https://orcid.org/0000-0002-1131-5273
Figures
Abstract
In this study, structural images of 1048 healthy subjects from the Human Connectome Project Young Adult study and 94 from ADNI-3 study were processed by an in-house tractography pipeline and analyzed together with pre-processed data of the same subjects from braingraph.org. Whole brain structural connectome features were used to build a simple correlation-based regression machine learning model to predict intelligence and age of healthy subjects. Our results showed that different forms of intelligence as well as age are predictable to a certain degree from diffusion tensor imaging detecting anatomical fiber tracts in the living human brain. Though we did not identify significant differences in the prediction capability for the investigated features depending on the imaging feature extraction method, we did find that crystallized intelligence was consistently better predictable than fluid intelligence from structural connectivity data through all datasets. Our findings suggest a practical and scalable processing and analysis framework to explore broader research topics employing brain MR imaging.
Citation: Kopetzky SJ, Li Y, Kaiser M, Butz-Ostendorf M, for the Alzheimer’s Disease Neuroimaging Initiative (2024) Predictability of intelligence and age from structural connectomes. PLoS ONE 19(4): e0301599. https://doi.org/10.1371/journal.pone.0301599
Editor: Constantine Dovrolis, Georgia Institute of Technology, UNITED STATES
Received: April 6, 2023; Accepted: March 19, 2024; Published: April 1, 2024
Copyright: © 2024 Kopetzky et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The authors do not have permission to share the data as it is 3rd party data. The authors did not have special access privileges that other researchers would not have. The HCP data can be accessed from https://db.humanconnectome.org/app/template/Login.vm. Preprocessed HCP data from braingraph.org can be accessed from https://braingraph.org/cms/download-pit-group-connectomes/. ADNI-3 data can be accessed from https://adni.loni.usc.edu/data-samples/access-data/. The lists of the 1064 HCP and 94 ADNI subject IDs used for the analysis are available as supplementary material. The code is available as supplementary material.
Funding: This research has received funding from the European Union’s Horizon 2020 Framework Programme for Research and Innovation under the Specific Grant Agreement No. 945539 (Human Brain Project SGA3). This work was supported by Eurostars project E! 113682 MS-CONNECT through the German Federal Ministry of Education and Research (BMBF) under grant number 01QE2025A. Marcus Kaiser was supported by a Medical Research Council (MRC) grant (MR/T004347/1) and by the Guangci Professorship Program of Ruijin Hospital (Shanghai Jiao Tong University). There was no additional external funding received for this study.
Competing interests: SJK and YL were employed with Labvantage – Biomax GmbH and therefore will be affected by any commercial implications caused by this manuscript. This does not alter our adherence to PLOS ONE policies on sharing data and materials.
Introduction
The connectome—the entire map of neural connections—uniquely represents every subject’s gender, age and intelligence like a fingerprint [1]. Intelligence is known to be affected by, e.g., topological properties of brain networks such as characteristic path length and global network efficiency, respectively [2, 3]. The association between a lower characteristic path length and IQ has also been described for resting state functional MR imaging (rs-fMRI) networks [4]. Predicting not only gender [5, 6] and age [7, 8] but also different forms of intelligence [5, 6, 9–11] in individual subjects made significant progress by using rs-fMRI. However, far less is known to what degree the underlying structural connectome, the backbone of the functional interactions, is also predictive of age and intelligence in cognitively normal adults. Several studies have tested the predictability of brain age using more advanced machine learning models. For example, Lin et al. predicted older individuals’ age using artificial neural networks [12] and Taoudi-Benchekroun et al. used deep neural networks and random forests to predict the age of infants [13].
Previous studies suggest a distinct nature with normal aging between crystallized and fluid intelligence [14]. Fluid intelligence shows one’s ability to acquire new knowledge and is reflected in problem-solving and adaptation to unknown environments therefore it examines cognitive tasks such as cognitive flexibility, working memory, and information processing speed, while crystallized intelligence more reflects experience-based knowledge and the ability to access it and is e.g. measured by vocabulary and decoding tasks [15–18]. Shokri-Kojori et al. [19] compared age-related variance between younger and older adults (100 subjects) for gray matter (GM) and white matter (WM) tissue-specific age scores. They found that the WM age score accounted for significantly more variance in chronological age and was negatively associated with crystalized intelligence in older adults. Góngora et al. [20] reconstructed 10 tracts by deterministic tractography in 83 healthy individuals from the Cuban Human Brain Mapping Project. Their results showed predictive effects of the forceps minor tract on crystallized intelligence and of the superior longitudinal fasciculus on fluid intelligence.
In this study, we explored how well different intelligence measures and age of cognitively normal adult subjects can be predicted from the structural connectome as quantified by diffusion-weighted imaging (DWI). Specifically, we processed DWI data of HCP young adult (1048 subjects) and ADNI-3 (94 cognitive normal elderly subjects) datasets to reconstruct whole brain structural connectome features by our in-house tool, NICARA. Then, we used NICARA extracted features to apply the correlation-based regression (CBR) machine learning method [21] to predict age as well as total, fluid and crystallized intelligence. A similar approach was also suggested by Shen et al. [22]. To further explore the predictive capabilityof the CBR ML model and allow for additional statistical comparisons, we also included the structural connectome features of the HCP dataset preprocessed by braingraph.org available with different parcellations.
Materials and methods
Ethics statement
According to national law and institutional rules research involving the analysis of existing data, where the data is either already publicly available or will be analyzed such that individual subjects cannot be identified is exempt from IRB oversight.
