PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2609.33428v1cs.AI

TL;DR

The paper asks whether multi-week wearable traces can provide a continuous signal for adolescent crystallized intelligence despite irregular observations and interacting behaviors. It models daily traces as temporal graphs with invalid-day pruning and demographic fusion. The conclusion presents missingness-aware temporal graphs as a framework for digital cognitive phenotyping, with reported behavioral signals and performance consistency across demographic strata.

  • Problem

    The study asks whether multi-week consumer wearable traces contain a reliable, continuous signal for standardized adolescent cognitive ability.

    Source: §I · p. 1

  • Method

    SATURN represents daily wearable observations as temporal graphs, prunes invalid-day edges during message passing, and fuses pooled temporal features with sociodemographic covariates.

    Source: §I · p. 1, E. SATURN Architecture · p. 3

  • Results

    The conclusion reports that SATURN recovers a behavioral hierarchy predictive of adolescent Gc, with absolute error consistent across household income, parental education, and adherence strata.

    Source: §IV · p. 9

  • Takeaways & Limitations

    The study presents missingness-aware temporal graphs as a rigorous substrate for continuous digital cognitive phenotyping that respects the sociotechnical structure of wearable data.

    Source: §IV · p. 9

    The study uses United States ABCD adolescents and daily summaries, so external validation is needed before generalizing across countries, ages, devices, or protocols, and within-day dynamics are not captured.

    Source: F. Limitations and Implications for… · p. 8

Abstract

from arXiv · show

Wearable actigraphy offers a scalable, ecologically valid alternative to episodic clinical assessment. However, predicting continuous adolescent crystallized intelligence ($G_c$) from such traces remains challenging due to irregular device adherence and complex behavioral-environmental interactions. We address this using daily summary data derived from 21-day Fitbit records of 6,091 adolescents in the Adolescent Brain Cognitive Development Study (Release 5.1). We propose SATURN, a Sleep-Activity Temporal Unified Regression Network. It represents participants as 21-node temporal graphs encoding daily behaviors and temporal adjacency. To prevent imputation artifacts, invalid-day edges are dynamically pruned during forward passes. Node embeddings are refined via residual GATv2 layers, aggregated through masked attention pooling, and fused with sociodemographic covariates. Under family-controlled, age-sex-BMI-stratified cross-validation, SATURN achieves $R^2 = 0.2783 \pm 0.0127$, consistently improving upon flattened machine learning (Gradient Boosting, $R^2 = 0.2372$) and sequential deep learning (BiLSTM, $R^2 = 0.2688$) baselines. Explainability analyses identify light activity, metabolic equivalents, and sleep duration as dominant predictors, while Monte Carlo dropout and subgroup analyses confirm equitable performance across sociodemographic strata. Ultimately, SATURN establishes a rigorous computational framework for digital cognitive phenotyping, offering a scalable pathway to complement traditional assessments by highlighting macro-level behavioral anomalies.

I. INTRODUCTION

The study asks whether multi-week wearable traces can provide a continuous signal for adolescent crystallized intelligence (Gc). It introduces SATURN, a temporal-graph approach designed to address missing observations, integrated daily behaviors, and demographic confounding.

I. INTRODUCTION

  • The central question is whether multi-week consumer wearable traces contain a reliable, continuous signal for standardized adolescent cognitive ability.The study focuses on Gc, a construct reflecting accumulated language, reading, and knowledge-based skills.Source: §I · p. 1
  • Wearable modeling must address irregular observations, interdependent sleep and activity, and demographic factors that co-vary with adherence and Gc.Treating absent days as zero activity can conflate measurement failure with genuine sedentary behavior.Source: §I · p. 1
  • SATURN represents daily observations as temporal graphs, pruning invalid days and fusing pooled behavioral information with sociodemographic covariates.The design targets cross-channel relationships while distinguishing informative absence from true low activity.Source: §I · p. 1
  • The study evaluates SATURN on ABCD data using family-controlled stratified cross-validation and subgroup-robustness and fairness analyses.Its stated aim is to establish missingness-aware temporal networks as a framework for extracting continuous cognitive-behavioral signals from consumer wearables.Source: §I · p. 2

II. MATERIALS AND METHODS

The methods use daily Fitbit activity and sleep data aligned to a fixed 21-day calendar, with demographic and developmental covariates included to assess incremental wearable signal. Participants are modeled using SATURN’s temporal-graph pipeline.

