Wearable Endpoints: Separating Non-Wear Time From True Inactivity

Choose and document a wear/non-wear classifier together with its version, parameters, input signal, placement, and valid-day rules before collecting endpoint data. Keep the protocol, statistical analysis plan (SAP), and data management specification consistent. Low activity counts alone do not prove device removal: quiet wear and an unworn stationary device can look similar. No published classifier is universally correct, and FDA does not prescribe a 60-minute or 90-minute rule. The practical decision is which supported rule defines this trial’s endpoint and how uncertainty in that decision will be assessed. Choi et al. (2011) FDA DHT guidance, section IV.E

A wear label is an inference from the available signal and the specified method. A removed monitor may record movement when it is carried or nudged; a worn monitor may record extended zeros while its wearer remains still. “Near zero” also needs a numerical definition: a run of exact zero counts is different from a run containing low but nonzero counts. Device output, epoch length, placement, and the surrounding signal all affect the classification. A count-based rule should not be applied directly to raw acceleration expressed in gravitational units. Choi et al. (2011) Zhou et al. (2015)

The 2011 Choi study used 49 adults and 76 youth during monitored room-calorimeter stays. Because monitors remained worn, intervals labeled non-wear exposed false removal classifications. The authors linked much of this error to sedentary behavior and estimated approximately 8% underestimation of sedentary time from the observed waking-time error. This controlled setting helps explain a failure mechanism; it does not establish the accuracy of a classifier for a patient population living at home. A rule can improve one type of error while still missing genuine removal episodes. Choi et al. (2011), methods and discussion

Separate classification from interpretation. First derive the wear mask using the chosen rule. Then derive daily and window-level measurements using that mask. Finally apply the statistical handling specified for unavailable observations. Retaining these intermediate results lets a reviewer locate whether disagreement arose in the signal, the classifier, the eligibility threshold, or the final summary.

The table summarizes published methods rather than recommending a default. The exact 60-minute parameters and calendar-boundary behavior below come from Choi’s documentation of the NHANES-era algorithm and its SAS implementation. The original Troiano study supplies historical context; its abstract is not evidence for all implementation details. These count-based comparisons use one-minute intervals. Choi et al. (2011), Table 2 and implementation discussion Troiano et al. (2008)

Choi’s 2012 study involved 29 adults aged 76–96, with wrist monitoring for approximately 24 hours per day and waist monitoring during waking hours for about seven days. Vector magnitude performed better than vertical-axis counts for wrist classification, and 90-minute windows improved the tested algorithms. Choi and Troiano methods outperformed the manufacturer’s ActiLife algorithm in that comparison. The reference relied on diaries and data plots; this does not establish that current vendor software is inaccurate or that the same settings are validated for every device or disease. Choi et al. (2012)

Zhou and colleagues discuss acceleration-only approaches and evaluate a method combining raw acceleration with skin-surface temperature. Such extra measurements may help resolve ambiguity, but the device must actually collect the required signal under the tested conditions. Separately, Syed and colleagues’ 2021 CNN method reported very high performance on its gold-standard dataset using patterns learned from a hip-worn monitor. That is evidence about a specific research method, not proof of fitness for this trial’s endpoint or endorsement by a regulator. Zhou et al. (2015) Syed et al. (2021)

Vanhelst and colleagues compared seven rules against non-wear diaries in 77 healthy participants aged 10–17 during four days of waist-worn monitoring. Wear-time estimates varied by 17%, sedentary-time estimates by 30%, and moderate-to-vigorous physical activity (MVPA) by 3.7%. The percentages meeting the 10-hour valid-day criterion also depended on the rule; the 10-minute and 60-minute continuous-zero rules differed significantly from the diary. These are study-specific differences, not expected treatment-effect bias or universal error rates. Vanhelst et al. (2019), results

The same study found strong sedentary-time agreement with the diary for several methods, including Choi, Troiano, and the 30- and 60-minute rules. MVPA estimates were much less affected than sedentary estimates. This counter-evidence matters: classifier choice can be consequential without changing every endpoint equally. Prioritize assessment of the low-activity observations and threshold crossings relevant to the planned measurement, rather than assuming the largest published percentage will apply to a trial.

