Linking Personalized Brain Atrophy to Schizophrenia Network and Treatment Response

Abstract

Background and Hypothesis

Schizophrenia manifests with marked heterogeneity in both clinical presentation and underlying biology. Modeling individual differences within clinical cohorts is critical to translate knowledge reliably into clinical practice. We hypothesized that individualized brain atrophy in patients with schizophrenia may explain the heterogeneous outcomes of repetitive transcranial magnetic stimulation (rTMS).

Study Design

The magnetic resonance imaging (MRI) data of 797 healthy subjects and 91 schizophrenia patients (between January 1, 2015, and December 31, 2020) were retrospectively selected from our hospital database. The healthy subjects were used to establish normative reference ranges for cortical thickness as a function of age and sex. Then, a schizophrenia patient’s personalized atrophy map was computed as vertex-wise deviations from the normative model. Each patient’s atrophy network was mapped using resting-state functional connectivity MRI from a subgroup of healthy subjects (n = 652). In total 52 of the 91 schizophrenia patients received rTMS in a randomized clinical trial (RCT). Their longitudinal symptom changes were adopted to test the clinical utility of the personalized atrophy map.

Results

The personalized atrophy maps were highly heterogeneous across patients, but functionally converged to a putative schizophrenia network that comprised regions implicated by previous group-level findings. More importantly, retrospective analysis of rTMS-RCT data indicated that functional connectivity of the personalized atrophy maps with rTMS targets was significantly associated with the symptom outcomes of schizophrenia patients.

Conclusions

Normative modeling can aid in mapping the personalized atrophy network associated with treatment outcomes of patients with schizophrenia.

Introduction

Numerous neuroimaging studies have investigated brain structure and function in schizophrenia,^1–3 but few have yielded clinical biomarkers to assist with diagnosis or treatment.⁴ One possible reason explaining the absence of biomarkers is that most of the research findings are between-group analyses with an assumption that patients form a homogenous population. However, mounting evidence calls this assumption into question, with recognition of marked heterogeneity among individuals comprising a given diagnostic category, both in terms of clinical presentation and underlying biology.^5,6 This heterogeneity suggests that findings at the group level are not necessarily representative of individuals. Thus, modeling individual differences within clinical cohorts is critical to translate knowledge reliably into clinical practice.^7–10

Recently, a growing number of studies have decoded individual differences in neuroimaging phenotypes.^4,11–14 Normative modeling is 1 such emerging approach enabling individual-level inference.^15,16 Using normative modeling of brain structure, magnetic resonance imaging (MRI) studies have identified locations of regional atrophy for individual patients with neurodegenerative diseases,^17,18 attention-deficit/hyperactivity disorder symptoms,¹⁹ autism spectrum disorder,²⁰ and corticobasal syndrome.²¹ In schizophrenia, brain regions outside established normative ranges are regionally heterogeneous and less than 20% of patients show common areas of atrophy.^22,23 Deviations in cortical thickness (CT) from normative ranges are associated with a higher polygenic risk for schizophrenia.²² However, the credibility and clinical utility of individual-specific atrophy patterns remain unclear. To address this question, this study investigated whether the personalized brain atrophy maps for individuals with schizophrenia (1) comprised a disorder-specific functional brain network, and (2) explained individual variation in clinical response to repetitive transcranial magnetic stimulation (rTMS). Each is now considered in turn.

The first aim was motivated by recent evidence suggesting that brain lesions located at any location within the same functional network could result in a common set of symptoms.^24,25 Indeed, several neuropsychiatric disorders are characterized by a disease-specific network, and while the anatomical location of lesions is diverse between individuals, these locations typically reside within a common network.^26–34 Thus, we predicted that regional atrophy in patients with schizophrenia would converge in a common brain network different from that in patients with nonpsychiatric disorders.

