Explainable Deep Learning of Resting-State Functional Connectomes Reveals Network Biomarkers of Adolescent Intelligence

Md. Tanvir Rahman; Nabil Anan Orka; Asaduzzaman Khan; Mohammad Ali Moni

q-bio.NCSep 27, 2026

Explainable Deep Learning of Resting-State Functional Connectomes Reveals Network Biomarkers of Adolescent Intelligence

Authors: Md. Tanvir Rahman, Nabil Anan Orka, Asaduzzaman Khan, Mohammad Ali Moni

Organizations: 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

Abstract

Mapping resting-state brain organization to individual differences in cognitive ability remains a major challenge in population neuroinformatics. Although deep learning enables flexible modeling of brain connectivity, limited interpretability restricts its scientific and clinical utility. To address this objective, we developed an explainable deep learning framework based on sparse projected residual networks to predict fluid, crystallized, and total intelligence from resting-state functional magnetic resonance imaging in 5,285 participants from the Adolescent Brain Cognitive Development study. We incorporated three complementary explainability methods (Integrated Gradients, Gradient Shapley Additive Explanations, and Occlusion) to interpret model behavior. The framework outperformed existing approaches, achieving Pearson correlations of 0.44, 0.58, and 0.56 for fluid, crystallized, and total intelligence, respectively, corresponding to predictive improvements of 6 to 9 percent. All three explainability methods produced near-identical feature rankings (pairwise rank correlations greater than 0.99). Consensus maps revealed a dual-layered functional architecture where primary predictive hubs localized within canonical systems, while the strongest global predictive pathways frequently bypassed these hubs through distributed, long-range relay connections. These findings suggest that intelligence emerges from the interaction between localized computational hubs and distributed communication pathways. Ultimately, these normative network architectures provide clinical reference maps to detect individual deviations, supporting earlier diagnosis, cognitive subtype stratification, and treatment monitoring in atypical neurodevelopment.

Figures & tables

Fold	Split	Subjects	Males	Females	Age (months)	$G_{f}$	$G_{c}$	$G_{t}$
1	Train	3758	1821	1936	119.86 $\pm$ 7.50	93.02 $\pm$ 10.09	87.23 $\pm$ 6.67	87.60 $\pm$ 8.52
	Validation	470	239	231	119.77 $\pm$ 7.46	93.69 $\pm$ 9.64	87.49 $\pm$ 6.64	88.16 $\pm$ 8.20
	Test	1057	500	557	119.66 $\pm$ 7.66	93.01 $\pm$ 10.06	87.12 $\pm$ 6.56	87.54 $\pm$ 8.44
2	Train	3765	1814	1950	119.73 $\pm$ 7.53	92.96 $\pm$ 10.13	87.18 $\pm$ 6.66	87.53 $\pm$ 8.56
	Validation	463	250	213	120.06 $\pm$ 7.48	92.67 $\pm$ 10.18	87.68 $\pm$ 6.86	87.65 $\pm$ 8.58
	Test	1057	496	561	119.98 $\pm$ 7.53	93.65 $\pm$ 9.67	87.21 $\pm$ 6.48	88.00 $\pm$ 8.15

TABLE I : Demographic and Cognitive Score Summary Across Five Family-Aware Cross-Validation Folds. Mean

\pm

SD are reported.

Threshold	$G_{f}$	$G_{c}$	$G_{t}$
30%	5015.2 $\pm$ 54.3	5281.0 $\pm$ 72.8	5200.6 $\pm$ 43.1
40%	3230.0 $\pm$ 54.9	3564.4 $\pm$ 49.3	3490.2 $\pm$ 42.6
50%	2127.0 $\pm$ 20.9	2394.8 $\pm$ 10.3	2361.4 $\pm$ 53.7
60%	1366.6 $\pm$ 15.6	1610.6 $\pm$ 20.0	1585.6 $\pm$ 53.3
70%	855.0 $\pm$ 11.5	1023.2 $\pm$ 26.6	1013.8 $\pm$ 34.6
80%	468.2 $\pm$ 19.9	599.0 $\pm$ 21.6	582.2 $\pm$ 25.5

TABLE II : Average Number of Stable Features Retained Across 5 Folds by Consensus Threshold. Mean

\pm

SD are reported.

Fig. 1 : Overview of the proposed methodology. Resting-state networks are constructed from 352 ROIs (333 cortical, 19 subcortical), yielding 61,776 static Functional Connectivity edges and 352 temporal variance features. Following family-aware stratification, bootstrapped Ridge consensus masking (50% threshold) reduces the feature space to approximately 2,100–2,400 stable features. The SPRN predicts cognitive scores through learned residual projections across descending latent dimensions (192, 96, 48, 24) with Gaussian noise augmentation, multi-level dropout ( $p=0.758$ ), Batch Normalization, and GELU activations. Covariates are incorporated via late fusion to prevent confounding of neural representations.

