Network-targeted transcranial magnetic stimulation (TMS) for mild cognitive impairment (MCI)

Highlights

• •We identified large-scale brain networks that are likely to be targeted by transcranial magnetic stimulation (TMS) therapy for MCI-AD.
• •The prefrontal TMS seed showed connectivity with the salience network primarily.
• •The parietal TMS seed showed connectivity primarily with the default network.
• •The parietal TMS map showed overlap with the frontoparietal and language networks as well.

Abstract

Background

Transcranial magnetic stimulation (TMS) is a promising non-pharmacological intervention for treatment of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD). Yet, we know little about precisely where stimulation would be ideal to improve cognitive function.

Objective

To examine the network functional connectivity (fc) characteristics of prefrontal and parietal stimulation sites, given that these sites have led to improved cognitive function in TMS studies involving MCI-AD and unimpaired participants.

Methods

Resting-state functional MRI data were acquired from 32 MCI participants at the baseline visit of an ongoing TMS trial and used to compute connectivity with prefrontal and parietal stimulation locations, selected on the basis of previous TMS studies. The TMS seed maps were examined for extent of spatial overlap with eight canonical networks. After identifying the network most likely to be targeted by TMS, we applied strategies that may provide purer targeting. Finally, we examined network connectivity in relation to participants’ behavioral characteristics because of the potential for TMS treatment to be personalized.

Results

The prefrontal TMS seed map overlapped primarily with the salience network. The prefrontal site is also notable for its anti-correlated connectivity with the AD-vulnerable posterior cingulate cortex (PCC). The parietal TMS seed map showed the expected strong positive connectivity with the PCC and other default network regions. Nonetheless, this particular parietal site may simultaneously modulate the fronto-parietal network. Strategies to improve network targeting and to personalize TMS are reported as secondary findings.

Conclusion

These results can be applied to network-targeted brain stimulation for MCI and early AD treatment. Greater precision and personalization of TMS offer the promise of achieving better outcomes for individuals with MCI or mild AD dementia.

1. Introduction

Currently there are no treatments for patients with mild cognitive impairment (MCI) that have well accepted clinically meaningful benefits (Ebell et al., 2024, Fitzpatrick-Lewis et al., 2015). Studies of non-invasive brain stimulation techniques, such as repetitive transcranial magnetic stimulation (rTMS) and theta-burst TMS, have shown promising improvements in cognitive function in studies of MCI and mild to moderate Alzheimer’s disease (AD) dementia. Recent meta-analyses and reviews of the effects of rTMS for MCI-AD reported moderate to large effect sizes for improvement of memory and global cognitive function (Jiang et al., 2021, Koch et al., 2024, Li et al., 2024, Wei et al., 2022).

Two popular cortical regions for TMS stimulation are the prefrontal and the parietal cortices. There is not yet enough evidence to recommend targeting TMS to one region vs. the other for treatment of MCI-AD because of differences in treatment duration and methodology. Compared to the prefrontal TMS randomized clinical trials, the parietal TMS trials typically did not provide as many TMS sessions to achieve a robust improvement in the active TMS group (Li et al., 2024). As described below, roughly half of prefrontal TMS trials used coarse methods of coil placement (Bagattini et al., 2021, Trapp et al., 2020) that lead to greater variability in response to TMS (Siddiqi et al., 2020). In addition, coarse placement methods are likely to impede the challenging work of elucidating therapeutic mechanisms and distinguishing among alternative hypothesized mechanisms.

Since the publication of the first randomized clinical TMS trial for AD in 2012, eight of the prefrontal studies (about half) used scalp measurements such as the 5-cm rule for coil placement (Ahmed et al., 2012, Cheng et al., 2022, Cui et al., 2019, Drumond Marra et al., 2015, Padala et al., 2020, Padala et al., 2018, Saitoh et al., 2022, Wu et al., 2015). Scalp-based methods are likely to cause greater variability in response due to variation in patients’ head sizes. Three prefrontal studies used the less coarse “F3/F4” method, in which coil placement is based on the electrode positions of the 10–20 EEG system (Bagattini et al., 2020, Tao et al., 2022, Turriziani et al., 2019). Five recent studies used T1-MRI neuronavigation to a specific coordinate mapped to the dorsolateral prefrontal cortex (DLPFC) (Li et al., 2021, Lu et al., 2022, Moussavi et al., 2024, Wu et al., 2022, Wu et al., 2024).

In contrast, none of the parietal TMS clinical trials employed scalp measurements and few used the EEG system (Hu et al., 2022). The majority (four of seven) used T1-MRI neuronavigation to locate either the precuneus region (Koch et al., 2018, Koch et al., 2022) or the left angular gyrus (Chen et al., 2023, Liu et al., 2022). In two other trials, neuronavigation was individualized by using single-subject fMRI scans (Jia et al., 2021, Jung et al., 2024). In every one of the parietal TMS trials for MCI-AD, the goal of TMS was to stimulate a key cortical node of the posterior default network. Targeting the default network has a compelling disease-based rationale, as this network is the first to show AD abnormalities (Palmqvist et al., 2017).

Historically, the randomized prefrontal TMS trials have not articulated a particular network or network node as the target. A rationale for prefrontal stimulation has been to increase synaptic plasticity and transmission in general (Ahmed et al., 2012, Li et al., 2021, Wu et al., 2024). Two studies noted that the dorsolateral prefrontal cortex (DLPFC) is a core region of the lateral frontoparietal network (FPN) (Bagattini et al., 2020, Wu et al., 2024). Yet, there is evidence that the DLPFC region is functionally heterogenous with multiple networks coursing through this region (Dixon et al., 2018, Vincent et al., 2008). It follows that a multiplicity of coil-positioning methods can make it challenging to interpret the findings of TMS trials. One model of DLPFC TMS examined 11 coil positions for treating depression (Opitz et al., 2016). The model predicted the stimulated network(s) at a given coil location and orientation by relating resting-state fMRI (rs-fMRI) data from the Human Connectome Project (HCP) to anatomically realistic finite-element models of the human head. Opitz et al. (2016) found that the 5-cm “average” scalp site and the EEG-based F3 sites fell predominately in the default network, and that the coordinates proposed by Fox et al. (2012) for TMS treatment of depression showed overlap with the frontoparietal and ventral attention networks--characterized by Yeo et al. (2011).

To advance TMS as a treatment for MCI and AD, we need to implement reproducible neuroscience-based TMS methods and recruit larger numbers of participants into controlled clinical trials (Jiang et al., 2021, Sanches et al., 2020). We think that bringing prefrontal and parietal TMS into a common framework of network-targeted TMS and working toward more precise coil placement methods can help. Deliberate targeting of distinct networks would encourage hypothesis-driven behavioral outcomes and related biomarkers of response. A common conceptual framework and harmonized methods would facilitate comparisons between parietal and the prefrontal TMS in terms of their efficacy and their therapeutic mechanisms. Such efforts may ultimately help advance TMS as an evidence-based treatment for MCI and AD.

A functional connectivity-based strategy for network-targeted TMS. Research using rs-fMRI data has defined multiple large-scale networks (Uddin et al., 2019) that show abnormalities in MCI-AD and neuropsychiatric populations in the strength of functional connectivity within and between networks. The conditions of MCI, AD, and major depressive disorder are all associated with abnormal connectivity within the default and salience networks (Javaheripour et al., 2021, Menon, 2011, Yu et al., 2021). Fox and colleagues examined the effectiveness of brain stimulation across several conditions including depression and AD (Fox et al., 2014). They found, for each disease, that the efficacious sites for deep brain stimulation and efficacious sites for noninvasive stimulation were within the same functional network as defined by Yeo et al. (Yeo et al., 2011). Moreover, ineffective sites fell outside the network. Fox et al. (2014) concluded that a functional connectivity strategy may be useful, not only for translating therapy between deep brain and noninvasive stimulation modalities, but also more broadly useful for implementing a perspective of targeted brain network modulation.