Datasets
We investigated the prediction capability for different intelligence measures and age of two different DTI pipelines (NICARA [23] and braingraph.org [24]) based on structural connectivity data of 1048 subjects from the Human Connectome Project (HCP) young adult study [25]. Investigated features were age, total intelligence, fluid intelligence, and crystallized intelligence (Table 1). The intelligence measures were unadjusted cognitive function composite score, fluid cognition composite score, and crystallized cognition composite score, respectively, based on the NIH toolbox [15].
The braingraph.org database was constructed by deterministic ROI-based fiber tracking (10 × averaged) of 1064 healthy subjects from the HCP Young Adult dataset. It was available with five different sets of ROIs (86, 129, 234, 463 and 1015 ROIs) and therefore also provides a good opportunity to investigate the influence of the parcellation method on the prediction outcome.
The full HCP dataset consisted of 1065 subjects, and we excluded subjects for whom no data was available from braingraph.org (2 subjects) and for whom not all investigated features were available (15 subjects). After this, 1048 subjects (S1 Table) remained for the analysis (484 male, 564 female).
Since the age range of subjects from the HCP Young Adult cohort is relatively small (22–37 years with a mean of 28.75 years and a standard deviation of 3.68 years; Table 1), we decided to additionally include a second cohort with adult and aged healthy subjects from the ADNI-3 study of the Alzheimer’s Disease Neuroimaging Initiative [26]. We used this second cohort (S2 Table) with 94 control subjects (40 male, 54 female) with an age range from 56.5 to 91.5 years (mean: 74.43 years, standard deviation: 7.86 years; Table 2) to confirm age predictability from connectomes extracted by our in-house pipeline. Fig 1 shows the age distribution of male and female subjects from both studies.
Data processing
We processed structural connectomes of 1048 subjects from the HCP young adult study as well as the 94 ADNI subjects. In-house processing pipelines used 379 ROIs with 360 cortical ROIs (HCP-MMP1.0, Human Connectome Project Multi-Modal Parcellation version 1.0 atlas [27] and 19 subcortical ROIs (Harvard-Oxford subcortical atlas [28–30]), and 50 million streamlines for the probabilistic whole brain tractography. As comparison datasets to better assess the influence of different parcellations, we downloaded preprocessed connectivity data derived from the same HCP subjects available from the braingraph.org database [24] that have been obtained with different processing routines. braingraph.org datasets were available with five different numbers of ROIs (86, 129, 234, 463 and 1015 ROIs).
In our in-house processing pipeline, at first the T1w images get defaced using SPM12 (https://www.fil.ion.ucl.ac.uk/spm/). The T1w and DTI images were co-registered, and the images in native space were transformed into the MNI152 space by normalization. Tissue-based segmentation was performed using the CAT12 toolbox [31]. Defaced T1w images got skull-stripped using the adaptive probability region-growing (APRG) approach and used as input for FreeSurfer [32], which was used to apply cortical & subcortical parcellation in order to project the HCP MMP 1.0 and Harvard-Oxford subcortical atlas to native space via “fsaverage”. Noise and distortion correction methods were applied to the DWI images using mrtrix3 [33], FSL [34], and ANT (Advanced Normalization Tools) [35]. Anatomically Constrained Tractography (ACT) was applied using mrtrix3 with 50 million streamlines yielding the 379 x 379 connectivity matrices used for the analysis. All processing steps were assembled to a stand-alone pipeline optimized for automated execution (referred to as NICARA or in-house pipeline) and run by our proprietary neuroimaging solution NICARA Version 2.0, Labvantage—Biomax GmbH, Planegg Germany (https://nicara.eu).
Feature prediction
In order to predict features based on the structural connectivity matrices, we applied the correlation-based regression algorithm proposed by Han et al. [21].
Given n subjects and m edges (depending on the brain parcellation), we first obtained a correlation vector R of size m from the correlation between the matrix A of unnormalized connection strengths of size n x m (containing the vectorized connectivity matrices per subject as rows) and the attribute of interest (Eq 1).
The correlation vector R contains the Pearson’s correlation coefficients r (Eq 2) between Ai (with Ai being the ith column of matrix A) and b, a vector of length n containing the values of the attribute of interest (e.g. age) for each subject.
Subsequently, we calculated the vector of predictor values S of size n for the regression analysis by summing up each subject’s connection strengths weighted by the respective correlation coefficient of the edge (Eq 3).
The predictor value Sj for subject j is the scalar product of the vectors Aj and Rx.
Based on the predictor values we fitted a simple linear least-squares regression model that was then used for predicting the feature value of interest. To assess and compare the prediction quality, a 10-fold cross-validation was performed for each dataset. R was calculated from the training data only and the Pearson’s r between predicted and actual values as well as the mean absolute error (Eq 4) and range-normalized mean absolute error were calculated from the ten iterations of the cross-validation.
The mean absolute error (MAE) is the norm of the difference of predicted values g and the real values h. The range-normalized mean absolute error (NMAE) can be obtained by dividing the MAE by the difference of the maximal and minimal value of the data and therefore allows for a better comparability between different datasets (unpaired data).
Statistical analysis
Statistical analysis was performed based on the paired absolute errors of subjects obtained from the cross-validation using the Wilcoxon signed-rank test. Comparisons were done between the different datasets of the same feature to determine differences in the as well as within each dataset between crystallized and fluid intelligence. We applied the Benjamini-Hochberg multiple testing correction with a false discovery rate (FDR) of 0.05 to the p-values of each analysis.
Results
We applied the correlation-based regression algorithm (S1 File) to predict age for different HCP datasets comprising 1048 subjects (NICARA, braingraph.org datasets) and a subset of the ADNI study comprising 94 subjects, as well as different intelligence measures in case of the HCP datasets. Group comparisons were performed based on the mean absolute error (Eq 4) as well as the Pearson correlation coefficient between the actual and predicted values based on a regression analysis of the total data as well as gender-specific subgroups. Then we applied the Wilcoxon signed-rank test for age, total, crystallized and fluid intelligence to identify differences in the paired absolute errors over the different pipeline conditions. Finally, Wilcoxon signed-rank tests of the paired absolute errors revealed significant differences between crystallized and fluid intelligence in all comparisons.