Second, rTMS has been investigated as a potential treatment for schizophrenia for decades,³⁵ but marked heterogeneity in clinical outcomes is evident among patients.^36,37 The selection of stimulation sites critically impacts rTMS efficacy. For instance, the functional connectivity of the stimulation site to pathological hub regions of major depression disorder significantly predicted individual responses to rTMS treatment.^38,39 This suggests that a more precise modulation of the pathological network may be associated with better clinical outcomes. Thus, we hypothesized that functional connectivity between personalized atrophy regions and the stimulation site would positively predict individual response to rTMS treatment. To this end, we adopted a randomized clinical trial dataset of rTMS in schizophrenia.

Materials and methods

Design

This study was approved by the Research Ethics Board for the Anhui Medical University, Hefei, China. We first identified personalized atrophy maps for each individual with schizophrenia using normative reference ranges for regional CT measurements. We then tested whether individual variation in deviations was associated with previous case–control studies and individual responses to rTMS treatment.

Participants

This study included 797 healthy adults (18 to 45 years old, 411 females) and 91 schizophrenia patients. All participants provided written informed consent before the experiments. High-resolution structural images of brain anatomy were acquired for all individuals. Additionally, resting-state functional MRI (rs-fMRI) data were acquired for 652 of the 797 healthy subjects. Schizophrenia patients were enrolled in the Anhui Mental Health Center (Hefei, China). The inclusion criteria for patients included: (1) diagnosis of schizophrenia using the Structured Clinical Interview for the Diagnostic and Statistical Manual of Mental Disorders, the 5th Edition, (2) no history of other neuropsychological diseases, and (3) 18 to 45-years-old. The exclusion criteria for both groups were: (1) history of significant head trauma, (2) alcohol or drug abuse, (3) focal brain lesions on T1- or T2-weighted fluid-attenuated inversion-recovery magnetic resonance images, (4) head motion exceeding 3 mm in translation or 3° in rotation during rs-fMRI scanning, or (5) Hamilton Anxiety Rating Scale (HAMA) or Hamilton Depression Rating Scale (HAMD) scores > 7.

rTMS treatment

In total, 52 of the 91 patients were from a randomized clinical trial (NCT02863094), which was not prospectively designed for the current study. Patients with schizophrenia at the Anhui Mental Health Center were enrolled to test the clinical efficacy of rTMS. Additional to the above criteria, these participants satisfied the following criteria: (1) stable doses of psychotropic medication for at least 8 weeks before the rTMS treatment. (2) no nonremovable metal objects in or around the head, (3) no prior history of seizure or history in first-degree relatives, and (4) did not receive previous rTMS or transcranial electronic stimulation.

Patients were randomly assigned to real or sham rTMS groups, and received continuous theta-burst stimulation (cTBS) for 15 days (1800 pulses per day). cTBS were delivered using a MagStim Rapid² stimulator (Magstim Company Ltd. Carmarthenshire, Wales, UK) with a 70-mm air-cooled figure-of-8 coil. One session of cTBS was 40 s in duration and consisted of 3-pulse bursts at 50 Hz repeated every 200 ms (5 Hz) until a total of 600 pulses was reached.⁴⁰ To achieve cumulative aftereffects, this protocol was repeated 3 times and separated by 2 15-min breaks (1800 pulses in total), in line with previous methodological studies.^41,42 cTBS was delivered at 80% of the resting motor threshold⁴³ or the highest intensity the stimulator could deliver for this protocol (50% of maximum output). The resting motor threshold was determined at each visit according to a 5-step procedure.⁴⁴

A stimulation target in the left temporoparietal junction (TPJ) area (Montreal Neurological Institute [MNI] coordinates, [−51, −31, 23]) ³⁷ was selected and transformed into each participant’s T1 space by applying an inverse matrix produced during T1 segmentation in SPM12 (www.fil.ion.ucl.ac.uk/spm). Each individual’s target was imported into a frameless neuronavigation system (Brainsight; Rogue Research, Montreal, Canada). The coil was held tangentially to the skull pointing forward, with the center over the sphere target. Patients in the sham control group received the same rTMS protocol and treatment duration as the real rTMS group. The only difference was the use of a sham coil (Magstim Company Ltd.) that produced a similar feeling on the participant’s scalp as the real coil but did not induce a current in the cortex. The positive and negative syndrome scale (PANSS) score was estimated before and (>12 h) after the treatment. Study participants, clinical raters, and all personnel responsible for the clinical care of the patient remained masked to the allocated conditions and allocation parameters. Only rTMS administrators had access to the randomization list; they had minimal contact with the patients, and no role in assessing clinical symptoms.