In this paper, we apply a functional connectivity approach to TMS for MCI-AD. We analyzed the pre-treatment rs-fMRI data from MCI participants. We had three aims: First, we sought to identify a large-scale network that can be targeted by DLPFC TMS, which has not been explicitly undertaken in TMS applications for MCI-AD. Given the functional heterogeneity of the DLPFC region (Cieslik et al., 2013, Opitz et al., 2016), we selected one prefrontal TMS site—the depression “efficacy” site (Fox et al., 2012). We predicted that TMS to this site targets the salience network based on: 1) the site’s location on an atlas of canonical functional networks (Whitfield-Gabrieli and Nieto-Castanon, 2012); 2) evidence that the salience network includes regions of the lateral prefrontal cortex (Touroutoglou and Dickerson, 2019, Uddin et al., 2019); 3) the previously mentioned modeling results of Opitz et al. (2016); and 4) a proposed taxonomy of macro‑scale functional networks, in which the salience network includes the ventral attention network (Uddin et al., 2019). In a complementary fashion, we mapped out network(s) that can be targeted by lateral parietal cortical (LPC) TMS. Although it has been previously shown that LPC TMS can modulate the cortical-hippocampal subsystem of the default network in healthy younger adults (Wang et al., 2014), our connectivity and behavioral data allowed us to meet our two secondary aims: to offer connectivity-based strategies toward improving the precision of TMS and for potentially personalizing TMS therapy for older adults with MCI-AD. Greater TMS precision and personalization hold promise for realizing better outcomes in older adults with MCI or mild AD dementia.

2. Materials and methods

Fig. 1 provides an overview of the methodological approach. An outline of the participant-selection criteria and behavioral measures are shown in the left column. Neuroimaging acquisition and fMRI preprocessing steps are summarized in middle column. A synopsis of the analyses conducted to address the primary aim and the two secondary aims is shown in the right-hand column.

Fig. 1.

Overview of Methods: Participants, Measures, and Analyses for each Aim.

2.1. Participants

Data of 32 participants (16 male, average age 75.9 ± 6.9 years) who met clinical criteria for amnestic MCI were analyzed in this study (see Table 1). For characterization of amnestic MCI, we used the Alzheimer’s Disease Neuroimaging Initiative (ADNI-3) classification criteria (Weiner and Petersen, 2016), which include: subjective memory concern; below normal memory function documented by scoring below the education-adjusted cutoffs on the Logical Memory II subscale (Delayed Paragraph Recall, Story A) from the Wechsler Memory Scale–Revised (WMS-R) (Wechsler, 1987); Mini-Mental State Exam score between 24 and 30 inclusive (Folstein et al., 1975); Clinical Dementia Rating of 0.5 (Morris, 1993), Geriatric Depression Scale (GDS) total score less than 6/15 (Sheikh and Yesavage, 1986), Modified Hachinski Ischemic score less than or equal to 4 (Rosen et al., 1980); no diagnosis of a non-AD neurological disease; no recent history of alcohol or substance abuse; and no diagnosis of dementia made by the study physician during the screening process. As in ADNI-3, all participants were required to have a study partner whose primary role was to provide collateral information about the participant's daily functioning, as well as noticeable changes in the participant’s health status during their participation in the study.

Table 1.

Demographics, clinical, behavioral,and fMRI in-scanner motion variables. All 32 participants had a clinical diagnosis of amnestic mild cognitive impairment (MCI).

Measure	Mean (SD)	Range
Age (years)	75.9 (6.9)	62–88
Sex	16F; 16 M	−
Education (years)	16.9 (2.1)	12–20
Logical Memory Delayed Recall, Story A (/25)	7.6 (2.8)	2–14
CVLT-II composite score (/16)	8.8 (3.2)	2.3–14.1
Trail Making Part B raw score (sec)	99.1 (63)	37–300
Trail Making Part B (log-transformed, reversed)	2.3 (0.50)	1.0–3.1
Geriatric Depression Scale self-rating (/16)	2.2 (1.4)	0–5
fMRI: mean framewise displacement (mm)	0.13 (0.08)	0.04–0.43

Abbreviations: CVLT-II, California Verbal Learning Test-II, Second edition

Notes: Trail Making Part B raw scores were log-transformed and reversed-scored so that higher values indicate better performance; transformed scores were used in correlational analyses. Two of the 32 participants are missing Trail Making Part B scores. Five participants have unknown values of framewise displacement.

The participants were enrollees of a randomized, controlled clinical trial of rTMS for MCI (NCT03331796). No participant had a known history of seizures or current major depressive disorder. The trial’s full inclusion/exclusion criteria can be found in the published protocol (Taylor et al., 2019). Participants provided written informed consent in accordance with the guidelines set by the Institutional Review Board of Stanford University.

2.2. Behavioral measures

For secondary aim B, we examined individual differences in behavioral measures in relation to strength of network connectivity. This aim was motivated by the heterogeneity of MCI and because TMS has the potential to be personalized to target the brain network most correlated with a clinical feature such as memory impairment, executive dysfunction, or loss of initiative. We selected four clinical assessments that are frequently used in MCI studies: the GDS; the WMS-R Logical Memory II Delayed Recall test (Wechsler, 1987); the California Verbal Learning Test, Second Edition (CVLT-II) (Delis et al., 2000); and the Trail Making test (Partington and Leiter, 1949, Reitan, 1958). The GDS and Story A of the WMS-R were administered at the time of screening for eligibility. The CVLT-II and the Trail Making test (Partington and Leiter, 1949, Reitan, 1958) were administered at the baseline visit and on the same day as the neuroimaging session.

Scoring of CVLT-II and Trail Making Part B tests. The CVLT-II is a word-list task that yields several indices of the ability to learn and remember verbal material (Stricker et al., 2002). The CVLT-II List A contains 16 words, with 4 words each from four semantic categories. The task procedures include 5 learning trials, 2 assessments of delayed free recall, and 2 assessments of cued recall. In view of the fact that the subscores on the learning and recall components were highly intercorrelated (r’s > 0.8) and the study’s modest sample size, we computed a summary measure of CVLT-II performance. It was the average of the three CVLT-II components: the mean number of words correctly recalled during the 5 learning trials, the mean correct on delayed free recall, and the mean correct on cued recall. The raw Trails B score is the time in seconds to connect a series of 25 circles by alternating between numbers and letters in order. Trails B is thought to require cognitive set-shifting as well as visual-scanning skills and processing speed (Strauss et al., 2006). The distribution of completion times was positively skewed (2.30) as is often the case other samples (Touroutoglou et al., 2012, Touroutoglou et al., 2018). The Trails B scores used for analysis were log transformed and reversed so that higher scores indicated better performance.

2.3. Neuroimaging

MRI acquisition and preprocessing. Neuroimaging data were collected using a 3 Tesla scanner (GE Healthcare, Chicago, Illinois) with a 32-channel head coil. Structural MRI data were acquired using a 3D T1-weighted gradient echo sequence [sagittal T1-weighted inversion-recovery-prepared, fast spoiled gradient recalled BRAin VOlume (BRAVO) imaging, TR = 8.2 ms; TE = 3.2 ms; flip angle = 9°, 1.0 mm isotropic voxels]. Whole-brain rs-fMRI data were acquired using a simultaneous multi-slice (SMS) acquisition gradient-echo echo-planar research sequence [“muxarcepi” developed by Kerr et al., CNI, Stanford University] using an SMS factor = 6. Other acquisition parameters were: 7 axial slices, TR = 490 ms, TE = 30 ms, flip angle = 45°; freq FOV = 22.2 and freq x phase = 74x74, which yielded 6 x 7 = 42 slices with a voxel size of 3 mm³. The functional scan was 8.9 min long and yielded 1,095 volumes. It was preceded by a 6 sec single-band scan and a 10 sec reversed phase encoding fMRI calibration scan that was used for geometric distortion correction. During the functional scan, participants were instructed to “close your eyes and do not think about any particular thing.” Measures were taken to ensure participants remained awake during the functional scan. These are described in the Supplementary material, Text 1. Head motion was minimized using head restraints and padding.

