Demographics alone already explain most of the variance
The paper's most important comparison is not between neural architectures but between wearable data and a strong non-wearable benchmark. Before any Fitbit signal entered the model, the authors fitted age, biological sex, BMI, study site, highest parental education, and combined household income to the uncorrected NIH Toolbox crystallized cognition score (Gc). Lasso regression on these six covariates alone reached R² = 0.2346 ± 0.0205 with r = 0.4868.
That result reframes the question. Crystallized intelligence reflects accumulated language, reading, and knowledge-based skill, and the paper notes it is strongly tied to developmental exposure and education-linked household context. So the meaningful test is whether activity and sleep sequences add reproducible predictive information beyond that demographic and site-adjusted benchmark, not whether a wearable can replace it.
The authors also report the wearable-only version of their model, without covariates, at R² = 0.1221 ± 0.0088. That confirms an independent behavioral signal exists, but it is clearly weaker than the demographic baseline on its own.
Days become graph nodes so missing days can be cut out
SATURN, the Sleep-Activity Temporal Unified Regression Network, represents each participant as a 21-node temporal graph. Each node is one calendar day in a fixed 21-day window starting at the Fitbit protocol's day 0, carrying 47 temporal features plus a validity channel, for 48 node-level input channels. Edges encode temporal adjacency between days, and residual GATv2 layers refine node embeddings before masked attention pooling produces one graph-level representation.
The design targets a specific failure mode. Free-living wearable streams are incompletely observed: device non-wear, school schedules, and family routines create gaps that the paper argues are not missing at random and carry behavioral meaning. Treating an absent day as a zero-activity observation conflates measurement failure with genuine sedentary behavior.
SATURN's answer is to build validity indicators before any missing values are filled, then dynamically prune invalid-day edges during forward passes. Invalid nodes are topologically isolated, so message passing runs only over observed days, and their global attention weights are forced to zero. Participants qualified with at least seven valid activity days or at least seven valid sleep days, which is why the cohort retains adolescents with incomplete activity adherence.
The graph advantage is real but narrow
All three model families were evaluated on the same family-aware five-fold cross-validation folds, with siblings and twins kept in the same split to prevent shared familial information from leaking across training and test partitions. Stratification crossed biological sex with age and BMI tertiles.
Flattened models treat the 21 days as independent tabular columns. The best of them, Gradient Boosting, reached R² = 0.2372 ± 0.0136, only marginally above the covariate-only Lasso result of 0.2346. The paper reads this as evidence that simple concatenation preserves neither adjacent-day continuity nor a workable treatment of intermittent missingness.
A bidirectional LSTM with masked mean pooling, which processes days chronologically, reached R² = 0.2688 ± 0.0154. Replacing its masking with conventional mean or forward-fill imputation degraded performance, which the authors attribute to stale states or population-average noise being propagated through synthetically filled gaps.
SATURN reached R² = 0.2783 ± 0.0127 with r = 0.5304 ± 0.0104. An ablation that fed imputed data into the GATv2 layers instead of pruning invalid edges fell to R² = 0.2735, which the paper uses to argue that topological pruning, not just the graph architecture, drives the gain. The margins over the BiLSTM are small, so the practical claim is a consistent ordering rather than a large jump.
The wearable signal overlaps heavily with social structure
The fully fused model's R² = 0.2783 is substantially less than the arithmetic sum of the covariate baseline (0.2346) and the wearable-only model (0.1221). The authors call this sub-additive variance overlap and interpret it as evidence that adolescent movement behavior, sleep hygiene, and device compliance are bounded and scaffolded by sociodemographic reality.
That interpretation is the paper's own framing rather than a causal finding. What the numbers establish is a predictive decomposition: the wearable stream carries information that partly duplicates what demographics already capture, and the residual increment is modest.
The paper is explicit that the design supports predictive association rather than causal inference, and that longitudinal trajectory modeling would be needed to move beyond it.
Light activity, metabolic equivalents, and sleep duration lead
Two attribution methods were run on held-out test subjects. Permutation feature importance measured the drop in test R² after shuffling a feature group across subjects, preserving each feature's within-subject temporal shape while breaking the subject-feature correspondence. Integrated Gradients measured gradient-based attribution along a baseline-to-input path with 50 interpolation steps, holding the graph-validity mask fixed so the baseline could not alter graph topology.
After excluding structural masks, temporal index variables, and device-compliance variables, the two methods showed substantial rank agreement (Spearman's ρ = 0.740, p < 0.001). The paper notes they should not be expected to produce identical scales because they answer different questions.
The dominant predictors identified were light activity, metabolic equivalents, and sleep duration. Final-layer GATv2 attention weights were also extracted and averaged across four heads, test subjects, and folds, but the paper treats these maps as diagnostics of learned message routing rather than causal explanations.
Subgroup R² tracks within-group variance, not just model quality
Pooled held-out predictions give an overall mean absolute error of 4.62 Gc points and RMSE of 5.90 against an observed mean of 91.26 ± 6.94 (range 66–125). Bias, defined as predicted minus observed, was −0.19 overall, indicating close calibration on average.
Subgroup results need careful reading because R² depends heavily on within-group target variance, as the paper's own table note states. The low parental-education group (≤12 years, N = 280) shows R² = 0.029 with bias 0.95, while the high parental-education group (≥17 years, N = 3905) shows R² = 0.146. The low-household-income group (<$25k, N = 585) reaches R² = 0.230. These differences mix model behavior with how much the target varies inside each group.
Wearable adherence splits in the opposite direction from what a simple quality story would predict: participants with fewer than 21 valid days (N = 2345) show R² = 0.317, while those with 21 or more valid days (N = 3746) show R² = 0.232. Monte Carlo dropout with 100 stochastic forward passes per test subject provided the uncertainty estimates used in these robustness checks.
The paper acknowledges that missingness itself may encode social structure, and that extensive edge pruning for highly non-compliant participants can reduce the graph to a sparse state and lose temporal resolution.
A complement to assessment, not a substitute
The paper positions SATURN as a framework for digital cognitive phenotyping that highlights macro-level behavioral anomalies, and the numbers support a narrow version of that claim. The incremental wearable signal over a strong demographic benchmark is about 0.04 in R², and the ordering of graph over sequential over flattened models held consistently across folds.
Several boundaries limit where that ordering transfers. The cohort is United States adolescents in the ABCD Study at the two-year follow-up, so other countries, age ranges, wearable devices, and collection protocols require external validation. The inputs are tabulated daily summaries rather than raw minute-level streams, so within-day activity timing, bout structure, and fine-grained sleep-stage dynamics are unavailable to the model. The target is uncorrected Gc only; fluid and total cognition were not evaluated.
For a reader deciding whether wearable traces are worth modeling for cognition, the paper's specific lesson is that the comparison baseline matters more than the architecture headline. A model reporting R² near 0.28 looks different once the covariate-only benchmark of 0.2346 is on the same table, and the defensible claim is a reproducible behavioral increment layered on strong social structure.
Sources used in this answer
Temporal Graph Learning of Wearable Actigraphy and Sleep Traces for Modelling Adolescent Crystallized Intelligence
SATURN, a 21-node temporal graph model with dynamic pruning of invalid-day edges, reached R² = 0.2783 ± 0.0127 for adolescent crystallized intelligence from 21-day Fitbit traces in 6,091 ABCD participants, above Gradient Boosting (0.2372) and BiLSTM (0.2688) but only modestly above a covariate-only Lasso baseline (0.2346).