PaperlayerView exact revision on arXiv

Source-linked AI summary

MAC-Net: A Multi-Task Deep Learning Framework for Modeling Cognitive Function From Task-Based fMRI

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

arXiv:2609.33440v1cs.AI

TL;DR

Individual cognitive function is difficult to estimate from task-based fMRI amid demographic and scanner-related variation, while complex models do not necessarily improve prediction. MAC-Net separates tfMRI feature learning from covariates and benchmarks the resulting framework. It outperformed conventional and advanced deep-learning baselines in the ABCD cohort, with N-back and MID forming a high-performing multi-task configuration.

  • Problem

    Demographic and scanner-related factors can inflate tfMRI prediction, and complex architectures require direct benchmarking because greater complexity does not ensure better performance.

    Source: §I · p. 2

  • Method

    MAC-Net separates task-fMRI representation learning from participant and acquisition variables, combining them through late fusion and evaluating the model against conventional and deep-learning baselines.

    Source: §IV · p. 9

  • Results

    MAC-Net outperformed conventional and advanced deep-learning baselines; N-back was the strongest single-task representation, and integrating MID produced a high-performing multi-task configuration.

    Source: §IV · p. 9

  • Takeaways & Limitations

    The framework provides a methodological foundation for future neural-engineering applications, with multi-site validation and longitudinal testing identified as next steps.

    Source: §IV · p. 9

    The analysis used cross-sectional baseline data from one developmental cohort, and performance differences are descriptive because task selection used the same cohort without formal paired testing.

    Source: E. Limitations and Translational… · p. 8

Abstract

from arXiv · show

Objective cognitive assessment from neural signals supports neurorehabilitation, but individual-level prediction from task-based fMRI (tfMRI) remains difficult because neural features coexist with substantial demographic and scanner-related variation. We present the Multi-task Activation and Contrast Network (MAC-Net), a covariate-aware deep learning framework for modeling individual cognitive function from regional tfMRI. By isolating tfMRI features into a dedicated neural pathway and restricting participant variables to a terminal late-fusion pathway, MAC-Net prevents dominant covariates from suppressing high-dimensional clinical representations during feature learning. Evaluating baseline data from 6,500 Adolescent Brain Cognitive Development Study participants under family-aware cross-validation, MAC-Net was benchmarked against linear models, random forests, and alternative deep architectures. The N-back plus Monetary Incentive Delay configuration achieved $R^{2}$ values of 0.174, 0.238, and 0.277 for fluid, crystallized, and total cognition, outperforming covariate-only baselines (0.178) and alternative deep models (0.217). N-back was the most informative paradigm, whereas incorporating the Stop Signal Task marginally degraded performance. Feature attributions via Integrated Gradients, DeepLIFT, and Input Gradient were highly concordant, localizing working-memory-related frontal, parietal, and cingulate regions. These findings demonstrate that covariate-aware multi-task modeling yields reproducible cognitive-function estimations, establishing a robust neural engineering framework for clinical translation.

I. INTRODUCTION

Task-based fMRI offers experimentally defined neural contexts for examining cognitive differences, but demographic and scanner factors can confound prediction. MAC-Net separates neural features from participant variables and evaluates their predictive value alongside benchmarking and attribution stability.

  • tfMRI engages defined cognitive systems, and combining task paradigms has been reported to improve predictive reliability over a single task state.These findings motivate integrating heterogeneous task-evoked representations while preserving information relevant to individual differences.Source: §I · p. 1
  • Demographic, anatomical, scanner, and motion factors can inflate prediction, motivating leakage controls and tests of predictive value beyond readily available covariates.The introduction also notes that feature-selection choices can yield divergent neurobiological interpretations.Source: §I · p. 2
  • Deep learning may integrate high-dimensional tfMRI features, but complexity does not guarantee better prediction, so architectures require direct benchmarking.Regularized linear models remain competitive, and neurocognitive deep-learning studies face inconsistent evaluation and limited explainability.Source: §I · p. 2
  • MAC-Net models regional tfMRI and participant variables separately, fusing covariates only at the final stage in an evaluation of prediction, task integration, and attribution stability.The study uses 6,500 ABCD participants and compares Integrated Gradients, DeepLIFT, and Input×Gradient attributions.Source: §I · p. 2
  • The work establishes a modeling framework rather than a clinical diagnostic system, with future investigation proposed in neurological and neurorehabilitation populations.Objective individualized characterization may complement behavioral and clinical assessments.Source: §I · pp. 2–3

II. METHODS

The analysis used 6,500 complete-case ABCD participants and family-aware five-fold cross-validation, grouping siblings and twins within the same fold.