Preprocessing of the imaging data was performed using both fMRIPrep 1.2.5 (Esteban et al., 2019) (RRID:SCR_016216) and the Conn Toolbox, release 18.b (Whitfield-Gabrieli and Nieto-Castanon, 2012) (RRID:SCR_009550 www.nitrc.org/projects/conn). Details are provided in the Supplementary material, Text S2. In brief, each T1-weighted volume was corrected for intensity non-uniformity and skull-stripped using fMRIPrep. Brain surfaces were reconstructed using recon-all from FreeSurfer 6.0.1. The brain-mask estimate was refined with a custom variation of the method to reconcile ants-derived and FreeSurfer-derived segmentations of the cortical gray-matter of Mindboggle (RRID:SCR_002438). Spatial normalization to the MNI-152 template was performed through nonlinear registration using antsRegistration using brain-extracted versions of the T1w volume and template. For tissue segmentation into cerebrospinal fluid (CSF), white-matter (WM) and gray-matter tissue maps, this version of fMRIPrep invoked fast (FSL v5.0.9). Preprocessing of fMRI data via fMRIPrep also included motion correction, susceptibility distortion correction, co-registration to T1, normalization to standard space, extraction of physiological noise regressors (using CompCor), and calculation of framewise displacements. Further preprocessing steps using Conn were removal of the initial 6 timepoints, spatial smoothing using a 4-mm FWHM Gaussian kernel, and ART-based detection of outliers. In the denoising step, linear regression was used to achieve scrubbing of outliers and to reduce the influences of: global, WM, and CSF BOLD signal; framewise displacement and global signal fluctuations; and subject motion (derived from six realignment parameters). Denoising was followed by band-pass filtering to select temporal frequencies in the range of (0.008 Hz < f < 0.09 Hz).

2.4. Functional connectivity analyses

2.4.1. Primary aim – Identification of large-scale networks

We focused on two stimulation sites—a prefrontal and a parietal site—and conducted seed-to-voxel whole-brain functional connectivity analyses to identify large-scale networks that can be targeted using TMS. We used the baseline rs-fMRI data of study participants who were scanned between February 2019 and March 2022. Participants in the trial of the efficacy of rTMS to improve memory in MCI were randomized to one of three groups: (1) Active DLPFC rTMS; (2) Active LPC rTMS; or (3) Sham rTMS (Taylor et al., 2019). Active DLPFC rTMS was targeted bilaterally to Montreal Neurological Institute (MNI) coordinates −38, 44, 26 and + 38, 44, 26. Participants in the active LPC arm received bilateral stimulation targeted to MNI −47, −68, 36 and MNI + 47, −68, 36. (Participants in the sham arm were equally distributed to the two cortical locations.) The left DLPFC coordinate is the depression “efficacy” site identified by Fox and colleagues (2012), who applied functional connectivity analyses to identify a potentially optimal site of stimulation for treatment of major depressive disorder. The left LPC coordinate was used by Wang and colleagues in an rTMS experiment for enhancing memory performance and hippocampal-cortical functional connectivity in healthy young adults (Wang et al., 2014).

For the functional connectivity analyses, seed regions of interest (ROIs) were constructed by centering a 10 mm radius sphere at each left MNI coordinate. This seed size is typical of the size used in other TMS studies on neuromodulation (Fox et al., 2014, Liston et al., 2014, Vidal-Pineiro et al., 2015). To generate seed-based functional connectivity maps for each TMS site and participant, we computed Pearson's product moment correlations, r, between the mean signal time course of each seed and the time course of all other brain voxels using the Conn Toolbox. The resulting correlation maps were converted to z-values, using Fisher's r-to-z transformation. Group-averaged maps were computed using general linear modeling (GLM); the GLMs included participants’ composite measure of motion “QA_MeanMotion” as a nuisance covariate. The voxel threshold was set at t = 3.71, p < 0.001 uncorrected (two-sided) to identify clusters of interest. To control for multiple comparisons at the cluster level, a False Discovery Rate (FDR) of p < 0.05, based on nonparametric permutation computations, was used to identify clusters with significant positive connectivity (minimum size of 60 voxels). All data processing, calculations, and thresholding were performed in Conn in standard MNI space.

To relate the two group-averaged L-DLPFC and L-LPC TMS seed maps (hereafter referred to as TMS maps) to canonical networks, we used the Conn Toolbox eight-network atlas. (The network atlas is based on a dataset of the Human Connectome Project (n = 497) and was derived using independent-component analysis https://web.conn-toolbox.org/resources/conn-in-pictures (Whitfield-Gabrieli and Nieto-Castanon, 2012).) The TMS maps are displayed on 3-D cortical surfaces of the brain. We also determined the spatial extent to which each TMS map overlapped with canonical networks to estimate the extent of “on-network” modulation in relation to the hypothesized network target, and identify the potential for “off-network” modulation. For each canonical network, we report the percentage of Conn-network voxels that lie within the TMS maps and the Dice coefficient (a metric of spatial overlap ranging from 0 to 1) (Dice, 1945, Zou et al., 2004).

2.4.2. Secondary aim A – Connectivity-based strategies for improving precision of TMS for treatment of MCI-AD

For refining coil placement, two dimensions can be considered: 1) use of an AD-focused vs. a cross-disciplinary approach and 2) use of individual subject vs. group-averaged connectivity data to identify an optimal TMS site. The two dimensions can be factorially crossed to generate four strategies. Here, we examined two of the four strategies. The first is an AD-focused strategy that uses individual connectivity data.

A.1 Use of knowledge about AD and individual connectivity data to inform decisions about coil placement. Functional connectivity and metabolism of the ventral posterior cingulate cortex (vPCC) region has been frequently identified as abnormal in AD (Lau et al., 2016, Mutlu et al., 2016). Accordingly, the vPCC region is a logical target for treatment of patients with MCI and mild dementia due to AD. To implement an AD-focused strategy using individual subjects’ connectivity data, we constructed a 4 mm spherical seed ROI centered on MNI 6, −52,16. This vPCC foci was identified in a meta-analysis of 21 rs-fMRI studies (Lau et al., 2016, Mutlu et al., 2016). Then for each MCI participant, we located the voxel within the left LPC that showed the strongest functional connectivity with the vPCC seed. The major steps were: (1) vPCC seed-based connectivity maps were calculated for each participant; (2) the search area for TMS coil localization was masked by the Brodmann anatomical atlas and constrained to the left LPC functional ROI of the Conn default network; (3) the MNI coordinate of the voxel with the highest correlation value was identified. The Yale BioImage Suite Web tool was used to identify the Brodmann area and cortical label corresponding to the individualized stimulation site (https://bioimagesuiteweb.github.io/bisweb-manual/tools/mni2tal.html). In the results section, the sites are overlaid on a 3-D semi-inflated surface of the brain for a third of the sample and tabulated. To examine the equivalence of individualized vs. group-based approaches, we computed Euclidean distances between participants’ individualized locations and a group-based coordinate such as MNI −47, −68, 36.

