Patients with AD were recruited in this study based on the National Institute of Neurological Disorders and Stroke-Alzheimer Disease and Related Disorders (NINCDS-ADRDA) criteria (Dubois et al., 2007). None of the participants had any history of head injury, stroke, depression, or tumor. All the patients had taken medicine such as acetylcholinesterase inhibitors (Donepezil) or N-Methyl-D-aspartate receptor antagonists (memantine) for at least 3 months at a stable dosage. In total, 51 participants with mild or moderate AD (clinical dementia rating ≤ 2) were recruited and randomly assigned into a real rTMS with cognitive training (n = 26) group or a sham group (only receiving cognitive training, n = 25) using a web-based randomization generator (http://www.jerrydallal.com/random/randomize.htm).

Before treatment, all patients underwent neuropsychological assessment and MRI scanning (T0). Then, the real rTMS or sham stimulation was repeated five times from Monday to Friday for 4 weeks. Another MRI scan and neuropsychological assessment were performed on Saturday morning at the end of the 4-week treatment (T4). To evaluate the long-time effect, an MRI scan and neuropsychological assessment were followed up 4 weeks after the end of treatment (T8). All neuropsychological measures before and after the treatment were assessed by a specialist with over 20 years of experience who was blinded to the allotment.

In the real group, rTMS was conducted in combination with cognitive training for up to 1 h each day. A butterfly coil (MCF-B65) with an inner diameter of 35 mm was used for the rTMS treatment, and the treatment was guided by an optical navigation system (Magventure, Germany). High-frequency rTMS pulses were delivered separately to the stimulation target first in the left DLPFC (Talairach coordinates: X = −35, Y = 24, Z = 48) and then in the left LTL (Talairach coordinates: X = −60, Y = −15, Z = −15). The following parameters were used: 20 trains (5 s duration at 10 Hz with an inter-train interval of 25 s), 100% resting motor threshold (RMT), 1,000 pulses, and 10 min in total for each target, and there were no maintenance sessions. The protocol in the sham condition was the same as that in the real condition, except that the coil was positioned with the lateral edge of one wing touching the scalp at 90°(Pascual-Leone et al., 1996). All participants were asked if they had any symptoms of discomfort after the stimulation.

The cognitive training was completed on an iPad tablet (version 9.1; Apple, USA) with several cognitive tasks selected by an experienced cognitive therapist. The memory tasks were completed during rTMS stimulation, while the other tasks, including attention tasks, mathematical calculations, agility drills, language tasks, and logic thinking tasks, were practiced after the stimulation ended (Zhang et al., 2019). Please refer to our previous article for detailed information.

The MRI data were acquired using a 3T scanner (Discovery 750, GE Healthcare) using a 32-channel head coil. To minimize head motion and reduce scanner noise, tight but comfortable foam padding and earplugs were used. During the resting-state fMRI scan, participants were instructed to close their eyes, not fall asleep, and not think of anything in particular. Gradient-echo echo-planar imaging (EPI) sequence was used to acquire fMRI images with the following parameters: repetition time/echo time (TR/TE) = 2000/30 ms, flip angle = 90°, matrix = 64 × 64, field of view = 240, slice thickness = 4 mm, interleaved acquisition, 29 axial slices in total, and 240 timepoints. A sagittal T1-weighted structure image was also acquired using a 3D brain-volume (3D-BRAVO) sequence with the following parameters: TR/TE = 8.2/3.2 ms, TI = 450 ms, slice gap = 1 mm, matrix = 256 × 256, flip angle = 8°, and voxel size = 1 × 1 mm.