Another frequently cited comparison needs its own qualifier. Choi 2011 Table 3 reports mean adult waking-time misclassification of 49 minutes for the current algorithm and 9 minutes for the revised algorithm, but that table specifies a 60-minute window for the revised comparison. It should not be described as the isolated effect of switching from 60 to 90 minutes. The paper separately recommends the 90-minute window together with revised interruption handling. Choi et al. (2011), Table 3

A coherent threshold example shows why the distinction matters. Suppose a participant truly wears a count-based monitor for 10 hours 15 minutes, including an isolated 75-minute zero-count period while seated. Assume activity outside that period prevents either algorithm from extending it. A 60-minute rule labels the 75 minutes non-wear and credits nine hours; a 90-minute rule retains that interval and credits 10 hours 15 minutes. With a 10-hour minimum, the day fails under the first rule and passes under the second. If there are exactly three other valid days, the seven-day window has three versus four qualifying days. This is hypothetical arithmetic, not evidence that the 90-minute label is correct whenever the actual wearing status is unknown.

For a window average, specify whether eligible days receive equal weight and how partial-day coverage affects the intended variable. A missing daily observation is not a measured day with zero steps. Nor does crossing a valid-day threshold validate the step-detection algorithm itself. The classifier, day rule, averaging rule, and step measurement each answer a different part of the derivation question.

FDA’s December 2023 DHT guidance contains nonbinding recommendations. Section IV.E recommends prespecifying endpoint definitions and source data in the SAP, planning to reduce and address missing data, and using the ICH E9(R1) estimand framework in late-phase trials. Illustrative options include monitoring and alerts, reminders, run-in periods, and investigator outreach. Appendix B’s sleep example describes repeated-assessment derivation, population-level summaries, and missingness within a day, a whole day, or a week. FDA DHT guidance, IV.E and Appendix B

Applying that framework to wearables supports an explicit classifier and valid-window specification. This application is the article’s methodological recommendation. FDA names no preferred wear-time algorithm or minimum number of valid days. Its population-specificity example concerns step capture in healthy people versus people with Parkinson’s disease and a shuffling gait; it does not show that a particular non-wear rule fails in heart failure. FDA DHT guidance, IV.C and IV.E

For the broader evidence supporting a device and its intended measurement, see the existing guide to DHT fit-for-purpose evidence. Here the narrower deliverable is the reproducible rule that turns recorded intervals into usable endpoint days. DHT endpoint evidence guide

ICH E6(R3) Annex 2 reached Step 4 on 3 June 2026; implementation follows jurisdiction-specific schedules. Sections 3.2.1 and 3.2.3 recommend describing and justifying design elements and data sources, and discussing relevant data variability in the protocol or related documents such as the SAP. Section 3.8 addresses sponsor oversight of trial conduct and service providers. ICH E6(R3) Annex 2, 3.2 and 3.8

Scope matters. Sections 3.4.1 and 3.5.1(c) discuss processing oversight and systematic resolution of linked-data inconsistencies in the RWD context. Section 3.5.2 says some RWD considerations may also apply to remote trial collection. These passages do not prescribe wearable classifier parameters or make every RWD provision automatically applicable to a prospectively collected sensor stream. A trial team can use them to inform its documented reconciliation and oversight approach while explaining the actual data sources and applicable framework. ICH E6(R3) Annex 2, 3.4.1 and 3.5

ICH E9(R1) distinguishes events affecting the interpretation or existence of relevant measurements from missing data. Neither study withdrawal nor other reasons for missing data are in themselves intercurrent events. Apply that distinction to the clinical question: non-wear generally leaves activity observations unavailable; stopping an assessment device does not automatically become an intercurrent event because the gap is permanent. Record why collection stopped, and evaluate any underlying event, such as treatment discontinuation, against the prespecified estimand. If device use is itself part of the assigned treatment, its discontinuation may require a different assessment. ICH E9(R1), A.3.1 and A.4

When intended observations are unavailable, the SAP should address the relevant missing-data assumptions. ICH E9(R1) sensitivity analyses target the same estimand. Re-deriving an endpoint with another classifier can examine robustness, but changing the variable or day-eligibility definition may instead produce a supplementary analysis. Label the purpose and limitations explicitly; an alternate rule is not a universally mandatory safeguard and agreement between two rules does not prove the treatment effect. ICH E9(R1), A.5.2 and A.5.3

A device-removal flag should therefore carry an operational reason separately from the analytical outcome. “Battery depleted,” “participant removed device,” and “reason unknown” describe collection conditions. Whether an unavailable period produces an invalid day or a missing visit measurement follows the endpoint specification. Whether a related clinical event needs an estimand strategy is a separate judgment. Do not fill a missing period with zeros merely because the recorded stream is quiet.