A.2 Cross-disciplinary approach: Use of group connectivity data to improve coil placement for TMS treatment of MCI-AD. Large-sample, connectome-quality data are advantageous for identifying promising TMS sites because they provide robust estimates of connectivity (Fox et al., 2012, Weigand et al., 2018). Moreover, advanced imaging is not routinely acquired in clinical TMS settings; the use of scalp measurements for coil placement gives rise to substantial variability in what part of the prefrontal cortex is stimulated. Fox and colleagues have shown that group-based estimates of functional connectivity can predict overall improvement and the pattern of symptom improvement following TMS treatment of major depressive disorder (Cash et al., 2021a, Cash et al., 2021b, Fox et al., 2012, Weigand et al., 2018). Specifically, the stronger the anticorrelated connectivity (i.e. negative connectivity) between the stimulation site and the subgenual cingulate, the better the response to TMS. Researchers who seek to advance TMS treatment of MCI-AD could apply knowledge gained in TMS for depression and test a cross-disciplinary approach. Interestingly, functional connectivity between regions within the lateral prefrontal cortex and the PCC are anticorrelated (Greicius et al., 2003) in a way that might be analogous to anticorrelated connectivity between the lateral prefrontal cortex and the subgenual cingulate. To offer an example of how a cross-disciplinary, group-based connectivity approach could be employed in TMS for MCI-AD, we turned to the Neurosynth database (Yarkoni et al., 2011), which hosts a group-average connectivity map based on rs-fMRI data of 1000 healthy subjects (Yeo et al., 2011). In the results, we report MNI coordinates within the left prefrontal cortex that are most strongly anticorrelated with the vPCC per Neurosynth.

2.4.3. Secondary aim B–Network connectivity, behavioral features, and personalized TMS

For the examination of linkages between abnormal network connectivity and behaviors that can potentially be addressed by personalized TMS therapy, we began by computing summary measures of the strength of functional connectivity within the salience and default networks. The selection of the salience and default networks is contingent upon the results of the Primary Aim; namely, we expected to observe strong connectivity of the salience network with the DLPFC stimulation site and strong connectivity of the default network with the LPC stimulation site.

Analytical approach for computation of summary measures of network connectivity. To compute summary measures of network connectivity, we first computed pairwise ROI-to-ROI connectivity scores, and then conducted principal component analyses (PCA) to guide the aggregation of scores into a smaller set of relatively non-redundant summary measures. We sought to generate summary scores that were functionally meaningful in relation to TMS, while recognizing that the size of our sample was modest. Therefore, rather than computing scores for all possible ROI-to-ROI pairs, we elected to compute ROI-to-ROI connectivity pairs between a network hub and other key regional nodes within that network.

Selection of hub ROIs. For the default network and salience hubs, we chose ROIs located in the ventral PCC and dorsal anterior cingulate/mid cingulate cortex (dACC/aMCC) respectively. The 4 mm spherical vPCC ROI was centered on MNI 1, −55, 17 (Vincent et al., 2008). This specific vPCC coordinate has been used in mechanistic TMS research, which found individual differences among healthy young participants in plasticity changes following TMS (Vidal-Pineiro et al., 2015). This MNI coordinate and nearby coordinates have been used as a vPCC seed ROI in studies on AD, MCI and “super” successful aging (Brier et al., 2012, Sheline and Raichle, 2013, Zhang et al., 2020).The dACC/aMCC ROI was a 4 mm radius spherical ROI centered on MNI 0, 22, 35, which is the peak coordinate of the ACC label of an ICA-derived salience network (n = 497; (Whitfield-Gabrieli and Nieto-Castanon, 2012)). Nearby ACC coordinates have been used in mechanistic and in biomarker-focused rs-fMRI studies of DLPFC TMS (Mitra et al., 2023, Tik et al., 2017). In the biomarker study, individual differences in pretreatment ACC signaling were predictive of response to DLPFC TMS treatment among patients with MDD (Mitra et al., 2023). In addition, the centroid of the dACC/aMCC ROI (MNI 0, 22, 35) that we selected is near the cingulate-centered hub of the salience network (MNI 4, 22,32) described by Touroutoglou and Dickerson (2019).

There are alternative hubs we could have selected for these two networks, such as the bilateral dorsal anterior insulae for the salience network hub (Uddin et al., 2019). Our selection of the vPCC and dACC/aMCC as the hub ROIs reflects a “cingulate-centered network” perspective (Touroutoglou and Dickerson, 2019). In an extensive review of neuroimaging studies, Touroutoglou and Dickerson (2019) highlight the roles of the PCC, the aMCC (aka dACC), subgenual ACC, and pregenual ACC as major network hubs that anchor their pertinent networks and flexibly interact with regions of other large-scale networks. The review also cites evidence that the cingulate hubs are structurally connected with each other. Taken together, a cingulate-centered approach may allow future studies of network-targeted TMS to parsimoniously test for changes in functional connectivity within a targeted network and explore, for example, changes in functional connectivity between network hubs.

Network node ROIs. For the salience network, six node ROIs were defined in the left and right DLPFC/anterior prefrontal cortex (aPFC), left and right anterior insula (aInsula), and supramarginal gyri (SMG). They are functional ROIs, as they are the ICA-derived labels of the CONN network atlas (Whitfield-Gabrieli and Nieto-Castanon, 2012). For the default network, five node ROIs were defined in the left and right LPC, left and right hippocampi, and the medial prefrontal cortex (MFPC). The parietal and prefrontal ROIs were the ICA-derived cortical labels of the CONN network; the LPC and MPFC are recognized core regions of the default network (Uddin et al., 2019). The hippocampal ROIs were defined using the probabilistic Harvard-Oxford Cortical Structural Atlas (RRID:SCR_001476). While the hippocampi are not recognized as nodes of the canonical default network (Uddin et al., 2019), they can be considered part of a subnetwork or fractionated network with respect to the canonical default network (Andrews-Hanna et al., 2010, Braga and Buckner, 2017).

Calculation of functional connectivity scores. ROI-to-ROI connectivity between a network hub and key nodes within the network were calculated using Conn’s ‘ROI-to-ROI’ analysis (Whitfield-Gabrieli and Nieto-Castanon 2012). For the salience network, six ROI-to-ROI connectivity values were calculated for each participant between the dACC/aMCC hub ROI and the six salience node ROIs. For the default network, five ROI-to-ROI connectivity values were calculated for each participant between the vPCC hub ROI and the five default node ROIs.

Aggregation of connectivity scores to create summary measures. A PCA of the six salience scores suggested a 2-factor solution, which accounted for 71 % of the variance. Loading strongly on Factor 1 were the two dACC-to-aInsula scores and the two dACC-to-DLPFC/aPFC scores (loadings were > 0.60). The average of these four scores was used as the summary measure of salience network connectivity. The other two scores (i.e. dACC-to-left-SMG and dACC-to-right-SMG) loaded more strongly on Factor 2 than Factor 1 (> 0.85 vs < 0.30). The two dACC-SMG scores were averaged and treated as a secondary measure of salience connectivity. The PCA of the five default network scores suggested a 1-factor solution, which accounted for 57 % of the variance. The two vPCC-to-LPC scores and the two vPCC-to-hippocampal scores had factor loadings ranging between 0.73 and 0.87. The vPCC-to-MPFC score had a factor loading of 0.66. Even though a loading of this magnitude is important in a strict sense of psychometrics, we did not include the vPCC-to-MPFC in the summary measure of default-network connectivity because the MPFC region is less sensitive to effects of parietal TMS and less vulnerable to early-stage AD than temporal-parietal regions (Jones et al., 2011, Lau et al., 2016, Mutlu et al., 2016, Vidal-Pineiro et al., 2015, Wang et al., 2014). The four vPCC-to-LPC and vPCC-to-hippocampal scores were averaged and used as a summary measure of vPCC-hippocampal-LPC connectivity.

In total, we computed a summary measure of dorsal salience network connectivity and a summary measure of posterior default network connectivity—signifying dACC-aInsula-prefrontal connectivity and vPCC-hippocampal-parietal connectivity respectively. We calculated Spearman correlation coefficients to examine the degree to which the strength of network connectivity was predictive of executive function impairment, verbal episodic memory, and depressive symptoms.