Age prediction
For the HCP datasets, the maximal observed Pearson correlation coefficients between the features and age per dataset ranged from 0.105 for the 129 node braingraph dataset to 0.227 for the in-house pipeline (S3 Table), while the minimal Pearson coefficients ranged from -0.104 for the in-house pipeline to -0.168 for the 86 and 129 node braingraph datasets (S4 Table). The maximum for the ADNI dataset was r = 0.619 and the minimum r = -0.550, therefore stronger compared to the values observed in the HCP datasets.
Our in-house pipeline achieved the highest Pearson correlation between actual and predicted age values (r = 0.21) and the lowest mean absolute error (MAE = 3.02) out of the datasets (Table 3; Fig 2). Within the braingraph datasets, the MAE slightly increased with the number of ROIs (3.02 to 3.13). Based on the Wilcoxon signed-rank test for the paired absolute errors of the full datasets, the difference between the NICARA dataset and the following braingraph datasets was significant before a multiple testing correction: braingraph 129 ROIs (p = 0.0251), braingraph 234 ROIs (p = 0.0245), braingraph 463 ROIs (p = 0.0127) and braingraph 1015 ROIs (p = 0.0085). Within the braingraph datasets, the following were significant: 86 vs 1015 ROIs (p = 0.0350), 234 vs 463 ROIs (p = 0.0478) and 234 vs 1015 ROIs (p = 0.0251). However, after applying the Benjamini-Hochberg multiple testing correction (FDR = 0.05), none of the p-values remained significant. In the gender-specific subgroups, the following two datasets showed a significant p-value for males before but not after multiple testing correction: 86 vs 129 ROIs (p = 0.035) and 86 vs 234 ROIs (p = 0.006).
The x-axis shows the actual age in years and the y-axis the predicted age in years during the cross-validation.
The prediction result for the ADNI data (Fig 3) achieved a higher correlation value between predicted and real values (r = 0.57) compared to the HCP datasets (rmax = 0.22) and a lower NMAE (0.14) compared to the best value out of the HCP datasets (0.20).
Prediction results for the ADNI dataset (age) for the total dataset (bottom) and the gender-specific subsets (top). The x-axis shows the actual age in years and the y-axis the predicted age in years during the cross-validation.
Dividing the data into gender-specific groups slightly improved the prediction quality for the ADNI data in case of females (rF = 0.68, NMAEF = 0.14), but slightly decreased it for the male-only subjects (rM = 0.52, NMAEM = 0.17; Table 4). For the HCP data, gender-specific analysis led to inconsistent results and mostly worse prediction outcome when compared to the total dataset, but the female subgroup tended to perform better than the male subgroup.
Intelligence prediction
The maximal observed Pearson correlation coefficients between the features and intelligence per dataset ranged from 0.127 (129 node braingraph dataset) to 0.168 (1015 ROI braingraph dataset) for total intelligence, from 0.108 (86 ROI braingraph dataset) to 0.139 (1015 ROI braingraph dataset) for fluid intelligence, and from 0.142 (86 ROI braingraph dataset) to 0.189 (1015 ROI braingraph dataset) for crystallized intelligence (S3 Table), The minimal observed Pearson correlation coefficients between the features and intelligence per dataset ranged from -0.153 (86 ROI braingraph dataset) to -0.235 (in-house pipeline) for total intelligence, from -0.129 (86 ROI braingraph dataset) to -0.185 (in-house pipeline) for fluid intelligence, and from -0.151 (129 ROI braingraph dataset) to -0.236 (NICARA) for crystallized intelligence (S4 Table).
The prediction for the different intelligence measures performed similarly between all available datasets. Data from the in-house pipeline showed the lowest MAE for total (11.47) and fluid (9.36) intelligence while the 129 ROI braingraph dataset had the lowest MAE (7.67) in case of crystallized intelligence (Tables 5–7).
The braingraph dataset with 1015 ROIs showed the highest correlation values for all intelligence measures for all subjects with r = 0.24 (Tables 5–7; Figs 4–6). Within the braingraph datasets, increasing ROI number showed the tendency to increase the MAE but also slightly the correlation between actual and predicted values. For total intelligence, only the absolute error differences between the NICARA and 1015 ROI braingraph dataset was significant based on the Wilcoxon signed-rank test (p = 0.0411). For fluid intelligence, there were only significant differences for 234 vs. 1015 ROI braingraph (p = 0.0341) and 463 vs 1015 ROI braingraph (p = 0.0140). Lastly, for crystallized intelligence, several absolute error differences were significant: NICARA vs. 463 ROI braingraph (p = 0.0489), NICARA vs. 1015 ROI braingraph (p = 0.0195), 86 vs 1015 ROI braingraph (p = 0.0280), 129 vs 234 ROI braingraph (p = 0.0399), 129 vs. 1015 ROI braingraph (p = 0.0125), 234 vs 1015 ROI braingraph (p = 0.0238), and 463 vs. 1015 ROI braingraph (p = 0.0273). None of the p-values remained significant after applying the Benjamini-Hochberg multiple testing correction (FDR = 0.05).
The x-axis shows the actual values and the y-axis the predicted values during the cross-validation.
The x-axis shows the actual values and the y-axis the predicted values during the cross-validation.
The x-axis shows the actual values and the y-axis the predicted values during the cross-validation.
