The Ultrahuman Ring's Updated Step Algorithm Misses Fewer Steps and Counts Fewer False Ones

Summary

• After members reported that the Ultrahuman Ring missed steps, Ultrahuman updated its step algorithm and measured what changed in members' days.

• After the update, 861 members who had reported undercounting saw a median 15.2% more daily steps (95% confidence interval (CI) 11.4 to 17.7), while members who did not receive the update changed by −0.7%.

• False steps recorded while resting fell by 21% in 774 members, a change of −7.1 per resting hour against members who did not receive the update (95% CI −11.3 to −2.9).

• In 121 structured sessions with 21 adults, the ring's median error against a hand count was 6.3% (95% CI 4.9 to 11.1) across walking, running, stairs, chores and gym work. It recorded no steps at rest or at a desk.

• The structured sessions measure accuracy against a hand count. Members' days have no hand count, so the member results show what members will see.

Background and Rationale

More daily steps are associated with lower all-cause mortality [1]. A step counter can be wrong in two ways. It can miss steps that were taken, or record steps that were not. The second error is easy to overlook, because an inflated count looks like an ordinary day.

On a treadmill, the accuracy of ten consumer activity trackers depended on walking speed [2]. Consumer trackers generally count steps accurately in the laboratory, but free-living evidence is sparser [3]. Free-living studies, including one checked against video recordings [4], found that a popular wrist-worn tracker tended to record more steps than were taken [5]. Activity monitors have also recorded false steps during bus and car travel [6]. Expert recommendations call for validation against directly observed steps, in both structured and free-living conditions [7].

Members told Ultrahuman that the ring missed steps they had taken. Ultrahuman updated the ring's step algorithm and released it to members in stages from September 2026. An algorithm that counts more steps could also count movement that is not walking. This paper therefore measures two changes in members' own data: in daily steps, and in false steps recorded while resting. Most false steps are likely to arise while the wearer sits but still moves, for example when commuting by car or bus [6]. Structured sessions compare the algorithm with a hand count during walking and with a known zero during motion without steps. A wrist-worn consumer wearable device, the benchmark device, was worn alongside.

Methods

Figure 1 shows the two tests.

Figure 1. Study design. In structured sessions, each ring reading was compared with the true count, and a wrist-worn consumer wearable device served as a benchmark. In members' data, each member's days after the update were compared with their own two weeks before it, beside randomly sampled members who did not receive the update.

The updated step algorithm

The algorithm detects rhythmic, walking-like movement in the ring's accelerometer signal and credits steps as cadence times walking time. A consistency check requires a steady walking rhythm before steps count, so that motion like vehicle vibration is not counted.

Structured sessions

Twenty-one adult volunteers aged 18 to 50, 13 men and 8 women, completed 121 sessions across 14 activities from 17 to 23 August 2026, during development of the updated algorithm (Table 1). All participants gave informed consent. Each participant wore two Ultrahuman Rings running the same step algorithm with identical settings, both on the non-dominant hand. The benchmark device was worn on the same wrist. In sessions with steps, the true count came from a hand-tally clicker pressed at each heel strike. In sessions with motion but no steps, the true count is zero by design, and no clicker was used.

Table 1. The 14 activities, grouped by condition. Each was performed by at least 5 participants. Eight car or bike sessions were on rough roads; all were kept.

Six data-entry errors were corrected before analysis. Without these corrections, the direction and significance of every comparison were unchanged. Six sessions without a usable hand count were removed. Of the remaining 115 sessions, 111 had readings from both rings and the benchmark device; these are the complete sessions.

Every valid ring reading in an analysed session is one observation of the algorithm, which gives 222 readings on the complete sessions. Paired tests compared per-session errors by the two-sided Wilcoxon signed-rank test. Participant-clustered 95% confidence intervals (CIs) came from 10,000 bootstrap resamples of participants.

Members before and after the update

The member analysis used de-identified data from members whose ring moved directly from the previous to the updated algorithm between 2 and 19 September 2026. They were members who had reported undercounting, and other early-release members, whom Ultrahuman chose to update first. The comparison group was a random sample of members who did not receive the update over the same weeks.

Each member's median daily steps over days 1 to 7 after the update was compared with their median over days −14 to −1. Days under 100 steps counted as days the ring was not worn [8]. Each comparison member was given a random reference day from the updated members' dates. Group medians and means carry 95% CIs from 4,000 bootstrap resamples of members. Activity bands followed published pedometer cut-offs [9].

