Predicting Cognitive Profiles from Structural MRI using Spatial Graph Representation Learning
Authors: Taslima Ferdaus Shuva, Md. Tanvir Rahman, Risala Tasin Khan, Mohammad Ali Moni, ASM Ashraf Mahmud
Organizations: Institute of Information Technology, Jahangirnagar University, Savar, Dhaka, 1342, Bangladesh · Department of Computer Science and Engineering, Daffodil International University, Daffodil Smart City (DSC), Birulia, Savar, Dhaka, 1216, Bangladesh · School of Health and Rehabilitation Sciences, The University of Queensland, QLD, 4072, Australia · Department of Information and Communication Technology, Mawlana Bhashani Science and Technology University, Tangail, 1902, Bangladesh · AI and Cyber Futures Institute, Charles Sturt University, Bathurst, 2795, New South Wales, Australia · Rural Health Research Institute, Charles Sturt University, Orange, 2800, New South Wales, Australia · Faculty of Computing Science and Engineering, University of the West of Scotland, High Street, Paisley, PA1 2BE, United Kingdom
Predicting cognitive abilities from structural neuroimaging remains challenging due to the high dimensionality and geometric complexity of brain morphology. Conventional models typically flatten regional morphometry into tabular features, failing to explicitly represent spatial relationships. We propose a Spatial-Topological Graph Neural Network (ST-GNN) to predict four distinct cognitive phenotypes across large pediatric cohorts (N = 2,378 to 2,424). By projecting regional morphometry onto a population-level standardized spatial graph, ST-GNN restricts message passing to defined neighboring regions. Evaluations demonstrate that ST-GNN effectively fuses spatial morphology with systemic covariates, outperforming linear baselines and topology-free deep models. Ablations reveal that imposing a sparse geometric topology (K-nearest neighbors) and utilizing global feature scaling - which preserves absolute morphometric scale - outperform fully connected networks and region-normalized approaches. These choices provide transferable architectural inductive biases, revealing differences between computational routing and morphometric feature attribution. Integrated Gradients analysis identified distinct morphological features consistent with neuroanatomical evidence, such as cerebellar volume for working memory and superior frontal surface area for executive function. These findings suggest standardized structural MRI provides useful information for task-specific cognitive prediction, establishing a foundation for MRI-based neurodevelopmental assessment.
Figures & tables
Characteristic
Cognitive Targets
Card Sort
Flanker
List Sort
Processing Speed
Sample Information
Subjects ( n )
2,424
2,422
2,378
2,398
Families ( n )
2,424
2,422
2,378
2,398
Demographics & Anthropometrics
Age (years)
10.5 ± 3.5
10.5 ± 3.5
10.6 ± 3.4
10.5 ± 3.5
Table 1: Participant characteristics for the target-specific cohorts included in the family-aware cross-validation analysis
Figure 1: Architecture of the proposed Spatial-Topological Graph Neural Network (ST-GNN) for cognitive score prediction from structural MRI and systemic covariates. The left panel illustrates the spatial graph construction, in which the 101 anatomical regions defined by the DKT and ASEG atlases are embedded in a spatial K-nearest-neighbor graph with K=3 . Morphological node features are combined with learnable node-identity embeddings to form the initial node representation. The middle panel depicts two graph-attention blocks with residual connections and regularization, followed by learned global attention pooling to obtain a whole-brain structural representation. The right panel shows the late-fusion stage, where demographic, socioeconomic, and acquisition-related covariates are processed through a parallel multilayer perceptron and combined with the graph representation to predict the continuous cognitive score.
Figure 2: Orthographic projections of the population-level spatial graph embedded in standardized MNI space. The graph comprises 101 neuroanatomical nodes corresponding to regions from the DKT and ASEG atlases, color-coded by macro-anatomical division. Undirected edges represent K-nearest spatial neighbors ( K=3 ) computed using Euclidean distance between regional centers of mass. The projections illustrate the spatial topology across the axial, coronal, and sagittal views, while the adjacency matrix highlights the sparse connectivity used to constrain ST-GNN message passing.