For the comparisons of absolute errors in the gender-specific subgroups for total intelligence, 129 vs 234 ROI braingraph (p = 0.045) was significant in the male subgroup and 86 vs 463 ROI (p = 0.0362), 86 vs 1015 ROI (p = 0.0128), 129 vs 1015 ROI (p = 0.0223), and 234 vs 1015 ROI (p = 0.0231) in the female subgroup before but not after multiple testing correction. For crystallized intelligence, 129 vs 234 ROI (p = 0.0417) showed significance in the male subgroup and the following in the female subgroup: NICARA vs 234 ROI (p = 0.0253), NICARA vs 463 ROI (p = 0.0252), and NICARA vs 1015 ROI (p = 0.0262). None of the p-values remained significant after multiple testing correction. For fluid intelligence in the male subgroup, the following comparisons were significant before multiple testing correction: NICARA vs 234 ROI (p = 0.0285), 86 vs 129 ROI (p = 0.0357), 86 vs 234 ROI (p = 0.0178), 129 vs 1015 ROI (p = 0.0140), 234 vs 463 ROI (p = 0.0107), 234 vs 1015 ROI (p = 0.0010), 463 vs 1015 ROI (p = 0.0332). The comparison between 234 and 1015 ROI braingraph datasets did remain significant even after the multiple testing correction. In the female subgroup for fluid intelligence, the following comparisons were significant before but not after multiple testing correction: 86 vs 463 ROI (p = 0.0227), 86 vs 1015 ROI (p = 0.0058), 129 vs 463 ROI (p = 0.0376), 129 vs 1015 ROI (p = 0.0110), 234 vs 463 ROI (p = 0.0496), 234 vs 1015 ROI (p = 0.0086), and 463 vs 1015 ROI (p = 0.0313).
Considering the relatively good Pearson correlation scores of NICARA among gender-specific groups versus pipeline conditions for both crystallized and fluid intelligence, we report two interesting trends from the gender-specific results by Wilcoxon signed-rank test of the absolute errors: The first trend shows difference in female group between NICARA and 234 ROI braingraph dataset for crystallized intelligence (p = 0.025; Fig 7). And the second trend reveals a difference in the male group for same comparisons of fluid intelligence (p = 0.0285; Fig 8).
Total dataset: green, Female subset: blue, Male subset: red. *: p < 0.05 without correction based on the Wilcoxon signed-rank test for paired absolute errors.
Total dataset: green, Female subset: blue, Male subset: red. *: p < 0.05 without correction based on the Wilcoxon signed-rank test for paired absolute errors.
Since the cognition fluid and crystallized composite score have the same range and we also applied the Wilcoxon signed-rank test to the absolute error differences between them. The differences were significant for all datasets (Table 8) and all of them remained significant after a Benjamini-Hochberg multiple testing correction (FDR = 0.05).
Gender-specific analysis led to a worse prediction outcome for males in case of total intelligence for all datasets compared to the results for all subjects. For females, the outcome improved for all datasets except for the 1015 ROI dataset based on the NMAE, but the correlation value only improved for the NICARA dataset compared to the total data. For crystallized and fluid intelligence, gender-specific analysis did not improve the outcome. In case of fluid intelligence, the correlation values between actual and predicted values for the male subgroup in the braingraph datasets were not even significant anymore (Table 7).
Discussion
In this study, we investigated the predictability of age and intelligence measures based on whole brain structural connectome measurement under various conditions of imaging processing and gender. First, we demonstrated that NICARA processing combined with our machine learning approach could well predict age of ADNI subjects but performed worse for HCP subjects consistent with braingraph.org results. Second, crystallized intelligence was better predicted than fluid intelligence in general. Finally, we found two interesting trends of gender effects for fluid and crystallized intelligence predictability that did not remain significant after multiple testing correction.
Using a simple machine learning approach based on whole brain structural connectivity only we were able to decently predict age for cognitively normal ADNI-3 control subjects (N = 94) with a distinct age range (r = 0.57, NMAE = 0.14). Our finding is similar to that of recent studies both in MAE and std of age prediction using structural connectivity data of ADNI [36, 37], although both studies employed different workflows to extract structural connectivity features and different prediction methods.
However, the prediction quality for age was worse in case of the HCP young adult dataset (r = 0.21, NMAE = 0.20), which only covered a narrow range of young subjects, for both investigated pipelines (in-house and braingraph.org) under different parcellation methods despite its large size (N = 1048). Multi-factor characteristics of HCP Young Adult dataset structural connectivity could account for limited predictability of age. This effect of the HCP dataset is evident by comparable age prediction quality from distinct processing approaches of NICARA versus braingraph.org.
For subjects of the HCP Young Adult study, total, fluid and crystallized intelligence values were measured as cognition composite scores. The overall low predictability of intelligence by whole brain structural connectome features confirmed the statement of Wu and colleagues’ work [38] that their prediction combining cortical and subcortical surfaces together yielded the highest accuracy of fluid intelligence for both ABCD (N = 8070, r = 0.314) and HCP datasets (N = 1097, r = 0.454), outperforming the state-of-the-art prediction of fluid intelligence from any other brain measures in the literature. Wu and colleagues developed a novel graph convolutional neural networks (gCNNs) for the analysis of localized anatomic shape and prediction of fluid intelligence. Krämer et al. [39] reported a similar outcome to this study for a small trend for a multimodal benefit therefore concluded that developing a biomarker for cognitive aging remained challenging. Their study employed multimodal information, i.e., region-wise grey matter volume (GMV), resting-state functional connectivity (RSFC), and structural connectivity (SC), and generalized results across different ML approaches in 594 healthy older adults (age range: 55–85 years) from the 1000BRAINS dataset.