The resting-state fMRI data were preprocessed using SPM8 (http://www.fil.ion.ucl.ac.uk/spm/). The first 10 timepoints of the rs-fMRI data were discarded, leaving the remaining 230 timepoints for slice timing and motion correction (threshold: translational or rotational motion parameters lower than 2 mm or 2°). The normalization included the following steps: (1) structural images were linearly coregistered to the mean functional image; (2) the transformed structural images were segmented, and then the gray matter was non-linearly coregistered to the Montreal Neurological Institute (MNI) space; and (3) the motion-corrected functional volumes were normalized to the MNI space using the parameters estimated during the non-linear coregistration. The functional images were resampled into a voxel size of 3 × 3 × 3 mm³ and then smoothed using a Gaussian kernel of 8 × 8 × 8 mm³ full width at half-maximum (FWHM). Finally, functional images were band-pass-filtered with a frequency from 0.01 to 0.1 Hz, and several nuisance covariates (24 head motion parameters, averaged signal from white matter, cerebrospinal fluid, and global signal) were regressed out by performing multiple linear regression analysis.

A 6 mm radius sphere centered on stimulation targets, that is, the left DLPFC and the left LTL, was used as the seed region. The coordinates were the same as those used for the rTMS treatment. Bivariate first-order coefficient-based voxel-wise GCA was performed to explore the influence of the stimulation target on the rest of the brain using the REST toolbox (http://www.restfmri.net).

Resting-state fMRI networks were identified by using the Group ICA program of the fMRI toolbox (http://www.nitrc.org/projects/cogicat/). Seven classic networks, including a visual network, a sensorimotor network, a dorsal attention network, a salience/ventral attention network, a limbic network, a control network, and a default mode network (DMN), were used as the networks of interest, which were obtained by a clustering approach across the cerebral cortex using resting-state functional connectivity MRI from 1,000 healthy subjects (Yeo et al., 2011). The generated network components were exported to calculate the inter-network effective connectivity.

The normal distribution of the data was determined by performing the Kolmogorov–Smirnov test. The χ² test or Fisher's exact test was performed to compare the categorical variables. An independent t-test or the Mann–Whitney test was performed for the quantitative parameters. EC changes at the voxel and network level were analyzed using a flexible factorial design installed in SPM12. We considered a 2 × 3 factorial design with “time” as the within-subjects factor (three levels: T0, T4, and T8) and “group” as the between-subjects factor (two levels: real rTMS + COG and sham rTMS + COG). If the group × time interaction was significant, post-hoc analysis was performed to explore the simple effects of time and group with the least significant difference (LSD) correction. All behavioral measures were analyzed using the IBM Statistical Package for the Social Sciences (SPSS) version 25, and graphs were made in GraphPad Prism 8. The relationship between EC changes and behavioral changes before and after the intervention was explored by performing a partial correlation analysis. Sex, age, education level, CDR, and disease duration were set as covariates in all tests.

In general, the left DLPFC seed from the GCA analysis showed significant causal outflow to five clusters (p < 0.05, FWE corrected), that is, the left inferior parietal lobe (IPL), the left superior frontal gyrus (SFG), the left postcentral gyrus (PoCG), the left thalamus, and the right parahippocampus (Figure 2A). However, the reverse influence after rTMS-COG treatment was not significant (p > 0.05).

Specifically, a significant group × time interaction effect (F = 14.057, p = 0.000) was observed for EC changes from the left DLPFC to the left IPL. Post-hoc analysis indicated that, compared with the baseline, high-frequency rTMS + COG produced a significant reduction in EC at timepoints T4 (p = 0.010) and T8 (p = 0.001), while sham rTMS + COG produced a significant increase in EC at timepoint T8 (p = 0.024) (Figure 2B, line chart). The EC changes differed significantly between the real and sham groups at T8 (p = 0.001) but not at T4 (p = 0.115) (Figure 2B, bar chart).

In terms of EC changes from the left DLPFC to the left SFG, a significant time × group effect was also identified (F = 10.766, p = 0.001). Compared with the baseline, high-frequency rTMS + COG produced a significant reduction in EC at timepoints T4 (p = 0.000) and T8 (p = 0.000) (Figure 2C, line chart). The EC changes differed significantly between the real and sham groups at both T4 (p = 0.003) and T8 (p = 0.002) (Figure 2C, bar chart).

For EC changes from the left DLPFC to the left PoCG, only the time main effect was significant (F = 4.053, p = 0.035). High-frequency rTMS + COG produced a significant reduction in EC at timepoint T8 (p = 0.013) but not at timepoint T4 (p = 0.164) (Figure 2D, line chart). The EC changes differed significantly between the real and sham groups at T8 (p = 0. 013) but not at T4 (p = 0.285) (Figure 2D, bar chart).