Hypothesized associations. The inter-connected nodes of the posterior default network are considered to play a vital role in the storage, consolidation, and retrieval of episodic memories (Sestieri et al., 2011, Squire et al., 2015, Vincent et al., 2008). In light of evidence that functional connectivity within this network is predictive of memory performance (Sperling et al., 2010, Touroutoglou and Dickerson, 2019), we predicted that stronger connectivity within the posterior default network would be associated with better memory scores. Key putative functions of the salience network include: detection of salient stimuli (Uddin et al., 2019); maintenance of a self-initiated task-set (Dosenbach et al., 2008); and dynamically switching between brain networks that support external and internal modes of attention (Sridharan et al., 2008). The Trail Making-Part B test purportedly assesses visual scanning to identify sequences of stimuli while mentally maintaining a task set of switching between sequences (Delis et al., 2003). In light of evidence that functional connectivity within the dorsal salience network is predictive of Trail Making-Part B performance (Touroutoglou et al., 2012, Touroutoglou et al., 2018), we predicted that stronger connectivity within the salience network would be associated with better Trail Making-Part B performance. The predicted positive correlation coefficients were tested using planned 1-tailed tests of association.

3. Results

3.1. Primary aim. Network targeting for treatment of MCI-AD: Prefrontal and parietal TMS

L-DLPFC TMS map. The L-DLPFC TMS seed region showed extensive connectivity with regions of the salience network (p FDR < 0.05). The highest fc-correlations were observed in the bilateral anterior cingulate and insular cortices, paracingulate gyri, frontal and central operculi, and frontal lobes—including the bilateral frontal poles, inferior frontal gyri (pars opercularis), and left middle frontal gyrus (MFG). The L-DLPFC TMS seed also showed connectivity with the parietal components of the salience network, especially the left anterior division of the supramarginal gyrus (SMG). See Fig. 2A, hot colors.

Fig. 2.

Resting-state functional connectivity maps relevant to (A) left prefrontal vs. (B) parietal transcranial magnetic stimulation (TMS). (A) The left dorsolateral prefrontal cortex (L-DLPFC) TMS seed (centered at MNI −x = -38, y = 44, z = 26) is preferentially connected with the anterior cingulate and other areas displayed as hot colors. (B) The left lateral parietal cortex (L-LPC) seed (centered at MNI x = -47, y = -68, z = 36) is preferentially connected to posterior cingulate, medial prefrontal and other areas shown in hot colors. Regions shown in purple to magenta showed negative temporal correlations with the TMS seeds (aqua dots). N = 32 older adults with Mild Cognitive Impairment; MRI acquired at pre-TMS baseline. Maps are displayed on 3-D cortical surfaces of the brain. Color bars represent t-statistics. FDR-corrected cluster threshold: p < 0.05. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Next we computed the degree of spatial overlap between the L-DLPFC TMS map and the canonical salience network. As shown in Fig. 3A, the centroid of the L-DLPFC TMS seed is located in the salience network atlas. In total, 87 % of the salience network atlas was covered by voxels of the L-DLPFC TMS map (Dice similarity coefficient (DSC) = 0.34). We also assessed overlap of the L-DLPFC TMS map with other networks. It is known, for example, that portions of the salience network are adjacent to the frontoparietal and dorsal attention networks. The L-DLPFC TMS map covered 17 % of the frontoparietal network and 19 % of the dorsal attention network (DSCs = 0.08 and 0.10 respectively). Overlap between the TMS map and the frontoparietal network was observed in the frontal poles bilaterally and the left MFG. Overlap of the L-DLPFC TMS map with the dorsal attention network was observed in the anterior SMG. (Note: A small area of anterior SMG belongs to both the salience and dorsal attention networks in the atlas.) The L-DLPFC TMS map covered less than 4 % of each of the other five networks (DSCs < 0.025). Taken together, this L-DLPFC TMS site showed functional connectivity chiefly with the salience network, and secondarily with the frontoparietal and dorsal attention networks.

Fig. 3.

Functional connectivity maps relevant to (A) left prefrontal and (B) left parietal TMS in relation to two of eight canonical networks provided by the CONN Network Atlas. (A) Overlap (purple) between the L-DLFPC TMS connectivity map (blue) and the Salience Network (red). (B) Overlap (purple) between the L-LPC TMS connectivity map (blue) and the Default Network (red). The TMS maps and network atlas are overlaid on a single-subject T1 template. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

L-LPC TMS map. As shown in Fig. 2B, the L-LPC seed showed connectivity with medial parietal regions (PCC and precuneus), medial frontal, paracingulate, and lateral parietal regions (angular gyri and left posterior SMG). The L-LPC seed also showed connectivity bilaterally with areas of the middle temporal gyri, lateral superior divisions of the occipital lobes, frontal poles, superior frontal and middle frontal gyri, as well as same-hemisphere connectivity with the left hippocampus, inferior temporal gyrus, inferior frontal, and frontal orbital cortices (p FDR < 0.05). Many of these regions are part of the default network.

In terms of spatial overlap, 48 % of the default network atlas was covered by voxels of the L-LPC TMS map (DSC = 0.28; see Fig. 3B). Overlap was highest (54 %) in the left parietal region (where the L-LPC TMS seed is situated). Overlap was lowest (38 %) in the medial prefrontal region of the default network. Next, we assessed overlap of the L-LPC TMS map with other canonical networks. Notably, in the lateral parietal lobe, the default network is contiguous with the frontoparietal and language networks. The L- LPC TMS map covered 32 % of the frontoparietal network and 16 % of the language network (DSCs = 0.12 and 0.06 respectively). Overlap between the TMS map and the frontoparietal network was observed bilaterally in parietal, occipital and frontal regions (specifically, the angular gyri, lateral occipital cortices, caudal end of the MFG, and left frontal pole). Overlap with the language network was observed in the left angular and left middle temporal gyri. Fig. S1 illustrates where the L-LPC TMS map intersects with the default, frontoparietal, and language networks (See supplementary material, Fig. S1). Spatial overlaps with the salience and other networks were less than 1 % (DSCs < 0.01). In summary, this L-LPC TMS site showed functional connectivity primarily with the default network, and secondarily with the frontoparietal and language networks.

Significance of primary aim results. Two takeaways are: (1) as shown by the hot vs. cool colors in Fig. 2A-B, the prefrontal and parietal stimulation sites reveal strikingly opposite, complementary patterns of functional connectivity; (2) our results suggest that when TMS is applied at either of these two sites at the MNI coordinates studied here, TMS might target the frontoparietal network secondarily. Continued refinement of connectivity-based strategies may be helpful toward optimizing coil placement and achieving greater network selectivity.

3.2. Secondary aim a. Connectivity-based strategies to improve the precision of TMS

0A.1. Use of knowledge about AD and individual functional connectivity data to identify a parietal stimulation that shows high connectivity with the AD-vulnerable ventral PCC. Fig. 4 shows individualized stimulation sites (red dots) in relation to the normative group-based site examined in Aim 1 (MNI −47, −68, 36; (Wang et al., 2014)) for 10 representative MCI participants. Of the individualized sites displayed in Fig. 4, six correspond to the angular gyrus (AG), two correspond to visual association cortex (VAC), and two are on the border between the AG and VAC (Table 2). For our MCI sample, the average distance of the individualized sites from the group-based parietal target was 14.25 mm (SD = 5.48, n = 32). This distance is significantly greater than 12 mm (t(31) = 2.32, p = 0.014), a plausible threshold that has been used to model the focal extent of TMS (Bagattini et al., 2021, Fox et al., 2013).

Fig. 4.

