Functional brain activity constrained by structural connectivity reveals cohort-specific features for serum neurofilament light chain.

Background: Neuro-axonal brain damage releases neurofilament light chain (NfL) proteins, which enter the blood. Serum NfL has recently emerged as a promising biomarker for grading axonal damage, monitoring treatment responses, and prognosis in neurological diseases. Importantly, serum NfL levels also increase with aging, and the interpretation of serum NfL levels in neurological diseases is incomplete due to lack of a reliable model for age-related variation in serum NfL levels in healthy subjects.
Methods: Graph signal processing (GSP) provides analytical tools, such as graph Fourier transform (GFT), to produce measures from functional dynamics of brain activity constrained by white matter anatomy. Here, we leveraged a set of features using GFT that quantified the coupling between blood oxygen level dependent signals and structural connectome to investigate their associations with serum NfL levels collected from healthy subjects and former athletes with history of concussions.
Results: Here we show that GSP feature from isthmus cingulate in the right hemisphere (r-iCg) is strongly linked with serum NfL in healthy controls. In contrast, GSP features from temporal lobe and lingual areas in the left hemisphere and posterior cingulate in the right hemisphere are the most associated with serum NfL in former athletes. Additional analysis reveals that the GSP feature from r-iCg is associated with behavioral and structural measures that predict aggressive behavior in healthy controls and former athletes. Conclusions: Our results suggest that GSP-derived brain features may be included in models of baseline variance when evaluating NfL as a biomarker of neurological diseases and studying their impact on personality traits.

Introduction

Neurofilaments are cytoskeleton proteins of neurons and are predominantly found in myelinated axons. NfL is one of three subunits of neurofilament proteins that are released into the cerebrospinal fluid (CSF) and eventually the blood in significant quantities following axonal damage or neurodegeneration[1-4]. Recent advances in immunoassay technologies have enabled reliable detection of NfL in blood and have been utilized in multiple studies to demonstrate high correlation between NfL levels in CSF and blood[5]. An increased concentration of NfL in blood (serum NfL level) or CSF has been reported in numerous studies of neurodegenerative diseases[6] as well as concussion[7,8]. Since collecting blood-based biomarkers is more practical and desirable for extensive clinical trials as compared to CSF-based biomarkers, numerous studies have analyzed serum NfL levels in the context of different neurological disorders such as multiple sclerosis[9,10], dementia[11], progressive supranuclear palsy[12], traumatic brain injury (TBI)[13], Parkinson’s disease[14], Alzheimer’s disease[15], and Huntington’s disease[16]. Changes in NfL levels have also been linked to aging[9,17] and regional atrophy in cortical brain areas[15,18] and are therefore relatable to brain atrophy in aging among people without a recognizable neurological disease. While existing studies provide convincing evidence that serum NfL level is a promising biomarker to detect neurodegeneration in a broad range of clinical applications, the interpretation of these results has been limited by focusing analysis only on detecting the abnormal increase in serum NfL levels and by a lack of studies aimed at identifying the relevant underlying features associated with serum NfL levels[6,19]. For instance, the fundamental mechanism that links aging with serum NfL levels, even among healthy subjects, is unknown[6]. Moreover, while serum NfL levels have been studied in controlled groups in the context of various neurological disorders, insight into the relationship between serum NfL levels and the onset of common neurological symptoms in a healthy group is still lacking[19]. Therefore, it is relevant to search for underlying features that can help ongoing neurological studies model the variance in serum NfL levels among healthy controls and thereby provide a better understanding of both baseline serum NfL levels and how abnormal variation in these levels is mechanistically linked to neurodegenerative disease through these features.

The analysis of brain imaging data has been extensively utilized in neuroscience for independent or joint studies of aging, cognition, and neurological disorders[20-22]. The application of graph signal processing (GSP) tools in neuroscience has recently gained traction because they provide an analytical framework for subject-specific decomposition of functional signals, wherein different components are associated with varying degrees of conformity to the subject’s own brain anatomical network[23-25]. The components of the BOLD signal extracted from task-based functional magnetic resonance imaging (fMRI) that are less aligned with, or ‘liberal’ with respect to, the underlying white matter architecture have been linked with cognitive flexibility[24]. Furthermore, GSP tools have been used to find discriminating features from resting state fMRI and diffusion MRI (dMRI) in autism spectrum disorder[26] and traumatic brain injury[27]. GSP tools have also been used to evaluate the extent of structure-function decoupling for different brain regions[28]. Recent work has also applied GSP tools for the statistical analysis of functional activity with functional connectivity as the underlying graph, thereby implementing a unimodal analysis[29].

