Why wear time decides what your data can say
Every wearable study has gaps: devices come off for charging, showers and sleep, sensors lose skin contact, and some participants stop wearing them. How you detect non-wear, define a "valid" day and handle what's missing can change your results as much as the sensor itself. These rules need to be written down before data collection, in the protocol and statistical analysis plan, not chosen after you see the data.
Detecting non-wear
| Method | How it works | Source |
|---|---|---|
| Troiano (count-based, NHANES) | Non-wear = at least 60 min of 0 counts/min, allowing up to 2 min of interruptions | As described in Migueles et al., Sports Med 2017 |
| Choi (count-based) | Zero-count threshold, a 90-min window of consecutive zero counts, and an allowance of 2 min of nonzero counts if there's a 30-min zero-count window upstream or downstream. Validated against 24-h room-calorimeter stays in 49 adults and 76 youth | Choi et al., Med Sci Sports Exerc 2011 |
| Device- or signal-reported status | Many consumer devices and APIs report wear, sensor or signal status directly (for example a skin-contact or signal-quality flag). Treat it as an input to your rule, document the vendor definition, and check it against a sample with known wear | Vendor documentation for your device |
Count-based algorithms were developed for research accelerometers (ActiGraph counts). For PPG, ECG or temperature-based wearables, the same logic applies, but thresholds must be derived and validated for your sensor and placement.
Valid-day and valid-week rules
| Rule | Common threshold | Source |
|---|---|---|
| Valid day (accelerometer) | ≥10 h/day of wear time | Practical considerations table in Migueles et al., Sports Med 2017 (ActiGraph GT3X/+, 24-h protocol) |
| Valid week (accelerometer) | ≥4 valid days | Migueles et al., Sports Med 2017 |
| Protocol matters | In NHANES 2003–2006 (waking-hours wear), only 40–70% of participants achieved ≥10 h/day for ≥6 days; in NHANES 2011–2012 (24-h wear), wear time was 21–22 h/day for ≥6 days | Migueles et al., Sports Med 2017 |
| Valid CGM period | 14 days of wear with ~70% of possible data | Battelino et al., Diabetes Care 2019; see our CGM data guide |
Stricter rules give more reliable per-person estimates, but they exclude more participants. Migueles et al., Sports Med 2017 reports that raising valid-day and valid-week requirements reduced sample size and statistical power. Pre-specify the rule, and report how many participants and days each rule keeps.
Adherence-reporting checklist
- Protocol: wear location, 24-h vs. waking-hours instructions, charging schedule, and planned monitoring period.
- Non-wear definition: algorithm name, parameters and version, and any device-reported status used.
- Validity rules: valid-day and valid-period thresholds, prespecified in the SAP.
- Wear time per participant: median (IQR) hours/day and valid days, overall and by key subgroups (age, sex, skin tone, study arm).
- Flow: how many enrolled participants, days and windows were excluded at each step, and why.
- Pattern of missingness: by time of day, day of study and participant characteristics. Missingness that differs by group is a bias warning.
- Comparison: characteristics of included vs. excluded participants.
Handling missing wearable data
| Approach | What it assumes | When it's biased |
|---|---|---|
| Complete-case (drop incomplete days or people) | Data are missing completely at random | Can be biased when data are missing at random but not completely at random, and loses precision and power as exclusions add up (Sterne et al., BMJ 2009) |
| Multiple imputation under MAR | Missingness depends only on observed data (for example time of day, activity before the gap, participant traits) | Observed data can't tell MAR from missing-not-at-random. Bias from MNAR (for example, people removing the device when unwell) can only be addressed by sensitivity analysis (Sterne et al., BMJ 2009) |
| Personal-baseline approaches | Each person's own earlier data define "normal", and analysis looks for change from that baseline | Needs enough clean baseline data. Gaps that coincide with the change you're looking for can hide or mimic it, so report data coverage around any detected change |
For confirmatory trials, FDA E9(R1) guidance (May 2021) asks sponsors to state the assumptions behind the main analysis and prespecify sensitivity analyses that test how robust conclusions are to departures from them. Missing-data handling is part of that.
From our research
Long-run, person-level wearable data are only as useful as their continuity. Mike Hittle is a co-author on a case report in JACC: Case Reports (Lautman Z, Gravellier L, Malik S, Hittle M, Haddad F, Ræder H, Snyder MP. "Longitudinal Changes in Wearable-Captured Physiology Before Sudden Cardiac Death," 2026, doi:10.1016/j.jaccas.2026.110397). In the authors' words: "Change-point analysis of wearable-captured physiology revealed a coordinated step-change, with most changes clustering 4-6 months before the sudden cardiac death," and "The key signal was not any single abnormal value but a coordinated, sustained step-change from the patient's own baseline that preceded imaging."
This is a single-patient case report. We quote it to show why personal baselines and data continuity matter, not as evidence that wearables predict cardiac events.
Related
Frequently asked questions
How many hours a day counts as a valid wear day?
A common accelerometer rule is at least 10 hours of wear per day, with at least 4 valid days per week (Migueles et al., Sports Med 2017). The right threshold depends on your device, protocol and outcome. Prespecify it and report how many participants it excludes.
What is the difference between non-wear detection and a valid-day rule?
Non-wear detection decides, minute by minute, whether the device was being worn. A valid-day rule then decides whether a day has enough wear time to be included in analysis.
Should I impute missing wearable data or drop incomplete days?
Dropping incomplete data can bias results unless data are missing completely at random. Multiple imputation can reduce that bias if data are missing at random, but neither fixes data that are missing for reasons tied to the outcome. Plan sensitivity analyses (Sterne et al., BMJ 2009).
Scope it in 30 minutes
Book a free 30-minute scoping call and we'll name the smallest engagement that moves you forward.
Book a free 30-min scoping callPrefer to write? Send a short note.