False steps were counted in resting periods between 08:00 and 22:00, in each member's local time zone. A resting period lasted at least 60 minutes. Throughout it, every heart-rate reading was within 10% of the member's own median daily resting heart rate over days −14 to −1. The ring's accelerometer activity count also had to average above zero, showing the ring was worn and moving a little, and below a walking threshold. That threshold was set so that 90% of half-hours with 150 or more steps sat above it, in comparison members before the update. Steps recorded in resting periods counted as false steps. False steps per resting hour were compared over the same two periods, in members with at least 1 resting hour in each.

Results

Overall error was similar for the ring and the benchmark device

Total absolute error adds up every step missed or added across all readings, as a share of the true steps. Across the 111 complete sessions, it was 17.7% for the ring and 17.4% for the benchmark device (Table 2). The difference, 0.3 points, was not statistically significant (two-sided p = 0.96). The participant-clustered 95% CI for the difference was −4.9 to 6.4 points. The two devices erred in opposite directions. The ring's net error was −7.2%, mostly from missed steps. The benchmark device over-counted in 68 sessions, and its net error of +4.7% was closer to zero than the ring's.

Table 2. Error across the 111 complete sessions, which contained 26,494 true steps: 222 ring readings and 111 sessions of the benchmark device. The first four rows are percentages of true steps; a negative net error means under-counting. The last row is the median across 21 participants of each participant's total absolute error. CIs are participant-clustered. Ring counts were read from the app; the benchmark device's count was the change in its daily total over the session.

In sessions with steps, the benchmark device was closer to the hand count

The 60 sessions with steps, other than cart pushing, covered walking, running, stairs, and chores and gym work. In them, the ring's median error was 6.3% (95% CI 4.9 to 11.1), against 5.6% (95% CI 2.1 to 8.0) for the benchmark device (Figure 2). The benchmark device was closer to the hand count in 40 of 60 sessions (two-sided p = 0.030), although the participant-clustered interval narrowly included zero (−0.4 to 5.4 points). It also had the narrower limits of agreement, the range that holds 95% of differences from the hand count [10]: −55 to +78 steps against −165 to +121 for the ring. A few ring readings fell well below the line of agreement, where the ring missed a large share of a walk.

Figure 2. Steps recorded against the hand count in the 60 sessions with steps, other than cart pushing: 118 ring readings (A) and 60 sessions of the benchmark device (B). Marker shape shows the activity; walking includes walking while holding a phone. The dashed line is perfect agreement. Median error is the median absolute percentage error, a reading's absolute difference from the hand count divided by the hand count; its CIs are participant-clustered.

Chores and gym work had higher error than walking, running and stairs, 34.1% for the ring and 31.6% for the benchmark device. No device detected cart or stroller pushing. In 4 sessions containing 1,320 true steps, the ring recorded no steps and the benchmark device recorded 36.

In sessions without steps, the ring recorded fewer false steps

In sessions without steps, the ring recorded exactly zero steps in 83.8% of 99 readings, against 36.7% of 49 sessions for the benchmark device (Figure 3). The ring was closer to zero in 25 sessions and the benchmark device in 6, with 18 ties (two-sided p = 0.002). At rest and at a desk, the benchmark device recorded 69 steps across 21 sessions, and the ring recorded none. Bike riding was the hardest case for the ring, with 68.9 false steps per reading against 66.7 per session for the benchmark device.

Figure 3. Share of ring readings, and of benchmark device sessions, that recorded exactly zero steps when no steps were taken; higher is better. Rough-road sessions are part of the car and bike row. The activity rows use complete sessions; the header uses every session without steps that had a reading. The p-value is from the two-sided Wilcoxon signed-rank test on per-session errors.

What members will see

Members who had reported undercounting saw more daily steps from the first day after the update, and the rise held for two weeks (Figure 4). In 861 such members, the median change was +15.2% (95% CI +11.4 to +17.7), from a median of 5,464 steps a day before the update to 6,231 after it. Two in three of these members (66%) saw higher counts. In the comparison group of 4,640 randomly sampled members who did not receive the update, the change was −0.7% (95% CI −1.4 to +0.1).