Our primary aim was to use statistical analyses to explore whether the associated features extracted from structural and functional brain imaging data are relevant in characterizing the serum NfL levels of healthy controls (HC) vs. former contact sports athletes with a prior history of concussions (ExPro). For both cohorts of subjects, we analyzed energy distributions of resting state fMRI after graph-informed filtering based on white matter connectivity extracted from dMRI. We hypothesized that measures of conformity of the BOLD activity with the underlying white matter anatomy in specific brain areas might reveal associations with serum NfL levels. GSP analytic tools leverage the spectral properties of the graph that represents the white matter anatomy to disentangle BOLD activity into components that either conform to or deviate from it. Low graph frequency components of the BOLD signal associated with a brain area characterize strong alignment between the functional coupling of this area to its underlying anatomical connectivity[24]. In contrast, high graph frequency components of the BOLD signal energy for an area signifies less intermodal alignment, i.e., loss or deviation of the functional coupling of a brain region to its anatomical connectivity[23]. The discriminating high and low graph frequency features between the two groups have been previously studied in Sihag et al.[30], and our focus here was on exploring their relationships with serum NfL levels in HC and ExPro cohorts.

To explore the clinical and neurological interpretation of the GSP features associated with serum NfL in our experiments, we also tested their associations with cognitive scores and structural measures. The two cohorts differed significantly in terms of aggression and depression related personality assessment scores (discussed in Section II.A.1). Therefore, it was of interest to explore the associations of GSP features with amygdala, since this region is instrumental in a broader neural circuit responsible for modulating aggression[31,32] and has been implicated in depression related disorders[33-35]. Furthermore, we hypothesized that the links between white matter degeneration and serum NfL levels might also be characterized by reduced cortical thickness[36]. In this context, we hypothesized that those GSP features aligned to our hypothesis of serum NfL being associated with conformity of BOLD activity to white matter anatomy might also be associated with cortical thickness measures. In a broader context, cortical thinning is also neurologically and clinically relevant, as it has been associated with structural abnormalities after TBI[37] as well as pathological personality traits[38].

Our results show that both low and high graph frequency components from different brain areas are relevant as features for the prediction of serum NfL levels in both HC and ExPro groups. Even more importantly, GSP features have different associations with serum NfL levels across the two groups and thus the statistical models for serum NfL levels are group specific: their performances do not broadly generalize to the combined dataset of HC and ExPro subjects. This observation is corroborated by our experiments on the behavioral scores under the umbrella of Personality Assessment Inventory (PAI) and structural metrics for both cohorts. The most striking observation is observed for the region of the isthmus cingulate in the right hemisphere, which shows distinct behavior in predicting NfL and associations with cognitive measures and structural measures in the two cohorts. Specifically, the GSP feature from this region is significantly associated with serum NfL levels in HC subjects but not in ExPro subjects. Moreover, our experiments indicate that this GSP feature has a suppressing statistical effect on the relationship between age and PAI Aggression score only among HC subjects. In ExPro subjects, this GSP feature has significant correlations with the volume of right amygdala (which is observed to be moderated by serum NfL levels in this cohort) and thickness of pericalcarine area in the right hemisphere (which is observed to be negatively correlated with serum NfL only in HC cohort). The two structural measures, volume of right amygdala and thickness of pericalcarine cortex, are associated with aggression as measured by PAI in both healthy subjects and pathological contexts[31,39,40], and therefore, our findings imply some significance to the GSP feature from isthmus cingulate in the right hemisphere in understanding behavior.

