Can covariate-aware multi-task deep learning estimate cognitive function from task fMRI?

What the R² values are compared against

The headline numbers are not measured against zero. A covariate-only model using interview age, biological sex, MRI device serial number, estimated total intracranial volume and task-matched mean framewise displacement already explained 11.4% to 17.8% of the variance in uncorrected NIH Toolbox scores. That makes 0.178 the real floor for any claim that task-fMRI adds information, and it is the number the neural model has to clear.

MAC-Net clears it, but by different margins depending on the target. For fluid intelligence the N-back plus Monetary Incentive Delay (MID) configuration reached R² = 0.174 ± 0.009, essentially level with the covariate-only baseline. Crystallized intelligence reached 0.238 ± 0.017 and total cognition 0.277 ± 0.008, both above that floor. The abstract's phrasing that the model outperformed covariate-only baselines (0.178) and alternative deep models (0.217) is best read as a comparison across targets, not a single like-for-like gain on every outcome.

The same table shows why architecture alone is not the story. Ridge, Lasso, ElasticNet and Random Forest were trained on the identical multimodal feature set and family-aware folds; ElasticNet reached 0.151 for fluid, 0.205 for crystallized and 0.240 for total cognition. Among deep alternatives, the Gated Tabular Network reached 0.136, 0.178 and 0.217. MAC-Net's advantage over the strongest conventional and deep baselines is therefore real but small, and largest on the composite score.

Why covariates get their own late pathway

The design problem is that demographic and scanner variables are strong, low-dimensional predictors that can swamp high-dimensional neural inputs during training. MAC-Net's response is structural: the flattened regional tfMRI activations pass through a projection layer that maps 94 anatomical regions into a 64-dimensional embedding, then through a three-layer feed-forward backbone (256→128→64 units) with batch normalization, SiLU nonlinearities and dropout of 0.7. The covariate vector bypasses this extractor entirely and is concatenated only at the terminal readout, where the prediction is formed from the neural embedding and the covariates together.

That constraint forces the upstream hidden layers to optimize for functional neural variance rather than for age or head motion. The paper frames this as preventing dominant covariates from suppressing clinical representations during feature learning. It is a design claim about training dynamics, not a demonstrated causal mechanism, and the evidence for it is the ablation structure rather than a direct measurement of what the hidden layers encode.

The input representation is deliberately coarse. Features are task activation (beta) estimates and functional contrasts averaged across 68 Desikan–Killiany cortical parcels and 26 bilateral ASEG subcortical segments, drawn from three ABCD paradigms: 9 N-back conditions (846 features), 10 MID conditions (940 features) and 7 Stop Signal Task conditions (658 features), for 26 unique task-evoked brain states. Predictions are therefore made from region-level summaries, not voxelwise or dynamic representations.

N-back carries the signal and SST does not

Task choice mattered more than feature volume. In single-task runs, N-back produced the highest mean R² across all three outcomes (0.178 ± 0.014 for fluid, 0.223 ± 0.022 for crystallized, 0.269 ± 0.016 for total), MID was intermediate, and the Stop Signal Task was lowest across every domain. The ordering held consistently, which the authors read as evidence that predictive utility depended on the cognitive demands of the paradigm rather than on how many features it supplied.

Adding MID to N-back improved crystallized and total cognition descriptively, which the paper attributes to reward-related and outcome-monitoring states carrying information complementary to working-memory activity. Adding SST instead moved R² to 0.173 ± 0.015, 0.233 ± 0.015 and 0.273 ± 0.014, marginally below the N-back plus MID values of 0.174, 0.238 and 0.277. SST was therefore excluded from the primary configuration.

The authors are careful about what that exclusion means. They note it reflects the present pipeline — the chosen inhibition contrasts, the reliability of regional SST measures, the smaller feature space, or this architecture's sensitivity — and is not evidence that inhibitory-control circuitry is irrelevant to cognition. The task-selection procedure was also informed by performance within the same cohort and lacked formal paired statistical testing, so these differences are descriptive rather than selection-independent estimates.