  • The complete-case cohort comprised 6,500 adolescents with cognitive outcomes, covariates, and regional tfMRI features after imaging and structural-protocol quality checks.Source: A. Cohort and Data Acquisition · p. 3
  • Five-fold family-aware cross-validation kept siblings and twins in the same partition, with separate training, validation, and held-out test groups in each iteration.Three groups trained the model, one supported validation and early stopping, and one served as the test set.Source: A. Cohort and Data Acquisition · p. 3

B. Cognitive Phenotyping and Covariate Integration

The methods define three NIH Toolbox cognitive outcomes, construct regional task-activation and contrast features, and introduce MAC-Net’s separate neural and covariate pathways.

B. Cognitive Phenotyping and Covariate Integration

  • Cognitive targets were fluid, crystallized, and total intelligence, with training-fold-only standardization preserving the uncorrected score scale.The covariate vector included age, sex, MRI device, eTIV, and task-matched framewise displacement.Source: B. Cognitive Phenotyping and Covariate… · p. 3

C. Task-fMRI Feature Engineering

D. MAC-Net Architecture and Optimization

  • MAC-Net routes high-dimensional tfMRI inputs and measured covariates through separate pathways, keeping covariates out of early neural feature extraction.The architecture combines the pathways through late fusion.Source: D. MAC-Net Architecture and Optimization · p. 3

1) Deep Feature Extraction:

MAC-Net extracts regional tfMRI representations before terminal covariate fusion, and the study compares task configurations and model architectures under shared evaluation conditions.

1) Deep Feature Extraction:

  • The tfMRI pathway projects regional activations into a 64-dimensional embedding before a three-layer backbone with progressively decreasing width.Batch normalization and dropout are applied at each hidden layer.Source: 1) Deep Feature Extraction · p. 3

2) Late Fusion of Covariates:

  • Encoded covariates bypass neural feature extraction and join the processed neural embedding only at the terminal readout.The resulting prediction is produced from the concatenated representation.Source: 2) Late Fusion of Covariates · p. 3

3) Training Protocol:

  • Training used AdamW with a maximum of 300 epochs, validation-based learning-rate reduction, and early stopping.Evaluation reported MAE, MSE, R2, and Pearson correlation after outputs were returned to the original score scale.Source: 3) Training Protocol · p. 3

E. Ablation Framework and Baselines

  • Baselines included standard regressors, individual-task and brain-only ablations, and three alternative deep architectures evaluated with family-aware partitions and the late-fusion mechanism.The alternatives were SpatialGATNet, a Tabular Transformer, and a Gated Tabular Network.Source: E. Ablation Framework and Baselines · pp. 3–4

F. Multi-Method Attribution Consensus

The attribution procedure standardized participant-level explanations against the training-partition mean and aggregated categorical dummy-variable attributions. It then formed a consensus across three methods and measured their agreement.

F. Multi-Method Attribution Consensus

  • Zero represented the training-partition mean for standardized features, and categorical dummy-variable attributions were aggregated across their columns.Attribution values were computed for held-out test participants using Integrated Gradients, DeepLIFT, and Input×Gradient.Source: F. Multi-Method Attribution Consensus · p. 4
  • Absolute attribution values were averaged across participants and min-max normalized to [0, 1].Source: F. Multi-Method Attribution Consensus · p. 5
  • The consensus score averaged normalized scores across three methods, while pairwise Spearman rank correlations quantified methodological convergence.Because covariates entered through late fusion, the resulting neural maps represent the fitted model’s attributions to tfMRI inputs.Source: F. Multi-Method Attribution Consensus · p. 5

III. RESULTS AND DISCUSSION

Across the evaluated tasks, N-back performed best and SST performed worst for all three cognitive outcomes. Adding SST to N-back+MID did not improve descriptive performance, while task-selection differences warrant caution because selection used the same cohort.