MRI Data and Preprocessing

Structural and fMRI data were obtained at the University of Science and Technology of China (Hefei, China) with a 3-T scanner (Discovery 750; GE Healthcare, Milwaukee, WI, USA). High-resolution T1-weighted images were acquired in the sagittal orientation using a magnetization-prepared rapid gradient-echo sequence (repetition/echo time, 8.16/3.18 ms; flip angle, 12; field of view, 256 × 256 mm²; 256 × 256 matrices; section thickness, 1 mm; voxel size, 1 × 1 × 1 mm³). Resting-state functional images were acquired using a single shot gradient-recalled echo planar imaging sequence (repetition/echo time, 2400/30 ms; flip angle, 90; field of view, 192 × 192 mm²; 64 × 64 in-plane matrix; section thickness, 3 mm; voxel size, 3 × 3 × 3 mm³; 46 transverse sections). A total of 217 volumes were acquired (~8.7 min). During scanning, participants were instructed to rest with their eyes closed without falling asleep.

Quantitative morphometric analysis was performed using FreeSurfer (version 6.0). In brief, images underwent preprocessing steps including intensity normalization, skull stripping, registration into Talairach space, and determination of the gray–white matter boundary and pial surface. CT was computed by measuring the distance between the white matter and pial surfaces at approximately 160 000 vertices. Each subject’s reconstructed brain was registered to an average spherical template space and smoothed with an 8 mm full width at half maximum kernel. The 8 mm kernel size was selected for consistency with most structural studies on patients with schizophrenia (supplementary table S1). The entire cortex of each subject was visually inspected, and inaccuracies in segmentation were manually edited.

The rs-fMRI data were preprocessed using SPM12 software (www.fil.ion.ucl.ac.uk/spm) and ANFI (https://afni.nimh.nih.gov/afni/). The processing steps were as follows: (1) delete the first 5 time points, (2) remove temporal spikes more than 2.5 SDs from the mean (AFNI’s 3dDespike), (3) slice timing correction, (4) head motion correction, (5) co-registration to the structural image, (6) regress out nuisance regressors (24 head motion parameters, and average signals in the cerebrospinal fluid, white matter, and whole brain), (7) spatial normalization to the MNI space using the matrix produced by structural image segmentation using DARTEL algorithm in SPM12,⁴⁵ and (8) spatial smoothing with a 4 mm full width at half maximum Gaussian kernel. This relatively small kernel (4 mm) was selected to maintain the sensitivity to detect small brain functional features as in our previous studies.^46–48

Atrophy Map and Network

Individual atrophy maps for each patient were defined as CT deviations from normative modeling.^{15,17,18,49,50} The CT rather than other structural measures were used here since it has been shown as a sensitive index for identifying individualized abnormality in schizophrenia.¹⁶ Briefly, we first established a vertex-wise general linear model (GLM) of CT for the healthy subjects using age and gender as covariates (figure 1A). Next, parameters of the normative model (ie, beta term maps for age and gender, and residuals) were used to calculate a vertex-wise W-score for the CT of each patient (a W-score is a z-score adjusted for age and gender) using the formula: W-score = (expected—actual)/RSD, where actual is the patient’s observed CT, expected is the predicted CT with the application of the nuisance factors and beta weights from the regression in healthy controls (ie, age×β_age+ gender×β_gender), and RSD is the residual standard deviation (SD) from the GLM. For each patient, locations with a positive W-score provided evidence of atrophy. To show the spatial consistency of atrophy loci maps across patients, we have to apply a cutoff value on the W maps. For consistency with previous studies,^18,21 we reported the atrophy loci with 2 W thresholds, 2 and 2.5. In contrast, the atrophy network did not need to be established on cutoff W maps. The whole brain W map was used as a weighted seed image to compute the atrophy network on functional data.

Fig. 1.