II. MATERIALS AND METHODS

  • SATURN maps daily behavioral streams to a fixed 21-day calendar and participant-specific temporal graph, then pools graph features before fusing sociodemographic covariates.Residual GATv2 message passing is applied only over valid-day edges.Source: §II · p. 2

A. Data Source, Target, and Covariates

  • The analysis uses ABCD Release 5.1 two-year follow-up Fitbit records and uncorrected NIH Toolbox Gc scores, with six demographic and developmental baseline covariates.Covariates include age, biological sex, BMI, study site, parental education, and household income.Source: A. Data Source, Target, and Covariates · p. 2

B. Calendarization and Daily Wearable Features

  • The cohort includes 6,091 adolescents with at least seven valid activity days or seven valid sleep days within the 21-day observation window.Daily features cover calendar context, activity, sleep duration and stages, heart-rate summaries, and sleep timing.Source: B. Calendarization and Daily Wearable… · pp. 2–3

C. Validity Masks, Encoding, and Missing-Value Handling

The preprocessing preserves observation validity through explicit masks, applies fold-specific standardization, and evaluates the model using family-aware stratified splits.

C. Validity Masks, Encoding, and Missing-Value Handling

  • A day is marked valid when usable activity wear-quality or sleep information is available, with separate indicators tracking activity, daytime activity, sleep, and wear quality.Remaining missing feature values are set to zero while the explicit validity mask distinguishes them from true observations.Source: C. Validity Masks, Encoding, and… · p. 3
  • Calendar and clock variables receive cyclic encodings, and the model uses 47 temporal node features plus the graph-validity mask as a 48th input channel.Continuous features are standardized using only valid training nodes; other missing values are set to the training-set mean after standardization.Source: C. Validity Masks, Encoding, and… · p. 3

D. Family-Aware Stratified Cross-Validation

  • Five-fold cross-validation keeps family members in the same split and stratifies by age, sex, and BMI tertiles.Validation data supports learning-rate scheduling, early stopping, and checkpoint selection, while the resulting folds were closely balanced on sample size, family count, sex, age, BMI, and Gc.Source: D. Family-Aware Stratified… · p. 3

E. SATURN Architecture

SATURN represents each 21-day wearable window as a temporal graph, with calendar days as nodes and temporal relations encoded on edges. It prunes communication through invalid days, then combines graph-derived wearable features with sociodemographic covariates to predict crystallized intelligence.

E. SATURN Architecture

  • Each participant’s wearable window is a directed 21-node graph with a fully connected base topology and 441 possible edges, including self-loops.The effective topology becomes participant-specific after invalid-day edges are pruned.Source: E. SATURN Architecture · p. 3
  • Edges encode signed direction, normalized distance, self-connections, adjacent-day relations, and weekly-cycle relations.A learnable distance-decay channel is added to discourage unsupported long-range routing.Source: E. SATURN Architecture · p. 3
  • During each forward pass, an edge is retained only when both endpoint days are valid, preventing invalid days from sending or receiving messages.The calendar-aligned graph excludes invalid-day communication while preserving the same base topology for all participants.Source: E. SATURN Architecture · p. 3
  • Node features pass through two residual GATv2 layers, each with four attention heads and 16 hidden dimensions per head.The layers use temporal edge attributes, and the residual update adds the layer output to the current node representation.Source: E. SATURN Architecture · pp. 3–4
  • Masked attention pooling uses a trainable query to aggregate day embeddings while excluding invalid nodes before softmax normalization.Invalid nodes receive negative infinity before softmax.Source: E. SATURN Architecture · p. 4
  • The graph-level wearable embedding is concatenated with encoded sociodemographic covariates and passed through a regression head to predict Gc.Age and BMI are standardized, while categorical covariates are one-hot encoded and passed through a 32-dimensional MLP branch.Source: E. SATURN Architecture · p. 4

F. Training and Evaluation Protocol

The evaluation fits data-dependent preprocessing within each training fold and transforms predictions back to the original score scale for held-out assessment. Training uses validation-based learning-rate scheduling and early stopping, with the best checkpoint evaluated once on the test fold.

F. Training and Evaluation Protocol

  • All data-dependent preprocessing was fit on each training split, and the target was quantile-transformed using training targets before predictions were inverse-transformed for evaluation.Validation and test transformations were applied using the corresponding training-fold fits.Source: F. Training and Evaluation Protocol · p. 4
  • SATURN was trained with AdamW using learning rate 10^-3, weight decay 10^-4, batch size 128, and a maximum of 300 epochs.Training minimized mean squared error in transformed target space, with gradient clipping and dropout regularization.Source: F. Training and Evaluation Protocol · p. 4
  • Validation R2 governed learning-rate reductions and early stopping, and the best-validation checkpoint was evaluated once on the held-out test fold.Early stopping used a patience of 30 epochs; the scheduler reduced the learning rate after 10 stagnant epochs.Source: F. Training and Evaluation Protocol · p. 4

G. Explainable AI and Uncertainty Quantification

The authors assess feature importance with permutation and Integrated Gradients, and inspect learned attention as a routing diagnostic rather than a causal explanation. Monte Carlo dropout supplies an uncertainty proxy from repeated stochastic predictions.