Methods

Participants

The study was approved by the research ethics boards of the University Health Network (IRB approval reference number: IRB 11-0088). Written consent was obtained from all subjects before participating in the study. The male healthy control subjects (number = 20, mean age = 49.38 years, standard deviation = 10.94 years) were recruited from the community. The subjects had no history of neurological disorders (e.g., seizure disorder), systemic illnesses known to affect the brain (e.g., diabetes and lupus), psychotic disorder, or known developmental disorders (e.g., attention deficit disorder, dyslexia) nor any lesions appearing on MRI. The male former athletes (number = 36, mean age = 50.64 years, standard deviation = 11.36 years) were former professional football, hockey or boxing athletes with history of multiple concussions (mean = 4.14, standard deviation = 1.7). There was no significant difference between the ages or serum NfL levels of the two groups (Mann Whitney U tests at 0.05 significance level). One subject each from the HC and ExPro group were excluded from the study since their serum NfL level was more than three deviations from the mean serum NfL level in their respective groups. There was no significant difference between the years of education for the two groups (HC: mean number of years of education = 16.4 years, standard deviation = 1.81 years, ExPro: mean number of years of education = 15.82 years, standard deviation = 1.68 years). Furthermore, no significant difference was observed in the cognitive scores in the contexts of memory, language, and visuospatial function for the two cohorts. Differences on inhibitory control, which is an executive function, have been reported previously on this sample[41].

PAI Assessments

The Personality Assessment Inventory (PAI) is a widely used and well-validated tool to study personality and psychopathology in brain injury[42]. It assesses Axis I and II disorders, including personality disorder, depression, aggression and anxiety, and includes indices of validity, such as positive and negative impression management. PAI scores on different subscales were evaluated for all participants. We observed statistically significant differences (p-value < 0.05 after false discovery rate (FDR) correction for multiple comparisons) in both raw and T-normalized PAI scores of the two cohorts for the subscales of somatic concerns, Depression, Schizophrenia and Aggression. Among these subscales, the Aggression and Depression subscales are considered valid in the context of TBI, as these were known to be not confounded by transdiagnostic measures characteristic of both psychopathology and neuropathology[42]. Table 1 summarizes the raw scores for HC and ExPro cohorts on the two subscales and the statistics for the Wilcoxon rank sum test for statistical differences between them.

Table 1

PAI Depression and Aggression subscale scores for HC and ExPro cohorts.

PAI scale	Score (HC)	Score (ExPro)	Rank sum statistic	p value (FDR corrected)
Depression	9.7 ± 10	17.81 ± 13	1218	0.0059
Aggression	9.2 ± 4.93	15.64 ± 9.54	1204.5	0.0234

Wilcoxon rank sum test (p value < 0.05 after FDR correction for multiple comparisons).

Diffusion magnetic resonance imaging acquisition and processing

All structural and resting state scans were performed on a 3 Tesla MRI Scanner (GE Signa HDx, Milwaukee, WI, USA) with a standard 8-channel head coil. A high resolution T1-weighted images were obtained using inversion recovery fast spoiled gradient echo (IR-FSPGR), with the following parameters: 180 slices with 1 mm thickness; 3 ms echo time (TE); 7.8 ms repetition time (TR); 450 ms inversion time (TI); 15 flip angle; 25.6 cm field of view (FOV); 256 × 256 matrix size; 1 × 1 × 1 mm3 voxel size. At least one DWI scan was obtained with diffusion gradients applied across 60 spatial directions (b = 1000 s/mm2) as well as 10 non-diffusion weighted (B0) scans. The DWI had the following parameters: 2.4 mm thick axial slices, TR = 14,000 ms, FOV = 23 cm, 2.4 × 2.4 mm2 in-plane resolution.

Diffusion MRI data were processed using the SCRIPTS pipeline with parameters as described therein[43]. Pre-processing involved correction for eddy-currents and head motions artifacts using FSL. After alignment of the co-registered dMRI to the T1 image, fiber tracking was performed using the MRtrix3 package. Fiber orientation estimation was performed using Constrained Spherical Deconvolution, and tracks were seeded from the white-gray matter interface. A propagation mask was applied through Anatomically Constrained Tractography (ACT) and streamlines were generated using a probabilistic algorithm using second-order integration over fiber orientation distributions (iFOD2) from 10 million seeds (step size 0.5 mm, maximum curvature 45, length 5–250 mm, FOD amplitude threshold 0.1). Streamlines were then selected using Spherical-deconvolution Informed Filtering of Tractograms (SIFT) to improve the fit between streamline reconstruction and the original dMRI image. The connectome weights were defined by the number of tracks going from one area of the parcellation mask to another, using the Desikan-Killiany atlas[44].