Attribution agreement is not biological validation

Three gradient-based attribution methods — Integrated Gradients with 50 approximation steps, DeepLIFT and Input×Gradient — produced nearly identical global feature rankings, with pairwise Spearman correlations above ρ = 0.996 for all three cognitive targets. The top neural features for fluid and total cognition were dominated by the N-back 2-back versus 0-back contrast, in regions including the left caudal middle frontal gyrus, left pars triangularis, rostral and caudal anterior cingulate cortices, and right inferior and superior parietal regions. That pattern overlaps distributed frontal–parietal systems associated with working memory and cognitive control.

The crystallized-intelligence profile was less clean: it mixed N-back features with MID loss-anticipation and feedback signals in pericalcarine, cingulate and frontal regions. Because MID does not directly measure crystallized knowledge, the authors interpret this as the model capturing shared variance in attention, motivation or general task engagement rather than a specific reward mechanism.

The paper is explicit that convergence among explainers does not establish biological validity. All three methods are gradient-based, applied to the same fitted model under the same zero-reference convention, and highly correlated features can trade attribution magnitude without changing predictions. Fold-wise stability showed the leading regions retained non-zero mean importance across all five folds, but exact rankings varied, so the broad spatial distribution is more reproducible than the precise ordering. The features are best treated as model-supported correlates, not biomarkers or causal mechanisms.

Covariate attributions are also magnitude-only. Age ranked highest for all three outcomes, followed in varying order by scanner device, eTIV, sex and task-specific motion. Because targets were uncorrected scores in a developing cohort, a dominant age contribution is expected; the normalized values quantify model sensitivity, not direction or causal effect.

What the family-aware folds protect

The evaluation used baseline data from 6,500 Adolescent Brain Cognitive Development Study participants who passed strict imaging quality control, held the recommended inclusion flags for the evaluated paradigms, and completed the FreeSurfer structural protocol. Five fixed, non-overlapping folds kept siblings and twins within the same fold, so shared genetic or environmental variance could not leak between training and test sets. Three folds trained, one validated with early stopping, and one was held out.

Preprocessing followed the same discipline. Cognitive targets were Z-scored using training-fold parameters only, continuous covariates were standardized and categorical ones one-hot encoded from the training partition, and outputs were inverse-transformed to the original uncorrected score scale before reporting mean absolute error, MSE, R² and Pearson correlation. Training used AdamW with a maximum learning rate of 8.7×10⁻⁴ and weight decay of 5×10⁻³, up to 300 epochs at batch size 128, with a plateau scheduler and early stopping at 30 epochs of patience.

The advanced baselines received identical training parameters, family-aware partitions, dropout and the same late-fusion covariate mechanism, which makes the architecture comparison fairer than a default-settings benchmark. The authors nonetheless caution that SpatialGATNet and the Tabular Transformer results do not show graph or self-attention mechanisms are unsuitable for task-fMRI, since their performance depends on graph construction, tokenization, positional encoding and tuning.

Where these estimates stop transferring

The cohort is cross-sectional baseline data from a single healthy developmental sample, so nothing here models within-person cognitive trajectories; the authors point to the longitudinal ABCD structure as the next step. The model was also developed in a healthy cohort, which the paper describes as a methodological foundation rather than immediate clinical utility. Deployment for objective cognitive evaluation would require independent multi-site validation and calibration in neurologically impaired and rehabilitation populations.

Two further constraints sit close to the results. The complete-case design means participants without complete cognitive outcomes, covariates and regional tfMRI features were excluded, so the sample is defined by data availability as well as by quality control. And the modest R² values, especially the fluid-intelligence result at parity with covariates alone, mean the practical value of the neural pathway is concentrated in the crystallized and total-cognition targets rather than spread evenly across all three.

The priority is the composite, not fluid intelligence

Read together, the results support a specific ordering of claims. The clearest incremental value of covariate-aware multi-task modeling in this cohort is on total cognition (R² = 0.277 against a 0.178 covariate floor) and crystallized intelligence (0.238), with N-back as the informative paradigm and MID adding a smaller complementary contribution. Fluid intelligence is the weakest case: 0.174 is indistinguishable from what age, sex, scanner, eTIV and motion already provide.

That split should drive how the framework is judged. A reader deciding whether task-fMRI adds anything beyond readily available participant variables should look at the gap over the covariate-only baseline per target, not at the best single R². On this evidence, the late-fusion design earns its keep mainly where the composite and crystallized targets are concerned, and the next informative test is whether those gaps survive nested cross-validation and participant-level paired inference rather than the descriptive comparisons reported here.

Sources used in this answer