Predictability of crystallized intelligence was better than that of fluid intelligence in all investigated datasets based on the Wilcoxon signed-rank test applied to the paired absolute errors (Table 8), which may imply a stronger relation between crystallized intelligence and whole brain white matter probabilistic tractography based connectivity than that of fluid intelligence. Similar findings have been published by investigating distinct multi-region neuroanatomical patterns extracted from grey matter surface as well as volumetric assessments in 1089 HCP subjects by employing an elastic net regression model [40]. Our finding is also consistent with another study that investigated 415 HCP subjects by a similar tractography method but using only 86 ROIs from FreeSurfer [41]. A much finer atlas parcellation with 439 ROIs created in that study did not show a significant difference in predictability between crystallized and fluid intelligence [41]. Other papers considered highly correlated neuroanatomical morphometry profiles, i.e. cortical surface area, and the environmental impact on the relevant neuroanatomical morphometry as explanations for better predictability of crystallized intelligence compared to fluid intelligence [40, 42].
Interestingly, we find tendencies of slightly better prediction of intelligence in NICARA by gender-specific subgroup analysis based on the range-normalized mean absolute errors and pearson correlation between actual and predicted values. This supports the finding that gender specific factors [43–48] may affect connectivity as well as the relationship between connectomics and cognition [11].
With the publicly available braingraph.org dataset, connectome data from a second structural connectome extraction method for the same HCP subjects was available with five different parcellations (86, 129, 234, 463 and 1015 ROIs). Within the braingraph datasets, the MAE showed the tendency to increase with a higher ROI number. This may be due to the parcellation method itself, or the fact that the resulting connectivity matrix might be too sparse using only 1 million streamlines for finer parcellation [41, 49–51]. However, the differences between the different braingraph datasets were not significant after correcting the p-values obtained by Wilcoxon signed-rank tests between the paired absolute errors for multiple testing using the Benjamini-Hochberg method (FDR = 0.05).
Limitations
An obvious shortcoming of the HCP Young Adult subjects for prediction of age and intelligence are the biases in the data distribution. The data only contained young subjects (aged 22–37 years), and the cognition composite scores are clearly biased towards higher values. 100 is the United States average but the mean values for the three scores of the 1048 subjects were all higher than that (Table 1). We could show the influence of the age range by also applying the approach to control subjects from the ADNI-3 study with an age range about twice as large (Δ = 35.0 years vs. Δ = 15 years) resulting in a better prediction performance (reflected by a lower NMAE and a higher correlation between actual and predicted values), even though these subjects were also biased as they contained elderly individuals only (56.5–91.5 years).
Apart from the bias, a possible explanation for the low predictability of intelligence in this study can be seen in the structural correlates of intelligent behavior. Human intelligence is thought to be associated with physiological and morphological properties of cortical pyramidal neurons [52], which, of course, can only be captured indirectly by whole brain DTI-based fiber tracking.
Future approaches
To certify the age prediction, it would be favorable to apply the proposed approach to a dataset with a wider age range such as the Lifespan Human Connectome Project in Aging [53] once it is complete and available. Moreover, it has been previously discussed whether sets of connections would uniquely map onto cognitive function [54]. Further work could elaborate on this idea to test whether specific functional brain networks [55] and their connectedness are more predictive than individual or sets of connections alone. From studies with multiple sclerosis patients, it is known that graph theoretical measures are better descriptors of cognitive decline [56, 57] than the strengths of individual connections. For future studies on prediction qualities of the connectome for cognitive measures it would therefore be favorable to include graph theoretical measures [58] as well. Combining structural and functional brain networks for intelligence prediction revealed so far ambivalent results [40]. Here is certainly more work needed, for example, by using brain multiplex networks [59] to increase predictive power of multi-modal connectomes. Being able to predict brain age and intelligence from brain connectivity has a large impact on monitoring disease progression in dementia or other brain diseases associated with cognitive decline, for example, multiple sclerosis.
Conclusion
In conclusion, this study explores the predictability of age and intelligence in cognitive normal subjects from the HCP and ADNI datasets by means of whole brain tractography-based structural connectome applying a simple and easy-to-use established machine-learning method. To our knowledge, this is the first study focusing on a single neuroimaging modality feature to predict age and intelligence from whole-brain tractography. The good predictability of age in ADNI and the finding that crystallized intelligence was better predictable than fluid intelligence in HCP datasets was possible by the combination of the NICARA structural connectome pipeline and a simple machine learning model. Therefore, we believe that such a combination could provide a reliable framework option to further narrow down the gap in prediction between neuroimaging features and subjects’ cognition as well as other biological features.
Supporting information
S1 Table. 1048 HCP subject IDs used for the analysis.
https://doi.org/10.1371/journal.pone.0301599.s001
(XLSX)
S2 Table. 94 ADNI subject IDs used for the analysis.
https://doi.org/10.1371/journal.pone.0301599.s002
(XLSX)
S3 Table. Maximal observed Pearson correlation coefficients (HCP datasets).
https://doi.org/10.1371/journal.pone.0301599.s003
(DOCX)
S4 Table. Minimal observed Pearson correlation coefficients (HCP datasets).
https://doi.org/10.1371/journal.pone.0301599.s004
(DOCX)
S1 File. Python3 implementation of the correlation-based regression algorithm.
https://doi.org/10.1371/journal.pone.0301599.s005
(PY)
Acknowledgments
Data were provided [in part] by the Human Connectome Project, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University.
Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: https://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf
The authors thank Josine Verhaal for constructive criticism of the manuscript.
References
- 1. Munsell BC, Gleichgerrcht E, Hofesmann E, Delgaizo J, McDonald CR, Marebwa B, et al. Personalized connectome fingerprints: Their importance in cognition from childhood to adult years. Neuroimage. 2020 Nov 1;221:117122. Epub 2020 Jul 4. pmid:32634596.