The EC changes from the left DLPFC to the left thalamus also showed a significant time main effect (F = 3.579, p = 0.049). The real rTMS + COG produced a significant reduction in EC at timepoints T4 (p = 0.001) and T8 (p = 0.024) (Figure 2E, line chart). The EC changes differed significantly between the real and sham groups at both T4 (p = 0.046) and T8 (p = 0.025) (Figure 2E, bar chart).

The EC changes from the left DLPFC to the right parahippocampus also showed a time main effect (F = 5.704, p = 0.012). Compared with the baseline, the high-frequency rTMS + COG produced a significant reduction in EC at timepoints T4 (p = 0.017) and T8 (p = 0.047), while the results of the sham group were not significant at either T4 or T8 (Figure 2F, line chart). The EC changes between the real and sham groups also showed no difference at either T4 (p = 0.247) or T8 (p = 0.140) (Figure 2F, bar chart).

The left LTL seed from the GCA analysis revealed two clusters with significant causal outflow (p < 0.05, FWE corrected), that is, the left orbital frontal cortex (OFC) and the right superior temporal gyrus (STG) (Figure 3A), with no significant reverse influence (p > 0.05). The EC changes from the left LTL out to the left OFC showed significant time × group interaction (F = 4.198, p = 0.032). Compared with the baseline, the high-frequency rTMS + COG produced a significant reduction in EC at timepoints T4 (p = 0.030) and T8 (p = 0.037). The sham group showed no significant difference at T4 (p = 0.072), but it produced a significant increase in EC at T8 (p = 0.024) (Figure 3B, line chart). The EC changes between the real and sham groups differed significantly at T8 (p = 0.008) but not at T4 (p = 0.491) (Figure 3B, bar chart).

In terms of the EC changes from the left LTL to the right STG, the time main effect was significant (F = 7.442, p = 0.016). The real rTMS + COG produced a significant reduction in EC at timepoint T4 (p = 0.004), but it showed no significant changes at T8 (p = 0.232). The sham group showed no significant changes at either T4 (p = 0.155) or T8 (p = 0.159) (Figure 3C, line chart). The EC changes between the real and sham groups showed no significant difference at either T4 (p = 0.138) or T8 (p = 0.121) (Figure 3C, line chart). Table 2 summarizes the voxel-level results of EC changes for directional influence to and from the left DLPFC and left LTL seeds.

For pairwise EC changes of the seven networks, a significant time × group interaction was only observed from the limbic network out to the default mode network (DMN) (F = 7.070, p = 0.005, Figure 4A). Compared with the baseline, the high-frequency rTMS + COG induced no significant EC changes at T4 (p = 0.231), but it produced a significant increase at T8 (p = 0.034). The sham group showed no significant EC changes at T4 (p = 0.381), but it showed a significant decrease at T8 (p = 0.014) (Figure 4B). The EC changes between the real and sham groups showed significant differences at T8 (p = 0.005) but not at T4 (p = 0.255) (Figure 4C).

For voxel-level EC changes, partial correlation analysis showed that the EC changes from the left DLPFC out to the left IPL, left SFG, left PoCG, left thalamus, and right parahippocampus were all positively correlated with ADL changes at T4 or T8 (P < 0.05), indicating that the reduction in EC induced by the real rTMS intervention improved the ability to perform activities of daily living. In particular, the decrease in EC from the left DLPFC out to the right parahippocampus was associated with a decrease in ADAS-cog score 4 weeks after the end of the treatment (T8), indicating that the reduction in EC from the left DLPFC out to the right parahippocampus was associated with improved cognitive function in patients with mild to moderate AD in the long term. For the network-level EC changes, the increase in EC from the limbic network out to the DMN at T8 was correlated with an increase in MoCA score after rTMS-COG treatment. Table 3 shows the partial correlation analysis results.

