Handling missing data in digital biomarker validation trials has shifted from a technical afterthought to a primary determinant of regulatory credibility. When participants wear wrist sensors for weeks, gaps happen: charging breaks, skin irritation, accidental submersion, or simply forgetting to put the device back on. If those gaps are imputed poorly, the entire endpoint analysis can drift into biased conclusions. This article walks through a step-by-step strategy to impute wearables gaps without biasing endpoint analysis, built around modern missingness typologies, transparent thresholds, and sensitivity testing.
Why Missing Wearable Data Is a Scientific Decision, Not a Cleanup Task
Missing data from wearables is often dismissed as a data-quality issue. In reality, the pattern of missingness itself carries physiological and behavioral information. A participant who removes a sensor during sleep may differ from one who removes it during high-intensity exercise. If the missingness mechanism is related to the outcome being measured — say, activity-derived heart rate variability or sleep fragmentation — then simply ignoring the gaps or carrying forward the last known value can silently undermine the endpoint analysis.
Digital biomarker validation trials are especially vulnerable because they depend on continuous, high-resolution data streams to establish measurement validity against a reference standard. A gap imputation method that works for sparse questionnaire data may be wholly inadequate for minute-level accelerometer and photoplethysmography signals. Therefore, the process must be pre-specified, reproducible, and aligned with the intended regulatory submission or clinical claim.
Step 1: Classify Missingness Before You Impute Anything
Before applying any imputation technique, you need to know why the data are missing. In wearables, three mechanisms typically apply:
- Missing completely at random (MCAR): The gap is unrelated to both observed and unobserved data. For example, a device malfunctions due to a random firmware bug, or a participant removes the sensor for a scheduled shower.
- Missing at random (MAR): The probability of missingness related to observed data, but not to the missing value itself. For instance, a participant may be more likely to remove the device during weekends because they are at home, and weekend activity levels are already recorded from previous days.
- Missing not at random (MNAR): The missingness is directly related to the unobserved value. This occurs when the sensor stops recording because the wearer is sweating heavily or moving vigorously — exactly the physiological state the trial is trying to capture.
This classification must be documented with evidence from device event logs, participant diaries, and contact logs. Without this step, later imputation decisions become guesswork. For validation trials, it is also advisable to report the percentage of missingness per participant and per time window, because high missingness can turn even the best imputation model into a false-confidence exercise.
Step 2: Define the Endpoint Window and Tolerable Missingness Threshold
Endpoint analysis in digital biomarker trials usually depends on averaging or aggregating data over a specific window, such as nocturnal heart rate, daily step count, or sleep onset latency. For each endpoint, define the minimum amount of raw data needed to calculate a reliable value. A common approach is the 70% wear-time rule, where a day is considered valid if the participant wore the device for at least 16 of 24 hours. But this threshold should be based on the endpoint’s stability characteristics, not a generic habit.
For example, if the endpoint is a night-time circadian feature, the relevant window may be only 8 hours, and a 70% threshold within that window is different from a full-day threshold. You should also specify what happens when a participant fails to meet the minimum threshold: will the entire day be excluded, or will imputation fill the gap? The answer should be written in the statistical analysis plan before unblinded data are reviewed.
This step prevents a common bias: selectively deciding which gaps are imputable after seeing the resulting endpoint distributions.
Step 3: Build a Missingness Model From Auxiliary Sensor Data
Once the missingness mechanism is classified and the thresholds are set, the imputation model should use all relevant auxiliary data. In wearables trials, that includes:
- Accelerometer-derived activity counts from adjacent time windows
- Heart rate or heart rate variability data from the same time of day on other days
- Skin temperature or electrodermal activity where available
- Participant-reported sleep times, mood scores, or symptom diaries
- Demographic variables such as age, sex, and BMI
Multiple imputation by chained equations (MICE) is a practical choice because it can handle the multivariate, longitudinal structure of wearable data. For each missing block, generate a small set of plausible values based on regression models that include the auxiliary variables and the participant-specific random effects. Avoid using the final endpoint value as a predictor if that endpoint is derived from the same missing variable; this can create circularity and over-optimistic precision.