Personalized atrophy mapping method. (A) Normative modeling of cortical thickness (CT) is generated based on 797 healthy subjects using age and gender as covariates. The functional connectivity map of a given seed is computed using resting-state functional magnetic resonance imaging (MRI) data of 652 healthy subjects (B). The vertex-wise atrophy map could be generated for individual patients using the normative modeling controlling for age and sex (C). Then, the connectivity pattern (or atrophy network) of the personalized atrophy map is represented by the t-map of 652 healthy connectomes.

Each patient’s atrophy network was defined by an “atrophy network mapping” approach (figure 1B).^18,21 This method was largely based on the idea of “lesion network mapping”.²⁶ Specifically, we first converted the W map from surface to volume space (mri_surf2vol function in Freesurfer). Then, the whole brain fMRI signals were weighted (by W-score) and averaged (to avoid the arbitrary selection of W threshold, we did not use regions with high W scores as seed regions). Next, this means signal was correlated with the whole brain voxels using Pearson’s correlation on the rs-fMRI data of 652 healthy subjects. Finally, one-sample t-tests were performed on these functional connectivity maps (Fisher’s r-to-z transformed). The t-map indicated the connectivity pattern of an individualized atrophy map, and thus was defined as the personalized atrophy network of a given patient.

Schizophrenia Network From Case-Control Studies

A “coordinate-based network mapping” method⁵¹ was adopted to obtain the schizophrenia network from case–control studies. We collected 27 contrasts (patients vs healthy controls) from 21 case–control studies according to 2 meta-analysis studies of schizophrenia^52,53 (supplementary table S1). Spherical seeds (3 mm diameter; recomputed using 6 and 9 mm) at each reported coordinate were combined to generate a contrast-specific seed. Then, we identified voxels functionally connected to each contrast’s combined seed using rs-fMRI data from 652 healthy subjects. Connectivity maps were thresholded at t > 4.7, which corrected for multiple comparisons (voxel-wise family-wise error [FWE] rate of P < .05). Binarized maps were then added together to identify regions significantly connected to coordinates of all, or most, of the contrasts. To ensure results were not dependent on the threshold, analyses were repeated for thresholds of t > 5.1 (P_FWE < .01) and t > 5.6 (P_FWE < .005).

Specificity Analysis of the Atrophy Network

To test whether the extent of regional overlap evident between the personalized atrophy networks was specific to schizophrenia, we compared the 91 atrophy networks to 49 coordinated-based networks from 4 nonpsychiatric disorders (Alzheimer’s disease, n = 8; behavioral variant frontotemporal dementia, n = 21; corticobasal syndrome, n = 12; and progressive non-fluent aphasia, n = 8) from the literature (supplementary table S2).⁵¹ This comparison was performed using a permutation-based 2-sample t-test in FSL v6.0 (PALM, 10 000 permutations, corrected by threshold-free cluster enhancement, P_FWE < .001).

Atrophy Network and rTMS Outcome

The same brain target was modulated in the 52 patients that received rTMS treatment; However, the clinical outcome was highly variable among the patients. To explain this variability, we related the personalized atrophy network and rTMS outcome. First, we extracted the t-value of the rTMS target (a 3 mm sphere in the TPJ) from the personalized atrophy network. Then, these individualized t values were correlated to the PANSS improvement after treatment (baseline minus post-rTMS scores, and divided by the baseline score).

Statistical Analysis

Associations between imaging and symptoms were performed using Pearson’s correlations. The treatment effect of rTMS was examined using linear mixed-effect models nested within individuals (GraphPad Prism 8.3.0). There were 2 factors (2 levels in each): Time (pre- and post-rTMS) and Group (active and sham rTMS). Post hoc analyses were performed using Sidak’s multiple comparison test. Baseline characteristics between the active and sham groups were compared using the χ² test for categorical variables and the Student’s t-test for continuous variables.

Spatial correlation between brain maps was performed across ROIs of the HCP-MMP atlas,⁵⁴ and corrected using “spin” tests (10 000 permutations). A spin-based method was used to correct for potential confounding effects of spatial autocorrelation. Specifically, we first generated 10 000 random spatial rotations (ie, spins) of the cortical ROIs to generate a null distribution (https://github.com/frantisekvasa/rotate_parcellation).^55,56 Then, the P_spin values were obtained by comparison with the null model (< 5th or > 95th percentile).

