JMIR Publications

ABSTRACT

Background:

No studies to date have evaluated the accuracy of consumer wearable activity monitors for step counting or activity classification in individuals with lung cancer. This study addressed that gap by assessing the criterion validity of the Fitbit Charge 6 against video-recorded direct observation measures in adults with lung cancer under controlled laboratory conditions.

Objective:

The primary objective was to describe the step-count agreement between the Fitbit Charge 6 and video-recorded direct observation across a range of walking bout durations and gait speeds. Additionally, we compared the Fitbit's accuracy in identifying active vs. sedentary minutes and descriptively analyzed the false-positive spurious step-count detection rate across a variety of activities.

Methods:

Fourteen adults diagnosed with stage I-IV lung cancer completed a cross-sectional, in-laboratory validation study at The Ohio State University Wexner Medical Center. Participants wore the Fitbit Charge 6 on the non-dominant wrist while completing variable-duration walking trials (5-, 15-, and 30-seconds), self-selected gait speed trials across eight progressively faster speeds, and standing, sitting, lying, and fidgeting tasks. All activities were video-recorded and coded at 1-second resolution. Step count agreement was evaluated using mean absolute percentage error (MAPE), intraclass correlation coefficients (ICC), and repeated-measures Bland-Altman analyses. Minute-level activity classification was assessed via a pooled confusion matrix using a ≥30-second majority-rule active-minute threshold. Spurious step detection was descriptively analyzed across non-walking minutes, stratified by fidgeting status.

Results:

Across 126 walking trials, the Fitbit showed poor agreement with video-coded step counts (ICC = 0.21; MAPE = 28.5%). Accuracy was duration-dependent: 5-second trials yielded a MAPE of 50.6% (ICC = 0.00), while 30-second trials showed a smaller mean bias (MAPE of 18.7%; ICC = 0.34) with persistent variability. Across 111 gait speed trials, the Fitbit significantly undercounted steps at gait speeds below 0.6 m/s (MAPE = 56.6%), with the lowest error near 1.2 m/s; a quadratic mixed-effects model confirmed a nonlinear speed–error relationship (p < .001). For activity classification, sensitivity was high (0.91) but specificity was modest (0.63) and positive predictive value was low (0.31), reflecting frequent misclassification of sedentary minutes as active. Among 267 non-walking minutes, 12.4% contained at least one spurious step, with higher false-positive rates during fidgeting (14.4%) than non-fidgeting (9.0%) periods.

Conclusions:

The Fitbit Charge 6 provides improved step count estimates during sustained walking at moderate speeds but introduces clinically meaningful error during short walking bouts and at the slower gait speeds characteristic of adults with lung cancer. High sensitivity but low specificity for activity classification suggests systematic overestimation of active time. These findings have implications for the design and interpretation of exercise oncology interventions that rely on consumer-wearable-derived endpoints in this population.