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

cs.AISep 27, 2026

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

Abstract

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.

Figures & tables

Fig. 1: Structural pipeline of the proposed SATURN architecture. Fitbit activity and sleep records are mapped to a fixed 21-day calendar with explicit validity indicators. Each day is represented as a graph node, and the fully connected temporal base graph contains 441 directed edges with five temporal edge attributes. Invalid days are pruned from the participant-specific effective edge set during GATv2 message passing. The masked temporal representation is refined by two residual GATv2 layers, summarized by learned attention pooling, fused with encoded sociodemographic covariates, and passed to a regression head for crystallized intelligence prediction.

Fold	Split	Subjects	Families	Males	Females	Age (months)	BMI	$G_{c}$
1	Train	3904	3331	1976	1928	143.32 $\pm$ 7.78	20.48 $\pm$ 4.80	91.28 $\pm$ 6.93
	Validation	967	832	521	446	143.27 $\pm$ 7.68	20.37 $\pm$ 4.74	91.21 $\pm$ 7.09
	Test	1220	1040	637	583	143.51 $\pm$ 7.95	20.35 $\pm$ 4.79	91.24 $\pm$ 6.88
2	Train	3907	3330	2020	1887	143.36 $\pm$ 7.79	20.34 $\pm$ 4.74	91.31 $\pm$ 7.00
	Validation	963	831	493	470	143.36 $\pm$ 7.70	20.72 $\pm$ 4.98	91.06 $\pm$ 6.80
	Test	1221	1042	621	600	143.33 $\pm$ 7.88	20.51 $\pm$ 4.78	91.27 $\pm$ 6.90

TABLE I: Demographic and Crystallized Intelligence Score Summary Across Five Family-Aware Cross-Validation Folds (Mean

\pm

SD)

Method	$R^{2}\uparrow$	r $\bm{\uparrow}$	MAE $\bm{\downarrow}$	MSE $\bm{\downarrow}$
Ridge Regression	0.2345 $\pm$ 0.0217	0.4866 $\pm$ 0.0240	4.7687 $\pm$ 0.1114	36.8723 $\pm$ 1.3702
Lasso Regression	0.2346 $\pm$ 0.0205	0.4868 $\pm$ 0.0228	4.7625 $\pm$ 0.1131	36.8663 $\pm$ 1.3438
ElasticNet	0.2340 $\pm$ 0.0230	0.4858 $\pm$ 0.0253	4.7676 $\pm$ 0.1174	36.8966 $\pm$ 1.4225
Random Forest	0.1964 $\pm$ 0.0226	0.4462 $\pm$ 0.0239	4.8769 $\pm$ 0.1326	38.7122 $\pm$ 1.6239
Gradient Boosting	0.2207 $\pm$ 0.0128	0.4727 $\pm$ 0.0146	4.7996 $\pm$ 0.1106	37.5413 $\pm$ 1.3297
Support Vector Regression	0.2071 $\pm$ 0.0196	0.4633 $\pm$ 0.0263	4.8371 $\pm$ 0.1140	38.1910 $\pm$ 1.3762

TABLE II: Baseline Predictive Performance for Crystallized Intelligence (

G_{c}

) Using Covariates Only.

Flattened Machine Learning (Zero-Padding)
Method Configuration	$R^{2}\uparrow$	r $\bm{\uparrow}$	MAE $\bm{\downarrow}$	MSE $\bm{\downarrow}$
– Ridge Regression	0.2171 $\pm$ 0.0135	0.4696 $\pm$ 0.0132	4.8077 $\pm$ 0.1270	37.7280 $\pm$ 1.6871
– Lasso Regression	0.2287 $\pm$ 0.0121	0.4825 $\pm$ 0.0142	4.7788 $\pm$ 0.1177	37.1644 $\pm$ 1.4363
– ElasticNet	0.2368 $\pm$ 0.0095	0.4897 $\pm$ 0.0111	4.7434 $\pm$ 0.1094	36.7727 $\pm$ 1.3978
– Random Forest	0.2071 $\pm$ 0.0104	0.4622 $\pm$ 0.0140	4.8461 $\pm$ 0.1279	38.2130 $\pm$ 1.6671
– Gradient Boosting	0.2372 $\pm$ 0.0136	0.4902 $\pm$ 0.0150	4.7523 $\pm$ 0.1001	36.7395 $\pm$ 1.1104

TABLE III: Comparison of Predictive Performance for Crystallized Intelligence (

G_{c}

) Across Methodological Families.

Fig. 2: Explainable AI (XAI) and temporal attention diagnostics for the SATURN architecture. Panel (a) shows top-predictor rank stability between Permutation Feature Importance (circle) and Integrated Gradients (diamond), with connecting segments showing method disagreement. Panel (b) shows global rank convergence between the two XAI methods (Spearman’s $\rho=0.740$ , $p<0.001$ ), separating covariates from temporal wearable features. Panel (c) visualizes the mean final-layer $21\times 21$ edge-attention matrix, where rows denote receiving days and columns denote transmitting days.

Fig. 3: Epistemic uncertainty estimation via 100 Monte Carlo dropout passes. Panel (a) compares true crystallized intelligence scores with the Monte Carlo predictive mean; vertical error bars and point color indicate the predictive standard deviation. Panel (b) plots true score against epistemic uncertainty using a bivariate density overlay and a quadratic trend line, showing lower uncertainty near the densely sampled population center and higher uncertainty toward the score tails.

Subgroup	N	$\bm{G_{c}}$ Mean $\pm$ SD [Range]	Bias	MAE $\bm{\downarrow}$	RMSE $\bm{\downarrow}$	r $\bm{\uparrow}$	$\bm{R^{2}\uparrow}$
Overall Cohort	6091	91.26 $\pm$ 6.94 [66–125]	-0.19	4.62	5.90	0.529	0.279
Biological Sex
– Female	2957	91.10 $\pm$ 7.00 [66–121]	-0.21	4.60	5.87	0.546	0.297
– Male	3134	91.41 $\pm$ 6.89 [67–125]	-0.18	4.64	5.92	0.512	0.261
Parental Education
– Low ( $\leq$ 12 yrs)	280	83.66 $\pm$ 6.15 [66–104]	0.95	4.76	6.05	0.324	0.029

TABLE IV: Subgroup Robustness and Calibration Analysis for SATURN (Pooled Held-Out Predictions)

Explore similar work

Explainable Deep Learning of Resting-State Functional Connectomes Reveals Network Biomarkers of Adolescent Intelligence
Sep 27, 2026Md. Tanvir Rahman, Nabil Anan Orka, Asaduzzaman Khan +1Developmental Trajectories Biomarker
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
TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding
Jul 31, 2026Sparsh Rastogi, Tanmay Kumar, Baiyu Chen +3Human Activity Recognition Reasoning Traces
University of New South Wales, Sydney, Australia · Thapar Institute of Engineering and Technology, Patiala, India
Uncovering Trajectory and Topological Signatures in Multimodal Pediatric Sleep Embeddings
May 13, 2026Scott Ye, Harlin LeeSleep Multimodal Embeddings
Department of Radiology, University of California, San Francisco · School of Data Science and Society, University of North Carolina at Chapel Hill

Sep 27, 2026, q-bio.NCJ/K move · Enter open · S save

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

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

Developmental Trajectories Biomarker Deep Learning Intelligence

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding

Uncovering Trajectory and Topological Signatures in Multimodal Pediatric Sleep Embeddings