Data Availability

Data are available from the corresponding authors upon request.

Results

Participants

A total of 91 patients with schizophrenia were included in this study (mean age ± SD = 26.0 ± 7.90; mean number of years of education ± SD = 12 ± 2.68; 54 women). These patients had high PANSS scores (mean ± SD = 69.9 ± 16.90) and low HAMA (mean ± SD = 5.7 ± 3.20) and HAMD (mean ± SD = 5.1 ± 2.92) scores, on average. The average risperidone equivalent was 21.6 (SD = 16.39).

Heterogeneous Structural Atrophy Loci

We constructed a normative model by using the demographic and structural MRI data of 797 healthy individuals. Besides the patients with schizophrenia (n = 91), we additionally collected structural MRI data of an independent healthy group (n = 60, mean age ± SD = 25.3 ± 7.40, 34 women) with the same scanner and sequence as the 797 healthy participants. These 2 validation groups were used to validate the normative model. Taking age and gender as inputs, we computed the predicted CT of each participant in the validation groups by using the normative model. The predicted brain CT was correlated with the actual CT across the vertex for each participant. The spatial correlation coefficients were significant in both the healthy group (mean ± SD = 0.74 ± 0.02, t = 263, P < .0001) and the schizophrenia group (mean ± SD = 0.75 ± 0.03, t = 309, P < .0001).

We defined the significant atrophy loci of each patient as regions with a W-score > 2, corresponding to a CT of 2 SDs below the mean of healthy subjects. All 91 schizophrenia patients revealed 1 or more atrophy loci. Binarized loci satisfying this criterion were summed across patients to show the overlapping degree in population. The highest overlapping of personalized atrophy loci was 46% (in the right insular cortex, figure 2). This overlapping degree dramatically decreased to 29% as the W threshold increased to 2.5.

Fig. 2.

Personalized atrophy loci. Examples of the location of atrophy loci (W threshold, 2.0) in 5 patients (A). The regions with overlapping degrees higher than 5% across 91 schizophrenia patients (B).

Common Functional Atrophy Network

The binarized (threshold t = 4.7, P_FWE< 0.05) atrophy networks showed a 88% maximal overlapping in the bilateral medial prefrontal cortex (MPFC; supplementary table S3, figure 3A, supplementary figure S1A). This maximal overlapping did not change as t thresholds increased to 5.1 (P_FWE< .01) or 5.6 (P_FWE< .005). The degree of overlap of 9 clusters exceeded 80% (supplementary table S3). To test whether the overlap was specific to schizophrenia, we compared the schizophrenia atrophy networks with those of nonpsychiatric disorders. All 9 clusters survived this specificity analysis (P_FWE < .001, figure S1B).

Fig. 3.

A common network of schizophrenia patients. The atrophy networks are computed by seed-to-whole brain functional connectivity maps on a large sample healthy functional magnetic resonance imaging (MRI) dataset (n = 652, see figure 1), where seeds are personalized atrophy maps (A) or literature coordinates (B). The significant positive areas in the connectivity maps are binarized and averaged across individual patients (A) or case-control contrasts (B). The scatter plot shows a high spatial correlation between coordinate-based and atrophy-based networks.

Schizophrenia Network From Case–Control Studies

Using coordinate-based network mapping, we identified networks of regional atrophy from 27 case-control contrasts. Summing the binarized connectivity maps (threshold t = 4.7, P_FWE< .05), we found relatively high overlapping in the MPFC and insular cortex (maximal overlap, 85% contrasts; figure 3B). This overlapping pattern was independent of the sphere diameter (supplementary figure S2) or t-threshold (supplementary figure S3) and spatially correlated to that computed from personalized atrophy networks (r = 0.67, P_spin < .0001; figure 3B).