Functional magnetic resonance imaging acquisition and processing

The resting state functional MRI (rs-fMRI) scan acquisition was 5 min 8 s using T2*-weighted echo-planar imaging with the following parameters: TR = 2000 ms, TE = 30 ms, 64 × 64 matrix, 20-cm FOV, flip angle = 85, 40 slices, 3.125 × 3.125 × 4 mm3 voxels. Prior to the resting-state functional MRI scan, participants were instructed to close their eyes, not think of anything in particular, and to not fall asleep. Participants were spoken to between each sequence, and prior to each rs-fMRI scan, they were asked if the session could continue. The technicians did not proceed if the participant didn’t respond. Functional MRI data were processed using fMRIPrep, an open-source pipeline integrating multiple state-of-the-art fMRI tools into a single software suite[45]. Motion artifact correction and denoising were performed using ICA-AROMA, and susceptibility distortion corrections were performed using the SyN “fieldmap-less” correction method implemented in Advanced Normalization Tools (ANTs). Details on fMRIPrep processing are available in Supplementary Note 1.

BOLD time series of length 308 s (154 time points) were exported in CIFTI format, and the first 18 seconds were discarded to remove initialization transient artifact. In addition, the BOLD time series were pre-processed by removal of any linear trends and constant offsets and passed through a band-pass frequency domain filter with range 0.009 Hz–0.1 Hz. To account for any variations in the fMRI data across the subjects due to physical and physiological aspects of MRI scanning, the BOLD time series per area were z-score normalized for all subjects.

Serum neurofilament light protein concentration acquisition

Venous blood samples were collected from participants. Serum NfL concentration was measured using the Human Neurology 4-Plex A assay (N4PA) on an HD-1 Single molecule array (Simoa) instrument according to instructions from the manufacturer (Quanterix, Billerica, MA).

Data analysis

Graph signal processing-based feature extraction

We modeled brain anatomical areas and connectivity using graph structures whose nodes represent the 66 cortical regions of the Desikan-Killiany atlas and whose edges were their pair-wise connections. Connection weights were computed based on the number of tractography streamlines connecting brain areas, a proxy for alignment and density of fibers in the neuropil, such as axons[46]. For every subject, we used the eigenmodes (i.e. the eigenvalue-eigenvector pairs) of the graph adjacency matrix of the structural connectome to decompose the BOLD signals into low and high graph frequency components according to their conformity to the underlying white matter anatomical network.

Graph Fourier transform (GFT) provides the necessary framework to encode the spatial variability of a graph signal into graph frequencies[47] that are derived from the spectrum of the graph connectivity matrix. In this study, for every subject, we treated the BOLD time series over the brain structural connectome as a graph signal over the brain anatomical network and used the subject-specific spectrum of the brain connectivity matrix to construct graph filters using GFT. The graph filters allowed us to extract different components of the BOLD time series in different brain areas according to their spatial variability. For instance, the energy of the extracted component corresponding to low spatial variability of the BOLD series in a brain area represented the extent to which the BOLD time series in that area conformed to the topology of the brain structural connectome[24].

We next provide the mathematical framework behind the GFT and its application for decomposition of BOLD time series. The brain structural connectome can be represented by an undirected graph G with n nodes, where each node is associated with a distinct brain area. The adjacency matrix of G is given by A, which is an n × n matrix whose off-diagonal entries is a proxy for the number of axonal connections between different pairs of brain areas. Since A is symmetric due to inherent limitations of tractography, it can be decomposed as , where the eigenvectors of A form the columns of and the eigenvalues of A are the elements of the diagonal matrix , s.t., . The formal definition of GFT based on the adjacency matrix is as follows[25]: Given a graph signal and the adjacency matrix , the GFT pair is given by

The eigenvectors of A form the spectral components of the graph and the eigenvalues of A form the graph frequencies. The eigenvector-eigenvalue pairs, , of A are termed as the eigenmodes of the graph G and are analytically related to the spatial variation of the graph signal (see Supplementary Note 2).

The application of GFT allows us to extract different components of the BOLD time series according to their spatial variation with the help of graph frequency filters of the formwhere is the frequency response for the eigenmode . For a given spatial vector x over the graph, its graph filtered output y is given by

As an example, the design of a high pass graph filter based on the adjacency matrix that passes the component corresponding to 10 highest graph frequencies is given by

Although the degree of spatial variability of the BOLD time series with respect to the brain anatomical network varies over a continuum of intermodal ‘alignment’[23] or conformity with the brain anatomy, previous studies have demonstrated that the components of graph signals with low or high spatial variability have better and more reliable performance in inference tasks based on neuroimaging data[23,25]. Therefore, in this study, we focused only on the components of BOLD time series with low or high spatial variability.