G. Explainable AI and Uncertainty Quantification

  • Permutation feature importance measures the test R2 decrease after shuffling feature groups, while preserving each wearable feature’s within-feature temporal shape.Grouped cyclic features, week indicators, and categorical covariates are permuted together.Source: G. Explainable AI and Uncertainty… · pp. 4–5
  • Integrated Gradients uses 50 interpolation steps and keeps the graph-validity mask fixed along the integration path.The baseline sets standardized biological temporal features and covariates to zero; absolute attributions are averaged across subjects and folds.Source: G. Explainable AI and Uncertainty… · p. 5
  • Attention weights are averaged across heads and test subjects to form 21 × 21 maps, which the authors interpret as diagnostics of learned message routing, not causal explanations.Source: G. Explainable AI and Uncertainty… · p. 5
  • Monte Carlo dropout estimates epistemic uncertainty with 100 stochastic forward passes per test subject, using prediction standard deviation as the uncertainty proxy.The mean of inverse-transformed predictions is used as the point prediction.Source: G. Explainable AI and Uncertainty… · p. 5

III. RESULTS AND DISCUSSION

Covariate-only models provide a strong Gc baseline, and SATURN achieves the highest overall performance among the compared modeling regimes. Its results indicate that wearable traces add behavioral signal beyond covariates, while that signal overlaps with sociodemographic information.

A. Covariate Baseline and Demographic Signal

  • R2 = 0.2346 ± 0.0205 and r = 0.4868 ± 0.0228: Lasso was the strongest covariate-only baseline, making incremental wearable information the relevant comparison.Age, sex, BMI, parental education, household income, and study site formed the covariate set.Source: A. Covariate Baseline and Demographic… · p. 5

B. Multimodal Model Comparison and Sequence Topologies

  • R2 = 0.2688 ± 0.0154: masked-mean BiLSTM exceeded the demographic baseline, whereas Gradient Boosting reached R2 = 0.2372 ± 0.0136.Replacing BiLSTM masking with conventional mean or forward-fill imputation degraded performance.Source: B. Multimodal Model Comparison and… · p. 5
  • R2 = 0.2783 ± 0.0127 and r = 0.5304 ± 0.0104: SATURN achieved the highest overall performance, while imputed inputs reduced its R2 to 0.2735.The comparison used family-aware cross-validation; SATURN dynamically pruned invalid nodes and used learnable distance decay.Source: B. Multimodal Model Comparison and… · p. 5
  • R2 = 0.1221 ± 0.0088 without covariates, while the fused model's R2 = 0.2783 remained below the sum of independent baseline R2 values.The authors interpret this sub-additive overlap as a modest but reproducible behavioral signal layered on strong sociodemographic structure.Source: B. Multimodal Model Comparison and… · p. 5

C. Explainable AI Convergence and Behavioral Interpretation

The two attribution methods agree on the broad hierarchy of model signals but differ in some feature-specific rankings. Temporal attention offers a separate diagnostic of how information is routed across days.

C. Explainable AI Convergence and Behavioral Interpretation

  • Spearman’s ρ = 0.740, p < 0.001: PFI and IG showed substantial rank agreement after structural masks, temporal indices, and device-compliance variables were excluded.PFI measures performance drop after feature disruption; IG measures local gradient-based attribution, so their scales need not match.Source: C. Explainable AI Convergence and… · p. 5
  • Light activity, average daily METs, sleep-period measures, and sedentary duration were consistent temporal predictors; parental education, age, income, and site ranked highly among covariates.The highest parental education was the dominant demographic predictor in the reported XAI analysis.Source: C. Explainable AI Convergence and… · p. 6
  • Covariates clustered near the highest ranks across both methods, while temporal-feature ranks were more dispersed, reflecting agreement on broad hierarchy but disagreement on some ordering.Correlated or redundant features can receive different PFI and IG rankings because the methods assess different quantities.Source: C. Explainable AI Convergence and… · p. 6
  • The final-layer attention map showed stronger mass in local chronological neighborhoods and broader off-diagonal structure across the 21-day window.These normalized, feature- and mask-dependent weights represent learned routing patterns, not standalone feature-importance scores.Source: C. Explainable AI Convergence and… · p. 6, Figure 2 · p. 7

D. MC Dropout Epistemic Uncertainty Mapping

Monte Carlo dropout uncertainty indicates that SATURN is more reliable for ranking and central-range estimation than for precise predictions at extreme Gc values. The estimates remain approximate and do not fully address interval calibration or unmeasured confounding.