Figure 4. Daily steps by day since the update, as the median member's change from their own median over the 14 days before it. Bold steps are weekly values and the thin line is day by day. Orange: members who had reported undercounting. Grey: randomly sampled members who did not receive the update, over the same calendar days. The header uses days 1 to 7, with 95% CIs from 4,000 bootstrap resamples of members. Members needed at least 7 counted days before the update and 4 after it; the update day is left out.

Other early-release members saw a similar rise, and it was largest for those who walked least (Figure 5). In 3,480 such members, daily steps rose by a median 15.8% (95% CI 14.4 to 17.3), from 5,465 to 6,448. The rise was 25.7% for members who had walked under 5,000 steps a day, 13.4% for 5,000 to 9,999, and 1.1% for 10,000 or more. Randomly sampled members who did not receive the update changed by +4.0%, −3.4% and −6.6% in the same bands. Against them, the rise was 21.7 percentage points larger for members under 5,000 steps a day (95% CI 17.5 to 25.2) and 7.7 points larger for those at 10,000 or more (95% CI 1.5 to 12.8).

Figure 5. How much daily steps changed, grouped by how many steps members took a day before the update. Orange: other early-release members. Grey: randomly sampled members who did not receive the update, in the same group. Dots are median changes, comparing days 1 to 7 with days −14 to −1. Lines are 95% CIs from 4,000 bootstrap resamples of members. The column on the right gives each group's median daily steps before and after.

False steps recorded while resting fell after the update (Figure 6). In 774 members whose rings updated, the mean fell from 24.1 to 19.0 per resting hour, 21% lower. In 1,339 randomly sampled members who did not receive the update, it rose from 23.2 to 25.3 over the same weeks. The difference between the two groups' changes was −7.1 false steps per resting hour (95% CI −11.3 to −2.9).

Figure 6. Mean false steps per resting hour over days −14 to −1 and days 1 to 7, for members whose rings updated (A) and randomly sampled members who did not receive the update (B). A resting period lasted at least 60 minutes between 08:00 and 22:00. Every heart-rate reading in it was within 10% of the member's own median daily resting heart rate. The ring's activity count was above zero, showing it was worn, and below the level typical of walking. The header's interval is a 95% CI from 4,000 bootstrap resamples of members.

Discussion, Limitations and Future Directions

After the update, members who had reported undercounting saw a median 15.2% more daily steps, and false steps recorded while resting fell by 21%. In structured sessions during the algorithm's development, the benchmark device was closer to the hand count in 40 of 60 sessions with steps (two-sided p = 0.030). The ring was closer to zero in sessions without steps.

In the structured sessions, most of the ring's error came from missed steps. It missed 12.4% of the true steps and added 5.2% that were not taken. The updated algorithm is built to count short walks as well as long ones.

The rise in daily steps was largest for those who walked least, for two likely reasons. Median daily steps rose by 756 for members under 5,000 steps a day and by 1,027 at 5,000 to 9,999. A gain of this size is a larger share of a smaller total. Less active members also probably take more steps in short walks. At 10,000 or more, median daily steps rose by 592. More of their day is probably long walks and runs, which were already counted.

Among randomly sampled members who did not receive the update, those in the lowest band rose and those in the highest band fell. Groups formed from one baseline period tend toward the middle, a pattern called regression to the mean. Each band is therefore compared with its own comparison band.

The fall in false steps while resting means the higher daily counts did not bring more of them. Over the same weeks, the comparison group's rose by 9%.

The benchmark device's tendency to add steps is consistent with earlier free-living findings for a wrist-worn tracker [5]. Wrist placement has recorded more steps than waist placement in free living [11].

Limitations and future directions

Members' days have no hand count. The member data therefore show what members will see, but not whether each added step was taken. Both updated groups were chosen to update first and may differ from other members. The reported changes cover one week after the update. Resting periods were inferred from heart rate and the ring's movement, so some steps counted as false may have been real. Resting periods of 30 minutes sit closer to walks and carry more steps. With them, the two groups did not differ.

Each participant contributed up to 18 sessions, which the Wilcoxon tests treated as independent. Their p-values are likely too small; the clustered CIs allow for this but are approximate. Riding a bike likely produces rhythmic motion that resembles walking, and pushing a cart leaves the hand still on the handle. Further structured sessions, including more car and bike travel, would extend these results.

Conclusion

Members who had reported undercounting saw about 15% more daily steps after the update, and false steps recorded while resting fell by a fifth. In structured sessions during its development, the updated algorithm's median error against a hand count was 6.3% in sessions with steps. It recorded no steps at rest or at a desk.