For each subject, we used the subject-specific graph filter that passed the 10 highest (or lowest) graph frequencies to extract the high (or low) graph frequency components of the BOLD time series. Note that the application of GFT leverages the connectivity of the brain anatomical network to decompose the BOLD time series signal in every TR and therefore, the output obtained after application of a graph frequency filter at any brain area is sensitive to the variation in the BOLD signal with respect to that in the other brain areas[23]. For every area, the application of a low pass graph frequency filter isolates the proportion of its BOLD time series that conforms to the topology of the anatomical network and that of a high pass graph frequency filter isolates the proportion of its BOLD time series that deviates significantly from the topology of the anatomical network. An example of application of high and low pass graph filters on BOLD data is illustrated in Supplementary Fig. 1.

For each brain area, we evaluated the energies of the components with low spatial variability and high spatial variability by calculating the norm of the respective graph frequency components of the BOLD time series. Therefore, two features were associated with every brain area for each subject leading to 132 GSP features (66 each from high and low graph frequency analysis) per subject. The group differences between the GSP features in this sample have been reported in our previous study[27].

Serum NfL level and GSP features

Prediction and inference form the two paradigms of statistical analysis that provide distinct insights into the relevance of variables depending on the actual modeling goal[48]. Inference helps in isolating individual variables that are significantly associated with the target variable (in this case, serum NfL) whereas prediction driven analysis guides the isolation of variables deemed relevant for predicting the target variable in unseen data. In this study, we aimed to explore the statistical correspondence of GSP features in the context of serum NfL levels in the two cohorts under both statistical paradigms. Due to lack of neuroimaging studies that link specific brain regions with serum NfL, we adopted data-driven approaches to isolate the GSP features that were most relevant to serum NfL from the complete set of 132 features.

For the inference paradigm, we adopted a standard linear model based univariate feature selection approach to isolate the GSP features most significantly associated with serum NfL[49]. This approach results in an F-value and a p-value for each GSP feature, whose statistical significance was determined after false discovery rate procedure for correction due to multiple comparisons. Similar approaches have been adopted previously to select the most relevant features from features extracted using GFT of neuroimaging data for various statistical inference tasks[50].

Given the fact that the number of GSP features (132) outnumbered the number of data samples in both cohorts (20 for HC and 36 for ExPro), we adopted PLSR analysis in our study for prediction paradigm of statistical analysis because of its recommended usage in the neuroimaging literature for scenarios with high multicollinearity among predictors and when the number of predictors outnumber the number of data samples[51,52].

The input and output features were z-score normalized for PLSR analysis. For each group, the PLSR model that fit all the GSP features to serum NfL levels of their respective groups was investigated first. In this context, the number of components for the PLSR model for a given set of features was selected based on the estimated mean square prediction error (MSEP) as the criterion which was evaluated by leave-one-out cross validation. Since the total number of GSP features (132) far exceeded the number of available observations for both groups, the PLSR model with a full set of independent variables was prone to overfitting, which was also confirmed by a nonparametric permutation test of the explained variance R2. The nonparametric permutation test for evaluating the significance of R2 for a PLSR model is described next.

Nonparametric permutation test for PLSR: The nonparametric permutation test was carried out by randomizing the serum NfL levels among the subjects and fitting them to the predictors using a PLSR model. The null distributions of R2 were obtained by evaluating the explained variance for 5000 random permutations of the NfL levels. The PLSR models were considered to be overfit for a given set of predictors if the null distribution of the explained variance R2 had a mean >0.5. The statistical significance of R2 values for a PLSR model at a given level was evaluated by counting the number of samples in the corresponding null distribution that exceeded it.

Variable selection for prediction model: For both groups, we adopted a variable importance in projection (VIP) based approach for selecting a subset of GSP features that could constitute the PLSR model that fits serum NfL without overfitting[53]. VIP score quantifies the relative importance of each predictor in fitting the PLSR model and was calculated for each predictor based on the PLSR model that fitted the serum NfL levels to 132 predictors for each group. A feature with a high VIP score is typically considered relatively more significant for the prediction performance of the PLSR model[54]. For each group, we varied the threshold of the VIP score and used the features with a VIP score greater than the selected threshold to form the PLSR model with one component. The number of components was set to one due to limited data size.

