Wearable data moves past prediction: swapping sedentary time for exercise yields sleep-mediated intervention contrasts
Synopsis
Wearable studies of student exercise, sleep and distress have mostly predicted associations, not what happens to distress when sedentary time is replaced by exercise. CISM writes such changes as reallocation operators that preserve each student's daily waking-time budget, generates next-night sleep and later distress from one temporally ordered stochastic response model, and estimates population and individualized policy contrasts with g-computation and sequentially augmented estimators. In the structural causal-recovery setting, full CISM reaches policy MAE 0.096, PEHE 0.129 and 94.1% interval coverage, while removing the time budget raises policy MAE to 0.151.
Interpretation
The exercise–sleep–mental-health question becomes a computable intervention problem: reallocate sedentary minutes to moderate activity while holding each student's available waking time fixed, then ask how later distress would change. Earlier wearable work mainly estimated associations between sensed behavioral variables and mental-health ratings, using gradient-boosted trees, sequence models or missingness-aware encoders; predictive accuracy and attention explanations do not supply a counterfactual estimand. In a structural simulator with confounder feedback and an activity/sleep budget, full CISM reaches policy MAE 0.096, PEHE 0.129 and 94.1% coverage, against 0.146, 0.194 and 89.8% for sequential g-formula and 0.226, 0.285 and 74.6% for naive regression.
One temporally ordered stochastic response model serves factual reconstruction, exercise interventions, sleep-mediator interventions and later distress estimation, rather than several disconnected predictive heads. Conventional regression that conditions on post-exercise sleep can block the mediated pathway, while natural indirect effects require cross-world counterfactual independence and are fragile when treatment-induced mediator–outcome confounding is present. Randomized mediator kernels define interventional mediation, and the split into direct and sleep-mediated parts is an algebraic identity for randomized mediator policies, not a claim to decompose the ordinary natural total effect; in the component ablation full CISM has mediation MAE 0.079, rising to 0.156 without the sleep kernel.
Estimates are reported only where the observational data support them, and the procedure abstains outside empirical support instead of treating prediction as causal conclusion. Large exercise increases have little comparable data in observational cohorts, and recorded exercise is self-selected: symptoms, academic deadlines and prior sleep can influence both today's activity and tomorrow's distress. In the ablation full CISM has 94.1% coverage and excludes 12.4% of candidate interventions by the support restriction, while removing the support filter drops coverage to 84.8% and exclusion to 0.0%; on factual prediction CISM reaches GLOBEM AUC 0.797 and SSAQS MAE 13.94, against 0.783 and 14.67 for TimeMixer.
Perspective
The evaluation covers two student wearable datasets and structural causal-recovery experiments: GLOBEM with 497 participants and 705 participant-years from 2018 to 2021, combining Fitbit activity and sleep with weekly surveys, and SSAQS with one semester of Fitbit Inspire 3 data from undergraduates at two Mexican universities plus daily stress and anxiety self-ratings on a 0–100 scale. Conclusions on real data are limited to chronological factual forecasting and assumption-indexed policy estimates, not verified clinical benefit.
Observational effect estimates remain conditional on identification assumptions including consistency, sequential exchangeability and positivity, which the authors state require sensitivity evaluation. SSAQS provides no depression diagnosis, and its daily stress and anxiety ratings are not interchangeable with GLOBEM's weekly depression-oriented outcomes, so cross-dataset transfer should be assessed only on comparable behavioral representations or within-outcome normalized proxies. Missing-not-at-random cannot be ruled out from the observed data, and SSAQS includes survey-only participants.