We validated the step-counting algorithm with older adults in the lab (using videorecordings as ground truth) and in 2 free-living studies (using the Fitbit as ground truth) lasting 2 and 12 days.

Our step-counting Amulet app estimates the number of steps taken over the course of a day and displays the information on the Amulet screen, similar to the functionality of a pedometer or of other wearable devices (Figure 1). The app continuously estimates the number of steps for each 5-second window, updating the count viewed by the user. It uses a 2-step process: machine learning is used to determine if physical activity occurred in the most recent 5 seconds of data, and then, the number of steps is estimated by counting the number of peaks.

We used a linear support vector machine that classifies each 5-second window of accelerometer data into low, moderate, or vigorous activity [16]. We trained the algorithm on data collected from older adults who performed various physical activities: sitting, standing, lying down, walking, and running [16,19]. Our evaluation of the algorithm produced an accuracy of 91.7% using leave-one-subject-out cross-validation. If the output of the algorithm is moderate or vigorous, the step-detection algorithm is run to determine the number of steps in the 5-second window. This 2-phase approach was necessary to reduce false positives by preventing various actions such as random hand motions from being counted as steps.

The step-detection algorithm estimates the number of steps in 5 seconds of acceleration data. The algorithm uses the magnitude of the acceleration of the 3-axis accelerometer of the Amulet. It is a 3-stage process consisting of zero-meaning, filtering, and peak detection, using an approach similar to that described in [13]. First, to ensure the data have a mean of zero, for each sample, we subtract the average of the preceding 20 data samples. Subsequently, a moving-average filter is applied, that is, each sample is replaced with the average of the 3 preceding samples. Finally, peaks in the filtered signal are identified by checking for change of slope. If the slope changes from positive to negative, and the peak value is above a certain threshold, then the peak is counted as a step. The cut-off threshold was initially empirically determined and then tuned.

Studies were approved by the Committee for the Protection of Human Subjects at Dartmouth College and the Dartmouth-Hitchcock Institutional Review Board (28905). All participants provided signed informed consent.

Participants were recruited through the Center for Health and Aging at Dartmouth and primary care practices at Dartmouth-Hitchcock using flyers, listservs, and posters. This was a convenience sample; our results may not necessarily be applicable to other groups.

Data for the development and evaluation of the step-detection algorithm were collected at our Dartmouth campus laboratory. Older adults (n=20, age ≥65 years) were asked to perform different types of walking (normal, fast, up and down a staircase, and intermittent) while wearing an Amulet. The Amulet collected raw acceleration data at a frequency of 20 Hz and logged the magnitude, which we later used to develop the step algorithm. The participants were videotaped. The video was independently reviewed to count steps by 2 individuals independently (JAB, RKM) and any differences were later reconciled. We used these data for the preliminary development of the step-detection algorithm and evaluated the algorithm using the error rate (the percentage difference between the algorithm’s estimated step count and the ground-truth step count measured from the video).

We conducted a 2-day field study in which older adults (n=7, age ≥65 years) wore an Amulet device (running our step counter app) and a Fitbit Flex 2 device (Fitbit Inc) on the same wrist for 2 days. We compared each participant’s step count estimated by their Amulet (exploring 5 different cut-off values) with their step count reported by the Fitbit (downloaded from the individual’s research-based Fitbit account). The error rate between Fitbit’s steps and Amulet’s steps for each of the 5 peak cut-off values was computed.

We conducted a field study (2 cohorts, 12 weeks each) to test the step-detection algorithm with the target population—older adults with obesity. This study was a subset of a larger study that evaluated the impact of a combined weight loss and exercise intervention in older adults with obesity [20]. The goal of this analysis was to compare the Amulet’s step-count estimate with the Fitbit’s step count over a long period in real-world settings. Participants (Table 1) from both rounds wore an Amulet and Fitbit on the same wrist for 12 weeks. The Amulet logged the summary steps on an SD memory card hourly and at the end of each day. A research assistant copied the data from each participant’s SD card on a biweekly basis. The Fitbit logged a summary of each day’s step count (computed by a proprietary algorithm) to the Fitbit app on the participant’s Android tablet, which uploaded the data to the Fitbit servers; we subsequently used the Fitbit research API to download participants’ data. After 10 weeks of monitoring data from the first cohort, we modified the step-detection peak cut-off value to minimize the error rate relative to ground-truth step count data from the Fitbit.

We calculated the mean with standard deviation for continuous measures and count with percentage for categorical measures. Although participants were instructed to wear the Amulet and Fitbit simultaneously, not all participants did so the entire time. We thus limited the data to days for which it was safe to assume that they were worn for nearly the same amount of time. The Amulet was able to detect if it was worn each hour, so we considered the Amulet to be worn if the Amulet was worn for 75% of a 15-hour day (675 minutes); the Fitbit only reported daily step count, so we considered the Fitbit to be worn if the step count was greater than 100. We selected these parameters from an understanding of the distribution of wear time over the course of a 24-hour day and the distribution of steps expected for this population per day [21]. Comparisons between Amulet and Fitbit were limited to data from days for which both were worn. We conducted bivariate linear regression to compare the association between Amulet and video-counted steps during various activities (laboratory study) and between Amulet and Fitbit steps (2-day and 12-week field studies). We compared Amulet steps to Fitbit steps using percentage difference (difference between Amulet and Fitbit steps divided by Fitbit steps).We used Bland-Altman plots to compare the agreement between Amulet and Fitbit steps. For our analysis, we defined significance as P<.05.

There was a strong linear association when participants walked normally (Figure 2; Table 2). For normal walking, the Amulet step-detection algorithm undermeasured the number of steps taken by an average of 6.7 steps (SD 32.6). Across all activities, the Amulet was on average 3.2 steps lower (SD 25.9) or 2.1% (SD 31.9%) lower than video-validated steps (Figure 2). The distribution was slightly positively skewed.

We discarded 1 participant’s data because the data indicated the devices had not been worn much (step counts were less than 350 per day). The associations between each participant’s daily step counts reported by the Amulet and Fitbit were high for all cut-off thresholds (Figure 3). The third cut-off threshold had the highest association between Amulet and Fitbit steps (R²=0.989). Cut-off threshold number 2 had the smallest mean percentage difference between Amulet and Fitbit (–3.1%, SD 25.1) of all cut-off thresholds (threshold 1: mean 15.27%, SD 33.19; threshold 3: mean –10.77, SD 23.43; threshold 4: mean 7.18, SD 29.28; threshold 5: mean –4.11, SD 24.94) (Figure 4).

Across both 12-day field studies, there were 297 participant-days for which both the Fitbit and Amulet had been worn. Cohort 2 used the modified app. For the first cut-off threshold (version 1), we recorded 86 participant-days; the average Fitbit step count was 5797 steps (SD 3296), and the average Amulet step count was 9780 steps (SD 3719). For the second cut-off threshold (version 2), we recorded 211 participant-days; the average Fitbit step count was 6415 per day (SD 3751), and the average Amulet step count was 7956 per day (SD 3324). The association between Fitbit steps and Amulet steps improved (first cut-off threshold: R²=0.386; second cut-off threshold: R²=0.669) (Figure 5). There was improved agreement between both measures with the second cut-off threshold (Figure 6). There were 9 observations by 5 unique participants with differences 2 standard deviations higher than the combined difference mean. These observations had an average Fitbit and Amulet step count of 1373 (SD 1988) and 10,689 (SD 1971), which suggested that participants may have taken their Fitbits off at some point during the day before removing their Amulet.