Data quality and standardization
Wearable data can be noisy, incomplete, and inconsistent across devices. Even within a single cohort, participants may use different brands or models, each with proprietary filtering and calibration. A SWAS should therefore report device types, firmware/algorithm versions when available, and implement quality control for artifacts, non-wear, and implausible values, alongside harmonization or stratification across devices. Practical lessons and reporting frameworks from digital biomarker development can be used to standardize these steps 15,17.
Contextual dependency
Sensor features are highly context-dependent: a spike in heart rate may reflect exercise, stress, dehydration, medication effects, or measurement error. Disentangling contexts requires auxiliary metadata (activity labels, location, schedule/season, self-report, smartphone logs) and careful temporal alignment. Without context, SWAS can produce spurious associations or miss genuine ones; negative-control outcomes and stratified analyses can help detect residual confounding.
Statistical and computational complexity
SWAS entails multiple testing on a vast scale, often with highly correlated features. Unlike GWAS, there is no meaningful universal significance threshold because SWAS feature libraries are cohort- and device-specific and can vary widely in size and dependence structure. Instead of relying on a GWAS-style threshold by analogy, SWAS should (i) define and version the feature space, and (ii) control and report false discoveries using strategies appropriate for dependence, such as false discovery rate control (Benjamini-Hochberg), q-values, permutation-based thresholds, or knockoff-style procedures. Importantly, broad feature engineering is still a form of hypothesis encoding; SWAS is best described as hypothesis-generating, not hypothesis-free.
SWAS analyses should explicitly address longitudinal structure and time-varying confounding. Within-person correlation can be handled using mixed models or generalized estimating equations18,19, and survival models with time-varying covariates20. Because non-wear and missingness can be informative, SWAS should avoid default listwise deletion and instead use principled approaches such as multiple imputation or inverse-probability weighting21. When causal interpretations are desired, SWAS findings should be paired with explicit causal-inference strategies for time-varying exposures, such as marginal structural models22. Temporal dependencies and drift can be modeled using state-space approaches or recurrent architectures 23,24.
Privacy and ethical considerations
Continuous sensor monitoring captures intimate details about daily life and can reveal habits, locations, and social interactions. Research institutions must implement privacy-by-design (data minimization, de-identification, secure enclaves) and ensure that data use aligns with meaningful, revocable consent and governance models, including emerging approaches to citizen data sovereignty25. Beyond privacy, SWAS must address representativeness and equity: device ownership, digital literacy, and connectivity vary by age, socioeconomic status, and geography, creating biased sampling and unequal benefit distribution. Commercial platform dependence is another core risk: access to raw data may be restricted, and vendor algorithms or firmware can change without notice, altering feature meaning over time. SWAS reports should therefore document device and algorithm versions, quantify differential missingness, and include fairness-aware analyses and community-informed governance where possible.