Atrophy Network and rTMS Response

A total of 52 of the 91 patients with schizophrenia received rTMS treatment. The overlap map indicated that loci and networks in this subgroup were similar to those in the entire group (supplementary figure S4). No significant difference was found in demographic and baseline symptoms between the active and sham rTMS groups (table 1). Two-way analysis of variance indicated a significant interaction of Time and Group with the PANSS score (F_(1,50)= 18.47, P < .0001, figure 4A). Post hoc analysis showed a significant decrease of PANSS score in the active group (mean ±  SD: From 73.4 ± 18.34 to 57.7 ± 13.39, P < .0001), but not the sham group (mean ± SD: From 71.4 ± 16.98 to 68.8 ± 21.01, P = .40).

Table 1.

Demographic and Clinical Information of Schizophrenia Patients

	Schizophrenia Patients (Cross Sectional, n = 91)	Schizophrenia Patients (rTMS Treatment)
		Active Group (n = 26)	Sham Group (n = 26)	Statistics/P-Value
Age (years)	26.0 ± 7.90	26.4 ± 8.63	26.5 ± 6.65	.05/.96
Gender (m/f)	37/54	12 ± 14	13 ± 13	.08/.78
Education (years)	12 ± 2.68	12.4 ± 2.89	12.0 ± 2.74	.49/.62
Disease Duration (years)	6.1 ± 4.95	6.7 ± 4.66	6.4 ± 5.40	.18/.86
PANSS	69.9 ± 16.90	73.4 ± 18.34	71.4 ± 16.98	.40/.69
HAMA	5.7 ± 3.20	5.0 ± 2.93	6.1 ± 3.48	1.17/.25
HAMD	5.1 ± 2.92	4.8 ± 2.36	5.7 ± 3.92	1.07/.29
Risperidone equivalent (mg)	21.6 ± 16.39	22.8 ± 15.6	23.4 ± 21.2	.12/.91

HAMA, Hamilton Anxiety Rating Scale; HAMD, Hamilton Depression Rating Scale; PANSS, the Positive and Negative Syndrome Scale; rTMS, repetitive Transcranial Magnetic Stimulation.

Fig. 4.

Relationship between symptom improvement and the personalized atrophy network. The positive and negative syndrome scale (PANSS) score was significantly improved after active rTMS, but not the sham condition (A). All patients received repetitive transcranial magnetic stimulation at the same anatomical site, but the connectivity between the target and personalized atrophy map varied among individuals. The connectivity strength was significantly associated with the improvement of the PANSS score in the active group but not the sham group (B). The brain map (C) shows 3 representative conditions from the active group (targets are marked by white circles). The t-maps are the results of one-sample t-tests on 652 seed-to-whole brain connectivity maps using patient-specific atrophy maps as the seed.

From the personalized atrophy networks, we extracted the connectivity strength of the rTMS target for each patient. The strengths were significantly correlated with the PANSS score improvement after active rTMS treatment (r = 0.54, P = .005), but not for the sham condition (r = 0.25, P = .24, figure 4B–C). This correlation remained significant when controlling for age, gender, education, HAMA, HAMD, and disease duration (r = 0.56, P = .01).

Discussion

This study adopted normative modeling to identify personalized atrophy maps for patients with schizophrenia. The credibility and clinical utility of these personalized atrophy maps were revealed in 2 findings. First, atrophy maps were highly heterogeneous across patients, but functionally converged to a putative schizophrenia network that comprised regions implicated by previous group-level findings. Second, the connectivity strength of the rTMS target in personalized atrophy networks significantly explained symptom improvement. Future studies may utilize the personalized atrophy networks to predict or improve the clinical efficiency of rTMS for schizophrenia.

Convergent Individual- and Group-Level Findings

In contrast to previous studies investigating the regional heterogeneity of CT atrophy in schizophrenia,^22,23 we used a relatively rough but concise GLM to establish a normative model. Probably because of these differences, we found a higher overlap ratio (46%, W threshold = 2.5) among the personalized atrophy regions in our patient group than the 20% reported in previous work.²² However, our overlap ratio markedly decreased, to 29%, when we increased the W threshold to 2.5. Wolfers et al., explained the heterogeneity of regions from the brain network perspective^23,57 whereby spatially discrete abnormalities may impair the same functional network via different mechanisms.^17,27 In line with this network hypothesis, we found that the atrophy maps in the majority of patients (ie, 88%) were functionally connected to the same region (ie, the MPFC).

