association with neurodegeneration and cognitive decline are frequency-specific. In particular, it is the alpha and beta hyposynchrony that precedes PHG atrophy and MMSE decline, whereas the delta-theta hypersynchrony does not seem to show such a precedence. This is consistent with previous findings that alpha and beta hyposynchrony is more tightly associated with tau accumulation, which is closely allied to neurodegeneration and cognitive decline (Pusil et al., 2019; Ranasinghe et al., 2020; Ranasinghe et al., 2021). Neural hyposynchrony in the alpha and beta bands may, therefore, represent harbingers of altered synaptic physiology associated with tau accumulation in AD. In fact, in human postmortem studies, the strongest correlate of cognitive deficits in AD patients is loss of synapse (DeKosky and Scheff, 1990; Terry et al., 1991). A study using transgenic AD mice has also shown that synaptotoxicity is an early phenomenon in AD pathophysiology (Zhou et al., 2017). In the context of fluid biomarkers to detect plasma amyloid, alpha and beta hyposynchrony can detect and quantify tau-associated neurodegenerative mechanisms, and hence may provide crucial information for early therapeutic interventions.

Previous studies have also shown that delta-theta oscillatory activity increases in AD and is strongly associated with amyloid accumulation (Ranasinghe et al., 2020; Ranasinghe et al., 2022b). In particular, increased delta-theta activity is a robust signal in individuals who are amyloid positive and cognitively unimpaired as well as those who harbor APOE-ϵ4 allele and an increased risk of AD (Cuesta et al., 2015; Nakamura et al., 2018). These previous findings indicate that delta-theta hypersynchrony is an early change in the AD spectrum and may even precede neurodegeneration and cognitive deficits. However, in the current results, the trajectory of the delta-theta hypersynchrony was identical to or lagged that of the PHG volume loss. This apparent controversy may be due to the possibility that oscillatory changes in the delta-theta band are more closely related to amyloid accumulations in AD, which become saturated early in the disease course and have a poor association with neurodegeneration and cognitive trajectories. It would be worth exploring how the trajectory of early saturated variables may be captured by EBM approaches.

The decrease in alpha and beta-band long-range metrics in the preclinical stages was much greater than that in the local metrics. This is consistent with the fact that AD-related abnormal brain activities are observed as disruptions of functional networks. Long-range cross-regional metrics, such as AECs, directly capture network disruptions involving all brain regions, while local metrics capture features of individual regions. From the definition, local synchrony describes collective neuronal oscillations in each local region, and thus the change along AD progression may depend mainly on long-term, slowly changing regional neuronal loss. On the other hand, long-range synchrony describes temporal coherence amongst regional collective neuronal oscillations and is vulnerable to altered neuronal oscillations. Therefore, long-range metrics are more sensitive to abnormal rhythms, collecting local abnormalities.

Preclinical neurophysiological markers that indicate the pathophysiology of AD are clinically important but have not been established. Aβ accumulation in preclinical stages is just a necessary condition for AD, and additional preclinical markers are required to fully predict the progression of AD. From this point of view, the present study indicates that alpha- and beta-band MEG metrics, especially long-range synchrony metrics (AEC), which were found to be sensitive to preclinical stages, could be promising candidates as additional markers.

A limitation of the current study is that there were differences in age between controls and AD patients. Although we adjusted the age of each metric using GLMs, age trajectories in neurophysiological measures have been reported to be nonlinear even in healthy aging (Sahoo et al., 2020). Age-related changes in brain atrophy have also been reported to follow a nonlinear time course depending on the brain areas (Coupé et al., 2019). These studies indicate that it may be better to employ a non-linear method beyond GLM to perfectly correct aging effects.

Another limitation is that we have not performed independent validations of the predicted trajectories and also have not examined the heterogeneity in AD progression, although we clarified for the first time the time courses of MEG neurophysiological metrics in AD progression. In fact, AD is a heterogeneous multifactorial disorder with various pathobiological subtypes (Jellinger, 2022). In this context, an EBM called Subtype and Stage Inference (SuStaIn) capable of capturing spatio-temporal heterogeneity of diseases (Young et al., 2018) has been proposed to subtyping neurodegenerative diseases including typical AD and has been applied to find different spatio-temporal trajectories of longitudinal tau-PET data in AD (Vogel et al., 2021). Since oscillatory rhythms are thought to depend on AD subtypes (Ranasinghe et al., 2017; Ranasinghe et al., 2022a), an extended trajectory analysis considering spatial and temporal variations of the MEG/EEG metrics is warranted in the future, and such analyses would provide distinct neurophysiological trajectories depending on AD subtypes. As a validation of the predicted trajectories, it would be necessary to investigate whether the predicted EBM stages are reliable and predictive of conversions (e.g. from control to MCI) while taking the AD subtypes into account.