In our study, rTMS-COG therapy induced a decrease in EC from the left DLPFC out to the left IPL and from the left LTL out to the left OFC, while cognitive training alone (sham condition) induced a significant increase in EC in the long term (T8). rTMS stimulation on the left DLPFC has been observed to have definite benefits for patients with mild AD (Di Lazzaro et al., 2021). As an important node of the ECN and the FPN (Vincent et al., 2008), the left DLPFC has been mostly used and has been proven to be a beneficial stimulation target for AD (Cotelli et al., 2008, 2011; Ahmed et al., 2012). The IPL related to bottom-up attention and episodic memory is an important node of the ECN (Xiao et al., 2022). Information flow from the IPL to the anterior DMN subsystem has been reported, along with extensive connections between the parietal cortex and the frontal cortex (the parieto-frontal circuit) (Buckner et al., 2009; Luo et al., 2019). Our study found a significant decrease in EC from the left DLPFC to the ipsilateral IPL, associated with improved ADL score both in the short and long term (T4 and T8). This is also in line with previous observations that the TMS of the left DLPFC could selectively modulate functional connectivity both within and between the ECN and DMN in patients with depression (Liston et al., 2014). rTMS over DLPFC modulated the coupling between the ECN and the DMN in healthy people and in heroin-dependent individuals (van der Werf, 2010; Chen et al., 2013; Jin et al., 2022). Stimulation of the left LTL with transcranial direct current stimulation (tDCS) improved recognition memory in AD patients (Ferrucci et al., 2008). Similarly, in our study, a decrease in EC from the left LTL out to the ipsilateral OFC improved episodic memory in the short term (T4) and the ability to perform activities of daily living in the long term (T8).

For EC changes from the left DLPFC out to the left SFG, rTMS-COG therapy induced a significant decrease in EC in both the short and long terms, with no significant changes for cognitive training (sham condition). Causal interactions between the frontoparietal central executive and the DMN were well identified in a TMS/fMRI study (Chen et al., 2013). Since the FPN covers several frontoparietal areas, including the SFG (Vincent et al., 2008), our study showed that rTMS targeted on the left DLPFC also induces information flow within the FPN, resulting in benefits for ADL in both the short and long terms.

rTMS is effective for increasing and decreasing (low-frequency: decrease, high-frequency: increase) functional coherence within the prefrontal-limbic network (Riedel et al., 2019). rTMS reportedly induces hypoconnectivity within the DMN in patients with amnestic mild cognitive impairment (aMCI)(Cui et al., 2019). However, the inter-network causal effect after multisite rTMS-COG treatment has yet to be expounded. In our study, compared with the cognitive training group (the sham condition), high-frequency rTMS-COG treatment induced a significant increase in EC from the limbic network out to the DMN at the 4-week follow-up after treatment (T8), indicating that multisite TMS-COG therapy had a long-term effect at the network level. Furthermore, the inter-network increase in EC was associated with cognition improvement (MoCA score increase).

It is particularly noteworthy that, although the cognitive training tasks were the same in the real rTMS and sham groups in our study, we could not ascribe the improvement in clinical outcomes in the real group to rTMS itself. It has been reported that TMS may have a synergistic effect with cognitive training (Rabey et al., 2013). Interlaced with cognitive training, rTMS has additional beneficial effects (Bentwich et al., 2011). Our results may aid the understanding of the neurobiological effects of rTMS-COG therapy on a macro-scale.

Our study has several limitations. First, due to the longitudinal design, the sample size and statistical efficiency were limited. Therefore, we cannot rule out the possibility that some of the changes we observed after rTMS-COG were not causally related to the treatment. Second, AD patients were recruited based on the NINCDS-ADRDA criteria instead of the amyloid/tau/neurodegenaration (ATN) diagnostic framework. These criteria were chosen because the cerebral spinal fluid (CSF) biomarker and molecular neuroimaging with positron emission tomography (PET) are usually unavailable in clinical practice due to their “invasive” characteristics. The heterogeneity across participants may have influenced the results, so our findings could be considered exploratory. When interpreting the findings of the present study, a great deal of caution is needed. A multi-center cohort study is needed to validate our pilot study.