To obtain the atrophy network at the group level, we summed the personalized atrophy connectivity maps.^18,21 The overlapping map showed a high spatial similarity to the map computed with coordinate network mapping.⁵¹ Both maps showed a high overlapping degree in the MPFC regions. These findings suggest that our individual-level findings were concordant with those of group-level studies. However, the individualized atrophy network was superior to group-level findings in providing more specific information explaining the symptoms of each patient.

Personalized Imaging Biomarkers

Numerous neuroimaging studies have sought to identify schizophrenia biomarkers,^1–3 but few of the findings have been successfully translated to clinical settings.⁴ A potential barrier to clinical translation is the use of group-level paradigms that overlook individual heterogeneity in symptoms and neuropathology. Recent work shows that CT atrophy associated with schizophrenia is regionally heterogeneous and may be nested within healthy variation.¹⁶ These features not only question the suitability of the case-control paradigm but also the existence of group-level schizophrenia biomarkers. On the contrary, normative modeling approaches can characterize individual variability among patients. Their deviations from normal ranges may be utilized as personalized imaging biomarker.¹⁸

Based on longitudinal rTMS data, we found that stimulation target connectivity strength in a personalized atrophy network was significantly associated with treatment outcome. The therapeutic utility of rTMS in schizophrenia has been investigated for decades,³⁵ but approaches to optimize target selection are lacking. The left TPJ area and dorsolateral prefrontal cortex are 2 of the most frequently targeted rTMS regions in schizophrenia studies.³⁶ The selection of these 2 regions was inferred from group-level findings. However, group information was not necessarily representative of any particular patient due to the high clinical and neuroanatomical heterogeneity among schizophrenia patients. A more personalized targeting approach may improve clinical outcomes. An interesting future direction for RCT studies may be selecting a patient-specific rTMS target from the personalized atrophy network. According to our findings, the TPJ point with the highest weight in the atrophy network may be an optimal choice.

Some limitations and future directions should be mentioned. First, the normative modeling was established using data from a single center, and this may not be generalized to other centers. Establishing multicenter modeling will be the next step to integrating normative modeling into clinical practice. Second, this study exclusively focused on schizophrenia patients. An increasing body of evidence implicates shared etiological and pathophysiological characteristics among mental disorders^2,58 and, thus, it would be interesting for future studies to generalize our findings to other psychiatric disorders. Third, we found that a negative correlation between the rTMS target and atrophy map was associated with poorer symptom improvement. Notably, this result was exclusively based on inhibitory rTMS data. If excitatory rTMS were used, it is possible that the negative correlation of the atrophy map may predict better treatment outcomes.^38,59 Fourth, we caution that the overlap ratio of the personalized atrophy maps was not comparable to that of the atrophy network maps, because the latter is much less sparse and therefore has higher overlap.

Conclusion

This study aimed to test the credibility and clinical utility of personalized atrophy maps in patients with schizophrenia. Our findings indicated that personalized atrophy maps showed high functional connectivity to schizophrenia-related hub regions revealed by previous group findings. More importantly, the personalized atrophy map, as an individualized biomarker, may explain the clinical outcomes of rTMS treatment. In summary, normative modeling can aid in mapping the personalized atrophy network associated with treatment outcomes of patients with schizophrenia.

Funding

This study was funded by the National Natural Science Foundation of China, Grant Numbers: 81971689 (G.J.), 82090034 (K.W.), 32071054 (Y.T.), 31571149 (K.W.), 31970979 (K.W.), and 82001429 (T.B.); the Science Fund for Distinguished Young Scholars of Anhui Province, Grant Number: 1808085J23; the Collaborative Innovation Center of Neuropsychiatric Disorders and Mental Health of Anhui Province; and the Youth Top-notch Talent Support Program of Anhui Medical University. A.Z. was supported by a research fellowship from the NHMRC (APP1118153).

Competing Interests

The authors declare no conflict of interest.