D. MC Dropout Epistemic Uncertainty Mapping

  • Lower true scores tended to be overestimated and higher true scores underestimated, making SATURN more reliable for ranking and central-range estimation than precise extreme-value prediction.Predictive standard deviation also increased toward both Gc tails, consistent with fewer extreme target values.Source: D. MC Dropout Epistemic Uncertainty… · p. 7
  • MC dropout captures one source of epistemic uncertainty but does not fully calibrate predictive intervals or account for unmeasured confounding.The authors call for calibration diagnostics, interval-coverage assessment, and subgroup-specific uncertainty analysis.Source: D. MC Dropout Epistemic Uncertainty… · p. 7

E. Subgroup Robustness and Calibration

SATURN shows comparable absolute errors across several subgroups, alongside a low-education calibration and within-group R² concern. The discussion limits generalization to a cross-sectional U.S. adolescent cohort and daily summaries, and notes that missingness and sparse graphs may constrain interpretation.

E. Subgroup Robustness and Calibration

  • Females and males had similar absolute error, with MAE values of 4.60 and 4.64, respectively, and small calibration biases.Pooled held-out performance was R² = 0.279, r = 0.529, MAE = 4.62, and RMSE = 5.90.Source: E. Subgroup Robustness and Calibration · p. 7, Table IV · p. 9
  • Figure 3 shows predictive uncertainty increasing toward the lower and upper Gc tails, where fewer observations are available.The authors say this makes the model more reliable for ranking and central-range estimation than for precise prediction of extreme Gc values.Source: Figure 3 · p. 8, E. Subgroup Robustness and Calibration · p. 8
  • MAE ranged from 4.56 to 4.76 across parental-education groups and from 4.52 to 4.83 across household-income groups, despite differing Gc distributions.The low-parental-education group had an average over-prediction of 0.95 points and within-group R² = 0.029; the authors note restricted target variance and smaller sample size (N=280) as partial explanations.Source: E. Subgroup Robustness and Calibration · pp. 7–8
  • Participants with fewer than 21 valid wearable days did not show degraded performance relative to the high-adherence group.Source: E. Subgroup Robustness and Calibration · p. 8

F. Limitations and Implications for Computational Social Systems

  • The study is cross-sectional and supports predictive association rather than causal inference; longitudinal trajectory modeling remains future work.Source: F. Limitations and Implications for… · p. 8
  • The cohort comprises U.S. adolescents in ABCD, and the analysis uses daily summaries rather than raw minute-level sensor streams.External validation is needed before generalizing across countries, age ranges, devices, or collection protocols; daily summaries omit within-day timing, bout structure, and finer sleep-stage dynamics.Source: F. Limitations and Implications for… · p. 8
  • Dynamic masking isolates unobserved days, but missingness may still encode social structure and extensive edge pruning can leave highly non-compliant participants with sparse graphs.The authors call for subgroup-specific reporting of performance, calibration, and uncertainty across demographic, site, and adherence strata.Source: F. Limitations and Implications for… · pp. 8–9
  • The framework links sociodemographic context to a temporal graph that captures incremental activity–sleep signal under explicit validity constraints.Source: F. Limitations and Implications for… · p. 9

IV. CONCLUSION

The conclusion presents missingness-aware temporal graphs as a framework for extracting behavioral signals from incomplete wearable traces to predict adolescent Gc. It also points to uncertainty estimates and macro-level behavioral deviations as potential aids for digital phenotyping and prioritizing traditional assessments.

IV. CONCLUSION

  • SATURN prunes invalid days within the temporal graph and recovers a behavioral hierarchy—including light activity, metabolic equivalents, and sleep duration—that predicts adolescent Gc above sociodemographic baselines.Subgroup analyses report consistent absolute error across household-income, parental-education, and adherence strata.Source: §IV · p. 9
  • Epistemic confidence tracks data density, offering a deployment signal rather than uniform confidence across predictions.Source: §IV · p. 9
  • The framework is proposed as a scalable computational substrate for identifying macro-level deviations in expected cognitive-behavioral trajectories and helping prioritize in-depth clinical assessments.Source: §IV · p. 9

Related papers

Loading 2609.33428v1…