For continuous physiological variables that show strong diurnal rhythms, include time-of-day terms and interaction terms with weekday versus weekend status. Predictive mean matching is a useful option because it preserves the natural distribution of the observed data, rather than forcing values into a normal distribution that does not fit skewed step-count or activity intensity data.
Step 4: Generate Multiple Imputation Sets and Perform Sensitivity Analysis
A single imputed dataset treats imputed values as known, which artificially reduces variance and can make p-values look more significant than they should. Instead, generate multiple imputed datasets — typically 20 to 50 for high-dimensional wearable data — and combine the results using Rubin’s rules. The combined estimates account for the uncertainty introduced by the missingness and the imputation process.
However, no imputation model is perfect. A robust validation trial must include sensitivity analyses that test different assumptions. A practical approach is to create a set of plausible imputation scenarios:
- Complete-case analysis: Only participants with no gaps in the endpoint window are included.
- MAR-based multiple imputation: The primary analysis, using the auxiliary model described above.
- MNAR-based tipping-point analysis: Re-run the imputation with shifted values — for example, assume the missing heart rate was 10 beats per minute higher than predicted — to see how much the endpoint estimate has to change before the conclusion flips.
If the tipping point is very close to the primary estimate, the endpoint is fragile. If the conclusion remains stable under a broad range of realistic MNAR adjustments, you have strong evidence that the missing data strategy did not bias the endpoint analysis.
Step 5: Validate the Imputation With a Simulated Wear Gap
The best way to test an imputation pipeline is to create artificial gaps in data that were actually observed. This is sometimes called a “masking” or “negative control” exercise. Take a subset of participants with complete data during a high-quality recording period, artificially remove 10%, 20%, and 30% of the data using patterns similar to real wear gaps, and then run the imputation and endpoint analysis. Compare the imputed endpoints to the true endpoints from the complete data.
This simulation gives you a direct estimate of imputation error and bias under realistic conditions. It also helps calibrate the tolerable missingness threshold. If the endpoint estimate remains within an acceptable bias margin when 20% of data are masked, but diverges at 30%, then a 30% missingness threshold should be formally rejected. Document these simulation results as part of the trial’s validation report. Regulators and clinical reviewers increasingly expect this kind of empirical justification rather than a generic imputation rationale.
This step also provides a practical opportunity to compare alternative imputation methods. A simple last-observation-carried-forward method will perform poorly for high-frequency vital signs, but you can quantify exactly how poorly. That objective evidence is far more persuasive than a theoretical argument about missingness mechanisms.
Transparent Reporting Is the Ultimate Bias Guard
Even the most sophisticated imputation pipeline will fail if it is not described clearly in the study protocol and final report. For every endpoint, document the number and percentage of participants with missing data, the missingness mechanism assumed, the auxiliary variables used, the number of imputation sets, and the results of sensitivity analyses. Use flow diagrams and tables to show how the sample size changes after each missing-data decision.
In 2026, the digital biomarker field is moving toward frameworks that require not just analytical correctness but auditability. A reviewer should be able to trace every imputed value back to the model that generated it. That level of transparency protects the trial from accusations of data fishing and makes the endpoint analysis reproducible by another team with the same raw data.
Conclusion
Missing data in wearable-based validation trials cannot be eliminated, but it can be handled with a disciplined, pre-specified strategy. By classifying missingness, setting meaningful endpoint thresholds, building an auxiliary-data-driven imputation model, combining multiple imputation sets, and validating the pipeline with simulated gaps, you can impute wearables gaps without biasing endpoint analysis. The goal is not to create perfect continuous data, but to ensure that the missingness process does not silently shape the conclusion.