Individual vs. normative group connectivity-based coil localization for targeted modulation of the posterior default network via rTMS applied to the left lateral parietal cortex (L-LPC). The aqua dot indicates the location of the group-based L-LPC TMS site under investigation (MNI −47, −68, 36; Wang 2014 Supplementary Materials). The red dots indicate individualized stimulation sites for 10 representative MCI participants. The individualized targets were identified using seed-based connectivity analyses performed on each subject’s rs-fMRI data. The a priori seed for these analyses is in the ventral PCC region of the default network (MNI 6, −52, 16; Lau 2016). The search space for individual targets of peak connectivity was constrained to the CONN Default Network atlas L-LPC region. (Note: Stimulation targets are overlaid on a 3-D semi-inflated surface of the brain for illustration and are not displayed at scale.). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Table 2.

Comparison of individualized connectivity vs. group-based approaches for parietal TMS coil positioning (Secondary Aim A) The columns x₁, y₁, z₁ are MNI coordinates for individualized coil positions for ten representative participants (n = 32). Each participant’s rs-fMRI data was used locate an MNI coordinate in the left lateral parietal region of the default network that exhibited the highest functional connectivity with the ventral PCC. The MNI coordinate of the vPCC seed was reported in a meta-analysis of resting-state abnormalities in amnestic MCI patients.¹ The remaining columns indicate how the MNI coordinates x₁, y₁, z₁ compare with results of four other positioning methods: (A) shows distances (mm) between coil locations depending on whether the ventral PCC or the left hippocampus was used as the seed for an individualized connectivity approach. (B and C) provide distance comparisons between the individualized ventral PCC method and two group-average connectivity methods. (D) shows distances between the individualized ventral PCC method and an EEG-based method for targeting the left temporoparietal area. The descriptive statistics in the lower half of the table represent data for all 32 participants.

Individual fMRI data connectivity approaches											Group-average connectivity approaches		EEG 10-20 system
	Ventral PCC seed to target:a							Hippocampal seed to target:a			rs-fMRI literature	CONN ICA network atlas	EEG electrode placement
Ppt #	x1	y1	z1	Brain regionb	BAb	(A) Distance between x1y1z1 and x2y2z2		x2	y2	z2	(B) Distance from group-based MNI -47, -68, 36*	(C) Distance from DN parietal peak MNI-39, -77, 33	(D) Distance from temporoparietal (TP3) position
1	-42	-78	32	AG	39	7.48		-44	-82	26	11.87	3.32	31.64
2	-36	-84	32	VAC	19	10.00		-36	-76	38	19.82	7.68	39.86
3	-42	-64	24	AG	39	2.00		-44	-64	24	13.60	16.09	21.00
4	-42	-76	38	AG	39	11.66		-50	-70	32	9.64	5.92	31.32
5	-38	-82	36	VAC/AG	19	7.21		-32	-78	36	16.64	5.92	37.75
6	-38	-72	36	AG	39	7.21		-44	-72	32	9.85	5.92	30.08
7	-48	-74	20	VAC/AG	19	6.32		-42	-72	20	17.12	16.09	26.32
8	-34	-88	26	VAC	19	14.70		-44	-78	30	25.87	13.96	44.15
9	-38	-80	38	AG	39	12.17		-38	-78	26	15.13	5.92	36.62
10	-36	-80	42	AG	39	10.39		-30	-86	36	17.35	9.95	39.00
Mean Dist (mm); N=32						10.74					14.25	9.31	32.91
Minimum						2.00					5.75	3.32	19.52
25th%ile						7.21					9.54	5.92	26.55
75th%ile						13.71					17.69	11.79	38.90
Maximum						26.31					25.87	16.58	44.15
p (12 mm focality)c						n.s. >0.90					.014	n.s. >0.99	<.0001

Note:

Abbreviations: MNI Montreal Neurological Institute; PCC posterior cingulate cortex, EEG electroencephalography; AG angular gyrus, VAC visual association cortex, BA Brodmann area.

^aOur individualized connectivity method used a seed in the ventral PCC (MNI 6, −52, 16) (Lau et al 2016) and a search area that was constrained to the left lateral parietal region within an atlas of the default network (Whitfield-Gabrieli & Nieto-Castanon 2012). An earlier TMS study of younger adults used the middle of the hippocampal body as the seed and a search area centered on MNI −47, −68, 36 (Wang et al 2014). We compared the two methods. For the MCI dataset (N = 32), the magnitude of the distances from MNI −47, −68, 36 were similar on average regardless of whether the seed was in the ventral PCC or the left hippocampus (vPCC: M = 14.25 mm, SD = 5.48; left hippocampus: M = 14.59 mm; SD = 5.33), one-sample t-test (p = 0.774).

^bThe Brodmann area and cortical label are reported for descriptive purposes. They were obtained using Yale BioImage Suite.

^cThe p values are results of one-sample t-tests tests of the null hypothesis that the distance “D” is within 12 mm, cf. Fox et al (2013).

0A.1 Comparison of the ventral PCC seed approach with three other methods of stimulation site localization. First, we varied the seed ROI one might use to identify an individualized left LPC stimulation site. We used the left hippocampus vs. the ventral PCC as seeds and computed the Euclidean distances between stimulation sites for each participant. The average distance between the vPCC-seeded and the hippocampal-seeded location was 10.74 mm (SD = 5.18), which is not significantly more than 12 mm (p = 0.91). Sixty-two percent of the sites (20 of 32 participants) were 12 mm or less apart. Thirty-one percent (10 of 32) were 12.2 to 16.5 mm apart; the remainder (2 of 32) were separated by more than 20 mm. Table 2, part A shows individual coordinate pairs and the accompanying distances for the same participants displayed in Fig. 4. Table S1 in the Supplementary material shows data for all 32 participants. If confirmed, these results suggest that one could opt to use either a vPCC or a hippocampal seed to identify a parietal stimulation site. Yet, the two seeds did not point to identical stimulation locations. This may be due to noise in single-subject single-occasion rs-fMRI data as well as the influences that distinct subsystems of the default network can have on ROI-to-ROI functional connectivity (Vincent et al., 2008, Zheng et al., 2021).

Second, we compared the individualized ventral PCC seeded locations with a group-based location (MNI −39, −77, 33). This MNI location is the peak coordinate of the Conn Default Network atlas’s Lateral Parietal region (atlas based on a Human Connectome Project dataset of 497 subjects). Group-averaged functional connectivity estimates are attractive because the signal to noise ratio is typically higher than that of single-subject data (Murphy et al., 2013, Ning et al., 2019, Power et al., 2017). The mean distance of the individualized sites from this peak coordinate was 9.31 mm (SD = 4.23, n = 32). This distance was not significantly greater than 12 mm (p = 0.994). Within the full sample, 78 % (25 of 32) of participants’ individualized sites were less than 12 mm from the Lateral Parietal atlas coordinate. Table 2, part C shows individual coordinate pairs and the accompanying distances for the ten participants displayed in Fig. 4. Table S1 in the Supplementary material shows data for all participants.

A possible stimulation site that would require no neuroimaging is the International 10–20 EEG System TP3 position. This EEG electrode position might allow TMS operators to reliably stimulate the angular-supramarginal gyral region of the parietal lobe (Herwig et al., 2003). Herwig and colleagues reported an average Talairach coordinate of y = -57, x = -49, z = 28 for the EEG TP3 position, based on an MRI-EEG mapping study involving 21 cognitively normal adults. The average distance of our MCI participants’ individualized sites from TP3 was 32.91 mm (SD = 7.20), based on Talairach to MNI conversion. The distances ranged from a minimum of 19.52 mm to a maximum of 44.15 mm. Given that the minimum distance was > 12 mm and most participants’ distances were > 20 mm, we do not recommend the EEG-based TP3 method to target the default network (cf. Bagattini et al., 2021).