- 2. Li Y, Liu Y, Li J, Qin W, Li K, Yu C, et al. Brain anatomical network and intelligence. PLoS Comput Biol. 2009 May;5(5):e1000395. Epub 2009 May 29. pmid:19492086; PMCID: PMC2683575.
- 3. Königs M, van Heurn LWE, Bakx R, Vermeulen RJ, Goslings JC, Poll-The BT, et al. The structural connectome of children with traumatic brain injury. Hum Brain Mapp. 2017 Jul;38(7):3603–3614. Epub 2017 Apr 21. pmid:28429381; PMCID: PMC6866988.
- 4. van den Heuvel MP, Stam CJ, Kahn RS, Hulshoff Pol HE. Efficiency of functional brain networks and intellectual performance. J Neurosci. 2009 Jun 10;29(23):7619–24. pmid:19515930; PMCID: PMC6665421.
- 5. Fan L, Su J, Qin J, Hu D, Shen H. A Deep Network Model on Dynamic Functional Connectivity With Applications to Gender Classification and Intelligence Prediction. Front Neurosci. 2020 Aug 18;14:881. pmid:33013292; PMCID: PMC7461846.
- 6. Sen B, Parhi KK. Predicting Biological Gender and Intelligence From fMRI via Dynamic Functional Connectivity. IEEE Trans Biomed Eng. 2021 Mar;68(3):815–825. Epub 2021 Feb 18. pmid:32746070.
- 7. Niu X, Zhang F, Kounios J, Liang H. Improved prediction of brain age using multimodal neuroimaging data. Hum Brain Mapp. 2020 Apr 15;41(6):1626–1643. Epub 2019 Dec 14. pmid:31837193; PMCID: PMC7267976.
- 8. Kardan O, Kaplan S, Wheelock MD, Feczko E, Day TKM, Miranda-Domínguez Ó, et al. Resting-state functional connectivity identifies individuals and predicts age in 8-to-26-month-olds. Dev Cogn Neurosci. 2022 Aug;56:101123. Epub 2022 Jun 15. pmid:35751994; PMCID: PMC9234342.
- 9. Ryman SG, Yeo RA, Witkiewitz K, Vakhtin AA, van den Heuvel M, de Reus M, et al. Fronto-Parietal gray matter and white matter efficiency differentially predict intelligence in males and females. Hum Brain Mapp. 2016 Nov;37(11):4006–4016. pmid:27329671; PMCID: PMC6867415.
- 10. Dubois J, Galdi P, Paul LK, Adolphs R. A distributed brain network predicts general intelligence from resting-state human neuroimaging data. Philos Trans R Soc Lond B Biol Sci. 2018 Sep 26;373(1756):20170284. pmid:30104429; PMCID: PMC6107566.
- 11. Jiang R, Calhoun VD, Fan L, Zuo N, Jung R, Qi S, et al. Gender Differences in Connectome-based Predictions of Individualized Intelligence Quotient and Sub-domain Scores. Cereb Cortex. 2020 Mar 14;30(3):888–900. pmid:31364696; PMCID: PMC7132922.
- 12. Lin L, Jin C, Fu Z, Zhang B, Bin G, Wu S. Predicting healthy older adult’s brain age based on structural connectivity networks using artificial neural networks. Comput Methods Programs Biomed. 2016 Mar;125:8–17. Epub 2015 Dec 8. pmid:26718834.
- 13. Taoudi-Benchekroun Y, Christiaens D, Grigorescu I, Gale-Grant O, Schuh A, Pietsch M, et al. Predicting age and clinical risk from the neonatal connectome. Neuroimage. 2022 Aug 15;257:119319. Epub 2022 May 16. pmid:35589001.
- 14. Cattell RB. The theory of fluid and crystallized general intelligence checked at the 5–6 year-old level. Br J Educ Psychol. 1967 Jun;37(2):209–24. pmid:6063107.
- 15. Mungas D, Heaton R, Tulsky D, Zelazo PD, Slotkin J, Blitz D, et al. Factor structure, convergent validity, and discriminant validity of the NIH Toolbox Cognitive Health Battery (NIHTB-CHB) in adults. J Int Neuropsychol Soc. 2014 Jul;20(6):579–87. Epub 2014 Jun 24. pmid:24960474; PMCID: PMC4103956.
- 16. Nisbett RE, Aronson J, Blair C, Dickens W, Flynn J, Halpern DF, et al. Intelligence: new findings and theoretical developments. Am Psychol. 2012 Feb-Mar;67(2):130–59. Epub 2012 Jan 2. Erratum in: Psychol Am. 2012 Feb;67(2):129. pmid:22233090.
- 17. Sternberg RJ. Intelligence. Dialogues Clin Neurosci. 2012 Mar;14(1):19–27. pmid:22577301; PMCID: PMC3341646.
- 18. Heaton RK, Akshoomoff N, Tulsky D, Mungas D, Weintraub S, Dikmen S, et al. Reliability and validity of composite scores from the NIH Toolbox Cognition Battery in adults. J Int Neuropsychol Soc. 2014 Jul;20(6):588–98. Epub 2014 Jun 24. pmid:24960398; PMCID: PMC4103963.
- 19. Shokri-Kojori E, Bennett IJ, Tomeldan ZA, Krawczyk DC, Rypma B. Estimates of brain age for gray matter and white matter in younger and older adults: Insights into human intelligence. Brain Res. 2021 Jul 15;1763:147431. Epub 2021 Mar 15. pmid:33737067; PMCID: PMC8428193.
- 20. Góngora D, Vega-Hernández M, Jahanshahi M, Valdés-Sosa PA, Bringas-Vega ML, et al. Crystallized and fluid intelligence are predicted by microstructure of specific white-matter tracts. Hum Brain Mapp. 2020 Mar;41(4):906–916. Epub 2019 Nov 5. pmid:32026600; PMCID: PMC7267934.