Target	SPRN Configuration	$R^{2}\uparrow$	r $\bm{\uparrow}$	MAE $\bm{\downarrow}$	MSE $\bm{\downarrow}$
$G_{f}$	SPRN (Proposed)	0.1906 $\pm$ 0.0175	0.4421 $\pm$ 0.0198	7.1437 $\pm$ 0.0850	81.5946 $\pm$ 1.7941
	– without Variance Data	0.1899 $\pm$ 0.0233	0.4412 $\pm$ 0.0269	7.1494 $\pm$ 0.0950	81.6464 $\pm$ 1.7613
	– without Residual	0.1539 $\pm$ 0.0487	0.4041 $\pm$ 0.0613	7.2739 $\pm$ 0.1937	85.3054 $\pm$ 5.0049
	– Covariates Only	0.1233 $\pm$ 0.0222	0.3560 $\pm$ 0.0310	7.4076 $\pm$ 0.0450	88.3668 $\pm$ 1.8614
$G_{c}$	SPRN (Proposed)	0.3361 $\pm$ 0.0250	0.5833 $\pm$ 0.0195	4.2274 $\pm$ 0.1878	29.2638 $\pm$ 2.8185
	– without Variance Data	0.3358 $\pm$ 0.0237	0.5818 $\pm$ 0.0203	4.2278 $\pm$ 0.1955	29.2777 $\pm$ 2.7990

TABLE III : Ablation Study on Architectural Components of Sparse Projected Residual Networks. Mean

\pm

SD across 5 Folds is reported. Bold indicates best per target.

Target	Bootstrap Selection Threshold	$R^{2}\uparrow$	r $\bm{\uparrow}$	MAE $\bm{\downarrow}$	MSE $\bm{\downarrow}$
$G_{f}$	30% Consensus	0.1822 $\pm$ 0.0178	0.4342 $\pm$ 0.0163	7.1658 $\pm$ 0.0886	82.4343 $\pm$ 1.4780
	40% Consensus	0.1915 $\pm$ 0.0220	0.4416 $\pm$ 0.0234	7.1276 $\pm$ 0.0468	81.4923 $\pm$ 1.5167
	50% Consensus (Proposed)	0.1906 $\pm$ 0.0175	0.4421 $\pm$ 0.0198	7.1437 $\pm$ 0.0850	81.5946 $\pm$ 1.7941
	60% Consensus	0.1948 $\pm$ 0.0215	0.4480 $\pm$ 0.0208	7.1310 $\pm$ 0.0848	81.1674 $\pm$ 1.9071
	70% Consensus	0.1912 $\pm$ 0.0246	0.4423 $\pm$ 0.0290	7.1330 $\pm$ 0.1068	81.5256 $\pm$ 2.2796
	80% Consensus	0.1879 $\pm$ 0.0253	0.4379 $\pm$ 0.0293	7.1365 $\pm$ 0.0905	81.8505 $\pm$ 1.8591

TABLE IV : Sensitivity Analysis of Bootstrapped Feature Selection Thresholds on SPRN Performance. Mean

\pm

SD across 5 Folds is reported.

Target	Method	$R^{2}\uparrow$	r $\bm{\uparrow}$	MAE $\bm{\downarrow}$	MSE $\bm{\downarrow}$
$G_{f}$	Ridge Regression	0.1028 $\pm$ 0.0279	0.3684 $\pm$ 0.0243	7.5259 $\pm$ 0.1298	90.4258 $\pm$ 2.1422
	Lasso Regression	-0.0087 $\pm$ 0.0055	0.0000 $\pm$ 0.0000	7.9626 $\pm$ 0.1329	101.7041 $\pm$ 1.8196
	ElasticNet	-0.1361 $\pm$ 0.0386	0.3061 $\pm$ 0.0255	8.4998 $\pm$ 0.1611	114.5058 $\pm$ 3.1759
	Random Forest	0.1256 $\pm$ 0.0136	0.3688 $\pm$ 0.0187	7.4148 $\pm$ 0.0821	88.1493 $\pm$ 1.6804
	Gradient Boosting	0.1393 $\pm$ 0.0098	0.3811 $\pm$ 0.0105	7.3559 $\pm$ 0.0760	86.7708 $\pm$ 1.0400
	Standard GCN	0.1289 $\pm$ 0.0192	0.3625 $\pm$ 0.0258	7.3879 $\pm$ 0.0597	87.8054 $\pm$ 1.6192

TABLE V : Comparison of Predictive Performance Across Methodological Families. Mean

\pm

SD across 5 Folds is reported.