0A.2 Cross-disciplinary approach: Use of group connectivity data to identify a prefrontal stimulation site that is strongly anticorrelated with the AD-vulnerable ventral PCC. Following the logic outlined in Section 2.4.2.A.2 of the Methods, we used Neurosynth (Yarkoni et al., 2011) and the 1000-subject fMRI data set (Yeo et al., 2011) to identify potential stimulation site(s) within the left prefrontal cortex that exhibit strong anticorrelations with the ventral PCC (Lau et al., 2016)). The Neurosynth tool revealed three coordinates near the surface of the left aPFC and bordering the left DLPFC: MNI coordinates [-38, 44, 28], [-40, 44, 28] and [-40, 44, 30], each having an anticorrelation of r = -0.30. The “depression efficacy” site that we mapped for Aim 1 (MNI −38, 44, 26; Fig. 2A), is in close proximity and has a Neurosynth anticorrelation coefficient of −0.28. We also note the presence of strongly anticorrelated region in the left supramarginal gyrus (r = -0.37; MNI −62, −38, 34) in the 1000-subject data set. All of these anticorrelated areas identified using Neurosynth participate in the salience network.

3.3. Secondary aim b. Behavioral measures in relation to network functional connectivity

As described in Section 2.4.3 of the Methods, we computed summary measures of functional connectivity for the dorsal salience and posterior default networks, and examined their relationships with behavioral measures of memory, executive function, and depressive symptoms. The default network scores were distributed as: M = 0.39, SD = 0.21; range = 0.03–1.11 on a z-scale. The salience network scores were distributed as: M = 0.36, SD = 0.19; range = 0.00–0.81 on a z-scale. Table 1 provides descriptive statistics for the behavioral measures. The full matrix of Spearman correlation coefficients is provided in Table S2 of the supplementary material. As expected, CVLT-II learning-and-memory composite scores correlated with stronger default network connectivity (Spearman rho = 0.33, p = 0.03, planned 1-tailed test for a positive relationship). On the other hand, the predicted correlation between Logical memory scores and default network connectivity did not reach statistical significance (rho = 0.23, n.s.). Also, the expected relationship between lower Trail Making scores and lower connectivity within the salience network was not evident (Spearman rho = 0.05, n.s.). Stronger connectivity within the salience network was not significantly associated with better memory scores (|rho| < 0.20, n.s). Interestingly, higher scores on the Geriatric Depression Scale were moderately associated with lower functional connectivity in both networks (salience network: rho = -0.48; p = 0.01; default network: rho = -0.41; p = 0.02; 2-tailed). These mood-connectivity relationships are remarkable considering that we designed the sampling criteria to exclude participants with untreated depression.

4. Discussion

In this paper we employed a connectivity approach (Fox et al., 2012) to map brain networks that can be targeted by prefrontal and parietal TMS for the aim of treating cognitive impairment. For the prefrontal TMS map, the seed we chose is a TMS site that was previously identified for efficacious treatment of depression (Fox et al., 2012). We selected a depression-centric approach because most TMS trials for MCI and mild AD employed coil-placement methods that were originally developed for TMS treatment of depression (Jiang et al., 2021, Wei et al., 2022). When we overlaid an atlas of canonical brain networks on the prefrontal TMS map that we computed using rs-fcMRI data of MCI participants, the prefrontal TMS map displayed substantial overlap with the salience network.

We also used the MCI rs-fcMRI data to map connectivity of a lateral parietal stimulation site that was shown to improve memory in healthy young adults and strengthen parietal-hippocampal connectivity (Freedberg et al., 2022, Wang et al., 2014). The TMS map of MCI data recapitulated most of the regions that were reported to be “stimulation-responsive” in the young-adult sample. The map also revealed that parietal TMS at this site may target the frontoparietal network secondarily (Fig. S1), in addition to the hippocampal and posterior regions of the default network.

These results are important because knowing what brain networks are likely to be targeted by TMS (or by other noninvasive modalities) is helpful toward achieving the kinds of behavioral outcomes most desired for patients with MCI and AD (Koch et al., 2024). Secondly, TMS research that is network-targeted could accelerate accumulation of knowledge by accomplishing more rigorous, reproducible trials in which specific hypotheses are tested about TMS effects on behavior and biomarkers. A framework of network-targeted TMS and harmonized methods, including methods of coil placement, would facilitate direct comparisons between parietal and the prefrontal TMS in terms of behavioral efficacy and biomarkers of response. These emphases may help optimize TMS for treatment of late-life cognitive impairment and foster more rigorous evidence for the use of TMS.

4.1. Network-targeted TMS and behavioral outcomes

The default network (DN) plays a vital role in episodic memory encoding and retrieval (Sestieri et al., 2011, Sperling et al., 2010), autobiographical memory (Viard et al., 2010), and imagining one’s future (Andrews-Hanna et al., 2014). Core regions of the DN are the medial prefrontal cortex, posterior cingulate, and the posterior extent of the inferior parietal lobule including the angular gyrus (Uddin et al., 2019, p. 936). The DN’s medial temporal subsystem includes the hippocampi (Andrews-Hanna et al., 2014). Evidence that parietal TMS can improve episodic memory in younger adults has been extended to at least three controlled trials involving cognitively impaired older adults (Jia et al., 2021, Jung et al., 2024, Koch et al., 2018). For example, in a sample of 69 AD dementia patients, active rTMS targeted to the L-LPC led to improved recall of a list of words in comparison to sham (Jia et al., 2021). TMS targeted to the precuneus also led to improved word-list recall in a smaller controlled study of patients with prodromal AD (Koch et al., 2018).

The salience network (SN)--also known as the midcingulo‑insular network--is functionally and anatomically distinct from the DN (Uddin et al., 2019). In addition to detection of salient stimuli, key putative functions of the SN include: maintenance of a self-initiated task-set, staying alert, and enabling flexible switching between networks (Coste and Kleinschmidt, 2016, Dosenbach et al., 2008, Menon, 2015, Sadaghiani and D'Esposito, 2015, Sestieri et al., 2017, Sridharan et al., 2008). These cognitive processes support executive function and performance of mentally effortful tasks including memory encoding and information retrieval (Sestieri et al., 2017, Touroutoglou et al., 2012, Touroutoglou et al., 2018, Zhang et al., 2020). Core regions of the salience network are the anterior midcingulate cortex (aMCC, aka dorsal ACC) and bilateral anterior insula (Touroutoglou et al., 2012, Uddin et al., 2019). Left DLPFC TMS has been shown to modulate the dACC (Tik et al., 2017) and salience network (Hawco et al., 2018, Mitra et al., 2023) using a variety of coil placement methods: 1) the 10–20 EEG system; 2) individualized left DLPFC targeting (defined as the site most functionally anticorrelated with the subgenual ACC); and 3) T1-MRI neuronavigation to MNI − 42, 28, 21.

Five studies of DLPFC TMS for treatment of AD dementia have also used T1-MRI neuronavigation (Li et al., 2021, Lu et al., 2022, Moussavi et al., 2024, Wu et al., 2022, Wu et al., 2024). Three of these TMS studies targeted the same MNI coordinate studied in the current paper, or Fox’s nearby group-based subgenual ACC anticorrelated coordinate (MNI −44, 38, 34) (Li et al., 2021, Wu et al., 2022, Wu et al., 2024). Compared to the control treatment, DLPFC TMS led to higher global cognitive function in all three studies, and better performance on tests of word-list recall in the two studies that reported memory effects (Wu et al., 2022, Wu et al., 2024). One of the five DLPFC TMS trials selected participants with comorbid dementia and depression (Lu et al., 2022). Both the active and the sham groups showed improvements in global cognitive function and clinician-rated symptoms of depression, but the active TMS group had a higher rate of remitted depression.