Data Partition
Cognitive Target Distributions [N (Mean ± SD)]
Card Sort
Flanker
List Sort
Processing Speed
Fold 1
Train
1648 (87.26 ± 19.18)
1647 (86.25 ± 15.53)
1616 (92.64 ± 17.47)
1630 (88.67 ± 19.78)
Val
291 (87.17 ± 20.29)
291 (85.69 ± 16.08)
286 (91.22 ± 19.28)
288 (87.74 ± 20.91)
Test
485 (86.48 ± 20.05)
484 (85.82 ± 16.19)
476 (92.36 ± 18.18)
480 (88.81 ± 20.20)
Fold 2
Train
1648 (87.17 ± 19.43)
1646 (86.08 ± 15.48)
1616 (92.33 ± 17.78)
1630 (88.51 ± 19.76)
Val
291 (86.05 ± 19.76)
291 (85.41 ± 15.79)
286 (92.80 ± 18.31)
288 (89.93 ± 20.64)
Table 2: Family-aware stratified cross-validation fold statistics across cognitive targets
Region Name
Left (MNI)
Right (MNI)
Region Name
Left (MNI)
Right (MNI)
X
Y
Z
X
Y
Z
X
Y
Z
X
Y
Z
Cortical Regions (DKT Atlas) [62 regions total: 31 Left, 31 Right]
Caudal Anterior Cingulate
-5.20
20.79
26.02
5.73
21.07
26.96
Pars Opercularis
-46.43
15.00
12.38
47.44
14.52
12.65
Caudal Middle Frontal
-35.91
11.64
45.33
35.98
12.41
46.12
Pars Orbitalis
-41.43
40.46
-12.12
42.39
40.94
-11.38
Cuneus
-7.06
-79.77
18.11
7.73
-79.12
19.40
Pars Triangularis
-44.56
30.95
2.54
46.83
29.95
4.32
Entorhinal
-23.84
-6.58
-34.46
24.71
-5.89
-33.84
Pericalcarine
-12.05
-80.75
5.83
11.98
-79.56
6.63
Table 3: Neuroanatomical regions and associated MNI spatial coordinates used for template graph construction
Cognitive Target
Linear Model
R2
r
RMSE
MAE
Card Sort
Ridge
0.3621 ± 0.0159
0.6041 ± 0.0132
15.55 ± 0.48
11.71 ± 0.36
Lasso
0.3625 ± 0.0161
0.6041 ± 0.0133
15.55 ± 0.48
11.71 ± 0.38
ElasticNet
0.3616 ± 0.0159
0.6041 ± 0.0134
15.56 ± 0.49
11.70 ± 0.38
Flanker
Ridge
0.3081 ± 0.0215
0.5594 ± 0.0225
13.06 ± 0.67
9.84 ± 0.49
Lasso
0.3078 ± 0.0210
0.5595 ± 0.0227
13.06 ± 0.67
9.83 ± 0.49
ElasticNet
0.3081 ± 0.0215
0.5594 ± 0.0225
13.06 ± 0.67
9.84 ± 0.49
Table 4: Baseline predictive performance using demographic and socioeconomic covariates only
Cognitive Target
ST-GNN Configuration
R2
r
RMSE
MAE
Card Sort
Graph Connectivity
– Spatial ( K=3 )
0.4433 ± 0.0183
0.6677 ± 0.0151
14.5239 ± 0.3818
11.0068 ± 0.3326
– Spatial ( K=5 )
0.4419 ± 0.0227
0.6676 ± 0.0173
14.5399 ± 0.3273
11.0116 ± 0.2484
– Spatial ( K=10 )
0.4443 ± 0.0206
0.6674 ± 0.0160
14.5106 ± 0.3928
11.0278 ± 0.2701
– Spatial ( K=15 )
0.4426 ± 0.0226
0.6661 ± 0.0175
14.5328 ± 0.4266
11.0425 ± 0.3077
– Fully Connected
0.4356 ± 0.0179
0.6607 ± 0.0135
14.6256 ± 0.4224
11.1633 ± 0.3507
Table 5: Ablation analysis of ST-GNN topology, feature scaling, and covariate fusion
Cognitive Target
Model
R2
r
RMSE
MAE
Card Sort
Ridge
0.3719 ± 0.0205
0.6117 ± 0.0179
15.4300 ± 0.4900
11.7000 ± 0.3300
Lasso
0.3720 ± 0.0178
0.6119 ± 0.0153
15.4300 ± 0.4700
11.6700 ± 0.3200
ElasticNet
0.3726 ± 0.0183
0.6128 ± 0.0167
15.4200 ± 0.4800
11.6700 ± 0.3400
MLP
0.2774 ± 0.0294
0.5483 ± 0.0309
16.5502 ± 0.5904
12.4426 ± 0.6814
Flanker
Ridge
0.3212 ± 0.0200
0.5708 ± 0.0181
12.9300 ± 0.5600
9.7400 ± 0.4200
Lasso
0.3120 ± 0.0340
0.5660 ± 0.0249
13.0100 ± 0.4900
9.8100 ± 0.3600
Table 6: Predictive performance of baseline models using unstructured (flattened) structural morphometry and covariates
Figure 3: Integrated Gradients attribution analysis across four cognitive targets: (a) dimensional change card sort, (b) flanker inhibitory control and attention, (c) list sorting, and (d) processing speed. The left panels show the five morphological brain regions with the highest mean absolute attribution for each target, with attribution decomposed into volume, surface area, and cortical thickness components. The right panels compare the aggregate attribution of structural brain morphology with that of demographic, socioeconomic, and acquisition-related covariates.
Figure 4: Relationship between attention-based routing and model attribution across the four cognitive targets. The scatter plots show spatial attention-based routing scores from the first graph-attention layer against absolute Integrated Gradients attribution. The plots illustrate that nodes with relatively high routing scores do not necessarily receive high predictive attribution. The high-routing, low-attribution nodes include centrally positioned fluid-space structures, whereas the high-attribution features include morphological regions identified by the model for each cognitive target.
Department of Cognitive Sciences, University of California, Irvine · Department of Neurobiology & Behavior, University of California, Irvine · Department of Psychological & Brain Sciences, University of California, Santa Barbara
Department of Computer Science, Georgia State University · Tri-Institutional Center for Translational Research in Neuroimaging and Data Science (TReNDS), Georgia State University, Georgia Institute of Technology, and Emory University · Neuroscience Institute, Georgia State University
School of Health and Rehabilitation Sciences, The University of Queensland, QLD 4072, Australia · Department of Information and Communication Technology, Mawlana Bhashani Science and Technology University, Tangail 1902, Bangladesh