A. Task-Specific Cognitive Modeling and Task Selection

  • Covariates alone accounted for 11.4% to 17.8% of variance in uncorrected cognitive scores, making them a substantial baseline for task-fMRI prediction.Source: A. Task-Specific Cognitive Modeling… · p. 5
  • N-back had the highest mean R2 across Gf, Gc, and Gt (0.178 ± 0.014, 0.223 ± 0.022, and 0.269 ± 0.016), while MID was intermediate and SST lowest.The consistent task ordering indicates that predictive utility was task-dependent rather than determined solely by feature volume.Source: A. Task-Specific Cognitive Modeling… · p. 5
  • Working-memory-related activation and contrast features were the most informative task-specific neural signal in this representation, but N-back performance does not reduce cognition to working memory.N-back responses also reflect visual processing, sustained attention, response selection, motivation, and general task engagement.Source: A. Task-Specific Cognitive Modeling… · p. 5
  • MID performed below N-back but retained predictive information; its inclusion descriptively improved Gc and Gt, though task selection used performance within the same cohort.The observed differences are therefore not selection-independent estimates.Source: A. Task-Specific Cognitive Modeling… · p. 5
  • SST was excluded from the primary multimodal configuration because it performed weakest across outcomes; possible explanations include measure reliability or the selected contrasts and representation.Its exclusion is an empirical property of this pipeline, not evidence that inhibitory-control circuitry is irrelevant to cognition.Source: A. Task-Specific Cognitive Modeling… · p. 5
  • Adding SST to N-back+MID changed mean R2 by approximately −0.001 for Gf, −0.005 for Gc, and −0.004 for Gt, with no descriptive evidence of incremental predictive information.The three-task configuration used 658 additional regional features; the authors suggest that adding features without substantial complementary variance may exacerbate the curse of dimensionality.Source: A. Task-Specific Cognitive Modeling… · p. 5

B. Multi-Task Cognitive Estimation

The N-back+MID MAC-Net configuration had the strongest overall performance among evaluated models, with phenotype-dependent gains over N-back alone. A brain-only ablation retained predictive utility, while late fusion limits covariate influence on early neural learning but does not guarantee deconfounding.

B. Multi-Task Cognitive Estimation

  • Mean R2 was 0.174 for Gf, 0.238 for Gc, and 0.277 for Gt with N-back+MID MAC-Net, exceeding ElasticNet for Gf and Lasso for Gc and Gt by 0.023, 0.033, and 0.034, respectively.The primary configuration produced the strongest overall performance among evaluated models.Source: B. Multi-Task Cognitive Estimation · p. 5
  • N-back+MID exceeded N-back alone by 0.015 in mean R2 for Gc and 0.008 for Gt, while Gf was similar (0.174 ± 0.009 versus 0.178 ± 0.014).These gains remain descriptive because participant-level paired uncertainty was not calculated.Source: B. Multi-Task Cognitive Estimation · p. 5
  • Without participant and acquisition covariates, MAC-Net achieved mean R2 of 0.110 for Gf, 0.134 for Gc, and 0.174 for Gt, exceeding the Random Forest covariate-only baseline across outcomes.This supports complementary contributions from tfMRI and covariate feature spaces, but is not a formal decomposition of statistically independent variance.Source: B. Multi-Task Cognitive Estimation · pp. 5–6
  • Late fusion routes tfMRI through a dedicated neural pathway and adds participant and acquisition variables only at the terminal prediction stage, limiting their direct influence on early neural representation learning.Source: B. Multi-Task Cognitive Estimation · p. 6
  • Late fusion constrains information flow but does not guarantee a covariate-independent neural representation, because tfMRI may encode measured and unmeasured sources of variation.The authors distinguish this architectural choice from statistical deconfounding or causal modeling.Source: B. Multi-Task Cognitive Estimation · p. 6

C. Architectural Benchmarking

MAC-Net outperformed the strongest alternative deep model in the architectural benchmark. Attribution analyses found highly concordant feature rankings, with neural patterns varying by cognitive outcome and showing fold-wise stability.

C. Architectural Benchmarking

  • The findings favor an architecture matched to atlas-averaged regional task-fMRI features over maximal architectural complexity.The lower performance of the tested graph and self-attention models does not establish that those mechanisms are intrinsically unsuitable for task-fMRI.Source: C. Architectural Benchmarking · pp. 6–7
  • MAC-Net exceeded the strongest alternative deep model by 0.038, 0.060, and 0.060 mean R² for Gf, Gc, and Gt, respectively.The Gated Tabular Network was the strongest alternative, with mean R² values of 0.136, 0.178, and 0.217 for Gf, Gc, and Gt.Source: C. Architectural Benchmarking · p. 6, Table III · p. 7