Prediction performance based on cross validation: We evaluated the prediction performance of the models based on leave-one-out cross validation procedure. In both cohorts, for every subject, the non-overfitted model trained on the rest of the subjects with the best prediction performance on that subject’s serum NfL was selected as the ‘best’ model. For both cohorts, the following procedure was followed to calculate the Q2 value for the model. At every instance of cross validation, the serum NfL level for one subject was estimated by the PLS model fitted to the data for the rest of the subjects. The variable selection procedure described above was performed within every instance of cross validation, i.e., VIP scores were evaluated using the serum NfL levels and the GSP features for the subjects in the training set at every cross-validation instance. The set of features for which the PLSR model was not overfitted was chosen to estimate the serum NfL level for the test subject. Therefore, there were 20 PLS models for the HC cohort corresponding to each instance of cross validation. Similarly, there were 36 PLS models for the ExPro cohort corresponding to each instance of cross validation in this cohort. We report the prediction performance from this cross-validation procedure for both cohorts which is quantified by their Q2 values. For any GSP feature, its frequency of inclusion in the PLS models for prediction of serum NfL for different subjects in cross validation, and similar trends in its respective weights across models, indicates its robustness as a predictor of serum NfL. An overview of the statistical analyses to investigate links between GSP features and serum NfL levels is provided in Fig. 1.

Fig. 1

Pipeline for statistical analysis for serum NfL levels and GSP-based features.

Brain imaging data (structural (dMRI) and functional (resting state fMRI)), age, and serum NfL levels were recorded for every subject in the cohorts of 20 healthy subjects and 36 former athletes. For each subject, the dMRI and fMRI were pre-processed to extract the structural connectome (in the form of a 66 × 66 adjacency matrix for 66 cortical brain areas) and BOLD signal (in the form of a time series of length 308 seconds at each brain area considered), respectively. Subject-specific graph filters derived from the eigen-decomposition of the structural connectome were constructed and used to extract different graph frequency components of the BOLD time series for all subjects. The energies (EHI and ELO, evaluated by norm) of the different graph frequency components at different brain areas were investigated (energies of 2 graph frequency components per area for 66 brain areas per subject, i.e., 132 GSP-based predictors) for association with serum NfL levels using univariate feature selection (UFS), and their prediction power evaluated using a rigorous statistical analysis of PLS regression models.

Clinical and neurological interpretations of GSP features linked with serum NfL

We used partial correlation, mediation, and moderation analyses to interpret the roles of GSP features that were relevant for serum NfL for both inference and prediction analyses in the two cohorts. Specifically, we investigated whether the GSP features mediated any associations between age, serum NfL, and PAI scores. We also investigated the relationships between the GSP features and structural measures such as cortical thickness and volumes of subcortical regions.

For mediation analysis, we used the mediation toolbox from Wager et al.[55]. The significance of the mediation was established using bootstrapping with 10000 samples. Moderation analysis was conducted based on linear regression and moderation effect was determined based on the significance of the interaction term in the linear model. The reporting of methods and results in this paper adhere to the STROBE guidelines[56].

Table 2

Statistics for mediation analysis between age and serum NfL with GSP feature from right isthmus cingulate as mediator variable for HC subjects.

	Coefficient	Std. error	p value(uncorrected)
Path a	−0.05	0.01	0.0038
Path b	−2.20	0.84	0.0001
Path c’ (adjusted effect)	0.16	0.06	0.0292
Mediation (ab)	0.11	0.03	0.0014
Path c (total effect) (age -> serum NfL)	0.2658	0.0646	0.0012

Table 3

Statistics for mediation analysis between age and PAI Aggression score with GSP feature from right isthmus cingulate as mediator variable for HC subjects.

	Coefficient	Std. error	p value (uncorrected)
Path a	−0.052	0.0139	0.0035
Path b	4.353	1.217	0.0032
Path c’ (adjusted effect)	0.2193	0.0806	0.0029
Mediation (ab)	−0.2326	0.0961	0.0122
Path c (total effect) (age->PAI Aggression score)	−0.0133	0.0988	0.9761

Table 4

Statistics for moderation analysis for interaction between GSP feature from right isthmus cingulate and right amygdala volume for ExPro subjects.

	Coefficient	Std. error	p-value (uncorrected)
Age	−11.092	25.51	0.0003
Serum NfL	8.0921	7.35	0.27
r-iCg GSP feature	30.249	27.869	0.28
Interaction (serum NfL X r-iCg feature)	16.495	5.51	0.0054
Intercept	1679.6	65.84	7.07e–35