The Lu et al. study (2022) on DLPFC TMS for comorbid depression and dementia is novel because it is one of the few published TMS trials that focused on this population. Depressive symptoms have been linked to lower functional connectivity within the salience network, as well as in the default network (Goldstein-Piekarski et al., 2022). In our MCI sample, higher scores on the Geriatric Depression Scale (GDS) were moderately associated with lower connectivity in these two networks, even though sampling criteria excluded participants with GDS scores > 5. Pending more evidence of efficacy and safety, DLPFC TMS could be the modality of choice for MCI and AD patients who have comorbid depressive symptoms because DLPFC TMS is already FDA-cleared for treatment of major depressive disorder. For MCI and AD patients in general, it would seem that parietal TMS to target the default network has the more compelling disease-based rationale because this network is the first large-scale cortical network to show early amyloid deposition and abnormal functional connectivity in AD (Jones et al., 2016, Palmqvist et al., 2017).

4.2. Suggestions for research rigor and optimization of TMS for treatment of MCI-AD

Connectivity-based coil placement has proof of concept for achieving network-specific modulation (Dave et al., 2022). There are two approaches for implementing connectivity-based coil placement: 1) use of group-average connectome-quality data, e.g. (Fox et al., 2012, Weigand et al., 2018); and 2) individual single-subject connectivity data, e.g. (Cash and Zalesky, 2024, Dave et al., 2022, Fox et al., 2013). Both approaches have the potential to increase the effectiveness of TMS (Cash et al., 2021a, Cash et al., 2021b, Fox et al., 2012, Weigand et al., 2018). In efforts to advance TMS as an evidence-based therapy for MCI-AD, potential advantages of incorporating connectivity-based strategies include: 1) more statistical power if there is greater average improvement following TMS treatment or smaller interindividual variability in the magnitude of change; 2) more reproducible results across TMS studies.

Evolution of connectivity-based coil placement methods. The 5-cm method for coil placement was designed to stimulate the left DLPFC for treatment of MDD (George et al., 1995). This genre of coil-positioning methods has been linked to variability in the magnitude of global improvement (Siddiqi et al., 2021). The F3 EEG method is intended for more accurate localization of the DLPFC (Fitzgerald et al., 2009). The connectivity-based DLPFC MNI coordinates are based on depression research (Fox et al., 2012). All of these methods have been used in prefrontal TMS trials for treatment of MCI or AD dementia. We do not know of any meta-analyses that have contrasted coil placement methods for differences in effect size.

A recent parietal TMS trial for treatment of MCI and AD targeted the left angular gyrus at (MNI −47, −67, 38) based on the authors’ previous cross-sectional functional connectivity results (Chen et al., 2023). In our examination of the nearby site (MNI −47, −68, 36), the results revealed two concerns about using these sites to target the posterior default network. First, the site examined in our study may target the frontoparietal network secondarily in that our L- LPC TMS seed map revealed connectivity with bilateral frontal, parietal, and occipital regions of the frontoparietal network. Secondly, the site was 14 mm away on average from the individualized connectivity-based sites we identified (Table 2). Many sites were located posterior to MNI −47, −68, 36. Similar locations were identified by Bagattini et al. (2021) in a methodological study on individualized network-based fMRI-guided TMS for AD. An alternative ICA-HCP coordinate (MNI −39, −77, 33) (Whitfield-Gabrieli and Nieto-Castanon, 2012) was 5 cm closer on average to the individualized connectivity-based sites. If researchers opt to use group-average connectivity-based neuronavigation, use of this coordinate appears to be justifiable.

Demonstrations of superior coil-placement methods and stimulation sites need to be confirmed in randomized trials of TMS. Much of the current evidence in support of individualized connectivity-based coil placement is based on retrospective data (Cash et al., 2021a, Cash et al., 2021b, Siddiqi et al., 2021). A recent multi-site clinical trial on TMS treatment of major depression did not find personalized connectivity-based coil placement to be superior to the F3 EEG method (Morriss et al., 2024).

Network-targeted TMS and therapeutic mechanisms. Another major motivation for connectivity-based coil placement is to elucidate the therapeutic mechanisms of TMS from a disorder-related perspective using reproducible methodology. Explicit identification of the targeted network of interest would also facilitate comparisons between DLPFC TMS studies for MCI-AD. Our predictive TMS map, and the predictive modeling of Opitz et al. (2016) suggests that the depression efficacy site (MNI −38, 44, 26) could be used as a connectivity-based coil placement method to target the salience network. A caveat is that, because of individual differences in brain structure and the lack of 1:1 structure–function relationships, other networks may be targeted in some individuals.

DLPFC TMS could potentially be a means to normalize the activity and functional connectivity of posterior regions such as the PCC. Here, we used Neurosynth to explore potential prefrontal TMS sites near the cortical surface that are reportedly the most anticorrelated with the vPCC (Yarkoni et al., 2011). We identified three MNI coordinates. They bordered the DLPFC and were located 2 to 4 mm superiorly relative to the “depression efficacy” site. In regard to therapeutic mechanisms of TMS, future research could test whether TMS delivered to the prefrontal “depression efficacy” site leads to stronger anticorrelation between prefrontal and PCC regions (cf. Fox et al., 2012). This cross-disciplinary approach could provide clues to understanding why DLPFC TMS is beneficial for both MDD and for MCI-AD. In summary, harmonized methods and an integrated framework of network-targeted TMS would facilitate comparisons between prefrontal and parietal TMS in terms of efficacy and therapeutic mechanisms.

4.3. Limitations

The primary aim was to identify the large-scale brain networks that are likely to be targeted by prefrontal and parietal TMS, based on the MNI coordinates of two influential TMS research papers (Fox et al., 2012, Wang et al., 2014). A limitation of the two TMS connectivity maps we computed is that the MCI data are in the absence of neuromodulation. Also, the TMS maps were averages of single-subject maps based on 9 min of fMRI data per individual. We lacked the recommended 15-min minimum to adequately represent individual differences in brain network architecture (Cash and Zalesky, 2024). Assumptions were made about the focality of TMS. The focality and volume of TMS-induced changes in neuronal activity is complex to model (Fox et al., 2013, Opitz et al., 2016). To refine the spatial extent of the maps, more robust pre- and post-TMS connectivity data will be needed. The TMS group maps presented here are meant to highlight the feasibility of using TMS to modulate separate brain networks for the objective of improving neurocognitive outcomes in MCI-AD. Another limitation is the lack of AD biomarkers in this sample.

4.4. Conclusions

Advances in TMS can pave the way for more effective and personalized treatment. Clinical studies in the field of depression show that group-averaged and single-subject connectivity mapping can be used to inform coil placement for the goal of improving effectiveness (Cash and Zalesky, 2024, Cash et al., 2019, Fox et al., 2012, Siddiqi et al., 2020, Weigand et al., 2018). We believe this approach is applicable to improving the effectiveness of TMS treatment for patients with MCI and AD dementia. Network-targeted TMS can also be applied in the service of more personalized treatment for this heterogenous population--such as salience network modulation for individuals with pronounced depression or apathy, or default network modulation for improvement of memory function. Our vision is a network-targeted TMS strategy that is relevant to the spectrum of Alzheimer’s.

CRediT authorship contribution statement

Joy L. Taylor: Writing – original draft, Supervision, Project administration, Methodology, Investigation, Funding acquisition, Conceptualization. Priyanka Bhatt: Writing – review & editing, Visualization, Validation, Software, Methodology, Formal analysis, Data curation. Beatriz Hernandez: Writing – review & editing, Validation, Formal analysis, Data curation. Michael Iv: Validation, Investigation. Maheen M. Adamson: Writing – review & editing, Conceptualization. Alesha Heath: Investigation, Conceptualization. Jerome A. Yesavage: Funding acquisition. Margaret Windy McNerney: Methodology, Investigation.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article:

Data availability

Data will be made available on request.