Study	Model Architecture	Modality	$G_{f}(r)$	$G_{c}(r)$	$G_{t}(r)$
Huang et al. (2022) [ 17 ]	ST-DAG-Att	rs-fMRI	0.288	-	-
Li et al. (2023) [ 22 ]	Bi-LSTM (Dynamic FC)	rs-fMRI (Rest Only)	0.408	0.536	0.529
Xia et al. (2023) [ 37 ]	ML-Joint-Att	rs-fMRI	0.298	0.391	-
Thapaliya et al. (2025) [ 34 ]	BrainRGIN	rs-fMRI	0.230	0.300	0.310
This Study	SPRN (Proposed)	rs-fMRI	0.442	0.583	0.562

TABLE VI : State-of-the-Art Comparison of Intelligence Prediction using rs-fMRI on the ABCD Dataset.

Fig. 2 : XAI Consensus visualization of neuroanatomical drivers for intelligence prediction. Left panels (a, c, e): Top 10 hubs (nodes) and their strongest inter-connections. Right panels (b, d, f): Top 10 global edges across the whole brain. Color intensity reflects attribution magnitude. Panels correspond to $G_{f}$ (a, b), $G_{c}$ (c, d), and $G_{t}$ (e, f).

Cognitive Target	Explainer Pair	Spearman $\bm{\rho}$	Kendall $\bm{\tau}$	Pearson r	Jaccard (Top 100)
$G_{f}$	IG vs. GradientSHAP	0.9998	0.9874	0.9999	0.9802
	IG vs. Occlusion	0.9995	0.9857	0.9997	0.9802
	GradientSHAP vs. Occlusion	0.9993	0.9809	0.9996	0.9608
$G_{c}$	IG vs. GradientSHAP	0.9998	0.9883	0.9999	0.9802
	IG vs. Occlusion	0.9992	0.9829	0.9996	0.9802
	GradientSHAP vs. Occlusion	0.9990	0.9793	0.9995	0.9802

TABLE VII : Multi-Algorithm Convergence Analysis of XAI Feature Attributions.

Explore similar work

RegionFM: Interpretable Region-Based Brain MRI Classification Using Foundation Model Embeddings
Jul 15, 2026Wei Zhang, Ming TangFunctional Magnetic Resonance Imaging Region
L3S Research Center, Hannover, Germany
Temporal Graph Learning of Wearable Actigraphy and Sleep Traces for Modelling Adolescent Crystallized Intelligence
Sep 27, 2026Md. Tanvir Rahman, Nabil Anan Orka, Asaduzzaman Khan +1Sleep General Intelligence
School of Health and Rehabilitation Sciences, The University of Queensland, QLD 4072, Australia
XAI-Refine: An Automated Explanation-Knowledge Loop for Brain-Age Prediction
Sep 8, 2026Yang Qiao, Junjie Wu, Deqiang Qiu +2Brain Age Explainable AI Methods
Department of Computer Science Emory University Atlanta, Georgia, USA · Department of Radiology and Imaging Sciences Computational NeuroImaging & Neuroscience Lab, Emory University Atlanta, Georgia, USA · Department of Neurology Emory University School of Medicine Atlanta, Georgia, USA

Jul 15, 2026, cs.CVJ/K move · Enter open · S save

RegionFM: Interpretable Region-Based Brain MRI Classification Using Foundation Model Embeddings

Wei Zhang, Ming Tang

L3S Research Center, Hannover, Germany

Functional Magnetic Resonance Imaging Region Logistic Regression

Open paper AskPDF arXiv

Foundation models provide powerful representations for brain MRI analysis, but their predictions remain difficult to interpret in anatomically meaningful terms. Clinical assessment of brain MRI is commonly organized around anatomically defined structures and regional abnormalities, whereas conventional explanation methods typically produce voxel- or patch-level importance maps that do not explicitly quantify the contributions of individual brain regions. To address this mismatch, we propose RegionFM, an interpretable framework that integrates anatomical segmentation with brain MRI foundation-model embeddings. RegionFM first divides each MRI scan into anatomical regions and constructs a separate MRI volume for each region. A frozen foundation model then encodes each region into an embedding, and a region-additive logistic model combines these embeddings such that every anatomical region contributes an explicit scalar term to the final prediction. This formulation supports both subject-level and cohort-level analyses of regional contributions. We evaluate RegionFM on cognitive-impairment classification using embeddings from multiple pretrained brain MRI foundation models. The results show that RegionFM maintains performance comparable to less interpretable fine-tuning approaches while providing anatomically grounded explanations. Randomized embedding ablations yield near-chance performance, indicating that the predictions rely on meaningful structure captured by the foundation-model embeddings rather than simple feature statistics. Overall, RegionFM better aligns model explanations with anatomy-based clinical reasoning while maintaining competitive predictive performance.

Abstract

Explore similar work

RegionFM: Interpretable Region-Based Brain MRI Classification Using Foundation Model Embeddings

Temporal Graph Learning of Wearable Actigraphy and Sleep Traces for Modelling Adolescent Crystallized Intelligence

XAI-Refine: An Automated Explanation-Knowledge Loop for Brain-Age Prediction