D. Model Attribution and Neural Feature Stability

  • Age ranked highest among covariate attributions for all three outcomes; normalized absolute attributions measure sensitivity, not variable direction or causal effect.Scanner device, eTIV, sex, and task-specific motion also had substantial attribution magnitude, with outcome-dependent ordering.Source: D. Model Attribution and Neural… · p. 7
  • N-back 2-back versus 0-back features dominated the leading Gf and Gt neural attributions, including frontal, cingulate, and parietal regions.The pattern broadly overlaps with distributed frontal–parietal systems implicated in working memory and higher-order cognitive control,,.Source: D. Model Attribution and Neural… · p. 7
  • Gc attributions combined N-back features with MID loss-anticipation and feedback signals in pericalcarine, cingulate, and frontal regions.The authors suggest this may reflect shared attention, motivation, or generalized task engagement rather than a specific reward-based mechanism.Source: D. Model Attribution and Neural… · pp. 7–8
  • Integrated Gradients, DeepLIFT, and Input×Gradient rankings correlated above ρ = 0.996 across outcomes, under the shared zero-reference baseline.Fig. 3 summarizes top task-condition-region features, normalized absolute covariate attributions, and pairwise rank correlations.Source: D. Model Attribution and Neural… · p. 7, Figure 3 · p. 8
  • Leading attributed regions retained non-zero mean importance across all five folds, although exact feature rankings varied between folds.The broad spatial distribution was more reproducible than the precise hierarchical ordering.Source: D. Model Attribution and Neural… · p. 8
  • Agreement does not establish biological validity: the features are model-supported neural correlates, not definitive biomarkers or causal mechanisms.The explainers share a gradient-based approach and reference convention, and correlated features can exchange attribution magnitude without materially changing predictions,.Source: D. Model Attribution and Neural… · p. 8

E. Limitations and Translational Implications

The findings are bounded by the study’s cross-sectional, single-cohort design and descriptive performance comparisons. Translational use will require independent validation and testing in neurologically impaired and rehabilitation populations.

E. Limitations and Translational Implications

  • The analysis used cross-sectional baseline data from one developmental cohort, and task-selection comparisons lacked formal paired statistical testing.The authors call for longitudinal modeling, nested cross-validation, and participant-level paired inference to obtain selection-independent estimates.Source: E. Limitations and Translational… · p. 8
  • Complete-case filtering may introduce sociodemographic selection bias, while atlas-averaged beta estimates discard fine-grained spatial variation and temporal dynamics.The authors call for evaluations using alternative missing-data strategies and empirical tests of more complex voxelwise or dynamic representations.Source: E. Limitations and Translational… · p. 8
  • Fig. 4 displays attributed regions and fold-wise importance; leading regions had non-zero mean importance across folds, but exact rankings varied.The broad spatial distribution was more reproducible than the precise hierarchical ordering.Source: E. Limitations and Translational… · p. 8, Figure 4 · p. 9

IV. CONCLUSION

MAC-Net separates task-fMRI representation learning from participant and acquisition variables through late fusion. In this evaluation, N-back was the strongest single task, and integrating MID yielded a high-performing multi-task configuration with concordant attributions to plausible neural networks.

IV. CONCLUSION

  • MAC-Net routes task-fMRI representation learning separately from participant and acquisition variables, combining them through late fusion.The framework was evaluated on 6,500 ABCD Study participants under family-aware cross-validation and outperformed conventional machine-learning and advanced deep-learning baselines.Source: §IV · p. 9
  • N-back provided the strongest single-task representation, and integrating MID established a parsimonious, high-performing multi-task configuration.Source: §IV · p. 9
  • Multi-method explainability analyses localized predictive features to N-back-related frontal, parietal, and cingulate regions.The attributions are associative correlates rather than causal biomarkers, and the features align with biologically plausible networks.Source: §IV · p. 9
  • MAC-Net provides a methodological foundation for future neural-engineering applications; clinical deployment requires prospective multi-site validation and longitudinal testing in neurologically impaired populations.Source: §IV · p. 9

Supplementary Material: A Multi-Task Deep Learning Framework for Modeling

The supplementary material specifies the regional atlas and task-feature construction, lists Stop Signal Task contrasts, and identifies a brain-only ablation table.

Supplementary Material: A Multi-Task Deep Learning Framework for Modeling

  • Task-fMRI representations used 94 anatomical regions: 68 Desikan–Killiany cortical parcels and 26 bilateral ASEG subcortical segments.The supplementary section enumerates the cortical and subcortical regions.Source: Supplementary Material: A Multi-Task… · p. 12
  • Feature extraction yielded 846 N-back, 940 MID, and 658 SST features; the listed SST contrasts distinguish successful and failed inhibition, execution errors, and stop-signal processing.Features were extracted for each task condition or contrast across the 94 regions.Source: Table S1 · p. 12, Supplementary Material: A Multi-Task… · p. 12
  • Table S2 describes brain-only ablation performance using N-back+MID inputs without participant or acquisition covariates, reported as mean ± standard deviation across five held-out test folds.Source: Table S2 · p. 13
Loading 2609.33440v1…