The general frameworks are covered in the site’s estimand and sensitivity-analysis guides; use them to specify the treatment-effect question before deciding how absent wearable observations enter its analysis. ICH E9(R1) estimand guide Missing-data sensitivity analysis guide

The following patterns assume one-minute activity counts and the published rule settings above. They describe classifier outputs, not verified physical states. Raw-axis variability cannot be inferred from a count stream, and sleep should not automatically be equated with removal. Apply the study’s wearing schedule and endpoint interval separately.

Return to the hypothetical heart-failure trial. Assume the three qualifying weekdays each have 3,650 steps, and the other two weekdays have only six hours of wear and are excluded. Weekend recording covers nearly the entire day, but classified wear is below two hours after excluding the long zero-count intervals and their allowed spikes. The example’s summary uses an equally weighted mean across valid days, conditional on at least four valid days.

The defective branch is an invented implementation failure, not an observed vendor behavior and not the expected difference between the published 60- and 90-minute algorithms. A few isolated spikes cannot explain many hours of extra classified wear when the remaining zero stretches still qualify as non-wear. A central reviewer should compare the saved interval mask, daily classified wear, eligibility flags, and averaging denominator to locate the discrepancy. The apparent 2,228-step result must not be interpreted as a valid week-12 activity measurement.

Authority to change the pipeline should be assigned in the trial’s governance plan. A useful division of work is for data management to document the discrepancy, the endpoint statistician to assess derivation and analysis implications, and the sponsor’s designated owners to authorize a controlled correction. This is a proposed workflow, not a regulator-mandated job-title sequence. Preserve the original output, the reason for correction, the approved specification, and the reproducible rerun. Assess all affected records consistently rather than rescuing selected participants with a different rule.

The failure case is a team manually changing weekend flags because the low average or missing week is inconvenient, without checking what the prespecified rule actually returns. That creates an unreproducible derivation. The cited methods papers and guidance do not establish specific inspection sanctions for this hypothetical. Investigate the discrepancy and explain any change and its analytical impact.

Use the following as a practical specification package. It translates the guidance and methods into reviewable implementation choices; the sample thresholds are not regulatory defaults. Specify the rule before endpoint collection where feasible, confirm execution before using study summaries, and manage later corrections with documented reasons. An implementation error can require correction; calling the rule “locked” should not prevent a justified, consistently applied fix.

1. Protocol: state the measurement objective, device placement, wearing schedule, monitoring window, and rationale for valid-day and valid-window thresholds. Keep availability criteria distinct from the treatment-effect question.

2. SAP: specify daily-to-window aggregation, how unavailable observations enter the analysis, and which assumptions each robustness analysis investigates. Explain whether another classifier preserves the estimand or defines a supplementary result.

3. Derivation specification: record algorithm and software versions, axis or vector-magnitude input, units, epoch length, zero threshold, minimum interval, permitted interruption, bounding windows, timestamps, and calendar-boundary behavior.

4. Implementation checks: compare output with independently expected classifications for short and long zero runs, tolerated and intolerable spikes, intervals across midnight, and days just above and below the eligibility threshold. A menu label is insufficient evidence of the implemented parameters.

5. Review records: retain the input data, interval-level wear mask, daily wear totals, eligibility flags, window denominator, configuration, execution records, and any authorized correction or rerun needed to reconstruct the result.

For each potentially missing window, a compact review record should show the number of qualifying days, which days failed, the relevant reason information, the algorithm configuration, and the prescribed analytical disposition. Participant reminders or site outreach can reduce future gaps; they cannot retrospectively establish that an ambiguous interval was worn. Keep explanations from diaries or sites attributable and apply the planned reconciliation rule consistently.

This article synthesizes public guidance and published methods; it reports no new dataset, vendor test, or patient-population validation. The older count-based studies and newer raw-data work are evidence to evaluate, not interchangeable implementations. Reassess the specification when the device, firmware, software, placement, population, or endpoint changes. The practical aim is a transparent derivation whose uncertainty can be examined, rather than a claim that any classifier reveals the participant’s physical state with certainty. SAP endpoint-definition guide