- 21.
Han CE, Peraza LR, Taylor J, & Kaiser M. Predicting age across human lifespan based on structural connectivity from diffusion tensor imaging. 2014. 2014 IEEE Biomedical Circuits and Systems Conference (BioCAS) Proceedings, Lausanne pp. 137–140. https://doi.org/10.1109/BioCAS.2014.6981664
- 22. Shen X, Finn ES, Scheinost D, Rosenberg MD, Chun MM, Papademetris X, et al. Using connectome-based predictive modeling to predict individual behavior from brain connectivity. Nat Protoc. 2017 Mar;12(3):506–518. Epub 2017 Feb 9. pmid:28182017; PMCID: PMC5526681.
- 23. Biomax. Biomax NICARATM integrates structural brain connectomics [White paper]. 2020. https://www.biomax.com/lib/products/nicara/Biomax_NICARA_WhitePaper.pdf
- 24. Kerepesi C, Szalkai B, Varga B, Grolmusz V. The braingraph.org database of high resolution structural connectomes and the brain graph tools. Cogn Neurodyn. 2017 Oct;11(5):483–486. Epub 2017 Jun 20. pmid:29067135; PMCID: PMC5637719.
- 25. Van Essen DC, Smith SM, Barch DM, Behrens TE, Yacoub E, Ugurbil K, et al. The WU-Minn Human Connectome Project: an overview. Neuroimage. 2013 Oct 15;80:62–79. Epub 2013 May 16. pmid:23684880; PMCID: PMC3724347.
- 26. Weiner MW, Veitch DP, Aisen PS, Beckett LA, Cairns NJ, Green RC, et al. The Alzheimer’s Disease Neuroimaging Initiative 3: Continued innovation for clinical trial improvement. Alzheimers Dement. 2017 May;13(5):561–571. Epub 2016 Dec 5. pmid:27931796; PMCID: PMC5536850.
- 27. Glasser MF, Coalson TS, Robinson EC, Hacker CD, Harwell J, Yacoub E, et al. A multi-modal parcellation of human cerebral cortex. Nature. 2016 Aug 11;536(7615):171–178. Epub 2016 Jul 20. pmid:27437579; PMCID: PMC4990127.
- 28. Makris N, Goldstein JM, Kennedy D, Hodge SM, Caviness VS, Faraone SV, et al. Decreased volume of left and total anterior insular lobule in schizophrenia. Schizophr Res. 2006 Apr;83(2–3):155–71. Epub 2006 Jan 31. pmid:16448806.
- 29. Frazier JA, Chiu S, Breeze JL, Makris N, Lange N, Kennedy DN, et al. Structural brain magnetic resonance imaging of limbic and thalamic volumes in pediatric bipolar disorder. Am J Psychiatry. 2005 Jul;162(7):1256–65. pmid:15994707.
- 30. Goldstein JM, Seidman LJ, Makris N, Ahern T, O’Brien LM, Caviness VS Jr, et al. Hypothalamic abnormalities in schizophrenia: sex effects and genetic vulnerability. Biol Psychiatry. 2007 Apr 15;61(8):935–45. Epub 2006 Oct 13. pmid:17046727.
- 31. Gaser C, Dahnke R, Thompson PM, Kurth F, Luders E, Alzheimer’s Disease Neuroimaging Initiative. CAT–A Computational Anatomy Toolbox for the Analysis of Structural MRI Data. Preprint. 2022. https://doi.org/10.1101/2022.06.11.495736
- 32. Fischl B. FreeSurfer. Neuroimage. 2012 Aug 15;62(2):774–81. Epub 2012 Jan 10. pmid:22248573; PMCID: PMC3685476.
- 33. Tournier JD, Smith R, Raffelt D, Tabbara R, Dhollander T, Pietsch M, et al. MRtrix3: A fast, flexible and open software framework for medical image processing and visualisation. Neuroimage. 2019 Nov 15;202:116137. Epub 2019 Aug 29. pmid:31473352.
- 34. Jenkinson M, Beckmann CF, Behrens TE, Woolrich MW, Smith SM. FSL. Neuroimage. 2012 Aug 15;62(2):782–90. Epub 2011 Sep 16. pmid:21979382.
- 35. Avants BB, Tustison NJ, Song G, Cook PA, Klein A, Gee JC. A reproducible evaluation of ANTs similarity metric performance in brain image registration. Neuroimage. 2011 Feb 1;54(3):2033–44. Epub 2010 Sep 17. pmid:20851191; PMCID: PMC3065962.
- 36. Yu J, Mathi Kanchi M, Rawtaer I, Feng L, Kumar AP, Kua EH, et al. Differences between multimodal brain-age and chronological-age are linked to telomere shortening. Neurobiol Aging. 2022 Jul;115:60–69. Epub 2022 Mar 31. pmid:35472831; PMCID: PMC9133148.
- 37. More S, Antonopoulos G, Hoffstaedter F, Caspers J, Eickhoff SB, Patil KR, et al. Brain-age prediction: A systematic comparison of machine learning workflows. Neuroimage. 2023 Apr 15;270:119947. Epub 2023 Feb 16. pmid:36801372.
- 38. Wu Y, Besson P, Azcona EA, Bandt SK, Parrish TB, Breiter HC, et al. A multicohort geometric deep learning study of age dependent cortical and subcortical morphologic interactions for fluid intelligence prediction. Sci Rep. 2022 Oct 22;12(1):17760. pmid:36273036; PMCID: PMC9588039.
- 39. Krämer C, Stumme J, da Costa Campos L, Dellani P, Rubbert C, Caspers J, et al. Prediction of cognitive performance differences in older age from multimodal neuroimaging data. Geroscience. 2023 Jun 13. Epub ahead of print. pmid:37308769.
- 40. Xu H, Xu C, Yang Z, Bai G, Yin B. Two sides of the same coin: distinct neuroanatomical patterns predict crystallized and fluid intelligence in adults. Front Neurosci. 2023 May 25;17:1199106. pmid:37304014; PMCID: PMC10249781.
- 41. Dhamala E, Jamison KW, Jaywant A, Dennis S, Kuceyeski A. Distinct functional and structural connections predict crystallised and fluid cognition in healthy adults. Hum Brain Mapp. 2021 Jul;42(10):3102–3118. Epub 2021 Apr 8. pmid:33830577; PMCID: PMC8193532.
- 42. Maggioni E, Squarcina L, Dusi N, Diwadkar VA, Brambilla P. Twin MRI studies on genetic and environmental determinants of brain morphology and function in the early lifespan. Neurosci Biobehav Rev. 2020 Feb;109:139–149. Epub 2020 Jan 3. pmid:31911159.
- 43. Gong G, He Y, Evans AC. Brain connectivity: gender makes a difference. Neuroscientist. 2011 Oct;17(5):575–91. Epub 2011 Apr 28. pmid:21527724.
- 44. Satterthwaite TD, Davatzikos C. Towards an Individualized Delineation of Functional Neuroanatomy. Neuron. 2015 Aug 5;87(3):471–3. pmid:26247857; PMCID: PMC5642037.
- 45. Jacobs EG, Weiss BK, Makris N, Whitfield-Gabrieli S, Buka SL, Klibanski A, et al. Impact of Sex and Menopausal Status on Episodic Memory Circuitry in Early Midlife. J Neurosci. 2016 Sep 28;36(39):10163–73. Epub 2016 Sep 28. pmid:27683911; PMCID: PMC5039260.
- 46. Gur RC, Gur RE. Complementarity of sex differences in brain and behavior: From laterality to multimodal neuroimaging. J Neurosci Res. 2017 Jan 2;95(1–2):189–199. pmid:27870413; PMCID: PMC5129843.
- 47. Jacobs EG, Goldstein JM. The Middle-Aged Brain: Biological sex and sex hormones shape memory circuitry. Curr Opin Behav Sci. 2018 Oct;23:84–91. Epub 2018 May 7. pmid:30271832; PMCID: PMC6157917.
- 48. Weis S, Patil KR, Hoffstaedter F, Nostro A, Yeo BTT, Eickhoff SB. Sex Classification by Resting State Brain Connectivity. Cereb Cortex. 2020 Mar 21;30(2):824–835. pmid:31251328; PMCID: PMC7444737.
- 49. Lord A, Ehrlich S, Borchardt V, Geisler D, Seidel M, Huber S, et al. Brain parcellation choice affects disease-related topology differences increasingly from global to local network levels. Psychiatry Res Neuroimaging. 2016 Mar 30;249:12–9. Epub 2016 Feb 12. pmid:27000302.
- 50. Messé A. Parcellation influence on the connectivity-based structure-function relationship in the human brain. Hum Brain Mapp. 2020 Apr 1;41(5):1167–1180. Epub 2019 Nov 19. pmid:31746083; PMCID: PMC7267927.
- 51. Wu Z, Xu D, Potter T, Zhang Y; Alzheimer’s Disease Neuroimaging Initiative. Effects of Brain Parcellation on the Characterization of Topological Deterioration in Alzheimer’s Disease. Front Aging Neurosci. 2019 May 21;11:113. pmid:31164815; PMCID: PMC6536693.
- 52. Goriounova NA, Heyer DB, Wilbers R, Verhoog MB, Giugliano M, Verbist C, et al. Large and fast human pyramidal neurons associate with intelligence. Elife. 2018 Dec 18;7:e41714. pmid:30561325; PMCID: PMC6363383.
- 53. Bookheimer SY, Salat DH, Terpstra M, Ances BM, Barch DM, Buckner RL, et al. The Lifespan Human Connectome Project in Aging: An overview. Neuroimage. 2019 Jan 15;185:335–348. Epub 2018 Oct 15. pmid:30332613; PMCID: PMC6649668.
- 54. Zimmermann J, Griffiths JD, McIntosh AR. Unique Mapping of Structural and Functional Connectivity on Cognition. J Neurosci. 2018 Nov 7;38(45):9658–9667. Epub 2018 Sep 24. pmid:30249801; PMCID: PMC6595988.
- 55. Yeo BT, Krienen FM, Sepulcre J, Sabuncu MR, Lashkari D, Hollinshead M, et al. The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J Neurophysiol. 2011 Sep;106(3):1125–65. Epub 2011 Jun 8. pmid:21653723; PMCID: PMC3174820.
- 56. Schoonheim MM, Broeders TAA, Geurts JJG. The network collapse in multiple sclerosis: An overview of novel concepts to address disease dynamics. Neuroimage Clin. 2022;35:103108. Epub 2022 Jul 14. pmid:35917719; PMCID: PMC9421449.
- 57. Schoonheim MM, Meijer KA, Geurts JJ. Network collapse and cognitive impairment in multiple sclerosis. Front Neurol. 2015 Apr 14;6:82. pmid:25926813; PMCID: PMC4396388.
- 58. Rubinov M, Sporns O. Complex network measures of brain connectivity: uses and interpretations. Neuroimage. 2010 Sep;52(3):1059–69. Epub 2009 Oct 9. pmid:19819337.
- 59. Nebli A, Rekik I. Adversarial brain multiplex prediction from a single brain network with application to gender fingerprinting. Med Image Anal. 2021 Jan;67:101843. Epub 2020 Oct 13. pmid:33129149.