Vitals SDK Accuracy: Clinical Validation vs Polar H10

The Vitals SDK was validated against the Polar H10 chest strap — the reference-standard wearable ECG used in HRV research — across 30 volunteers and 265 accepted readings. The study measured all six SDK metrics under two conditions (resting and paced breathing), on both iOS and Android devices, and across six Monk skin tones. This page summarizes the full results. For methodology details, statistical appendices, and Bland-Altman plots, see the full report linked at the bottom of this page.

​

Study Design

Both iOS and Android phones were used simultaneously with the chest strap during each reading, so the SDK and reference measurements are directly time-paired.

​

Full Accuracy Results

The table below reports mean absolute error (MAE), Pearson correlation (r), and 95% Bland-Altman limits of agreement for all six metrics across all 265 accepted readings.

​

Results by Subgroup

No subgroup produced a single reading off by 20 ms or more. The table below breaks down RMSSD error and the proportion of readings within 10 ms for each subgroup.

​

Quality Gating: Withheld Readings

The SDK includes an internal signal-quality evaluator that runs throughout every scan. When the final quality score falls below the acceptance threshold, the SDK withholds the result rather than returning a potentially misleading value. In this study, 3.3% of readings were withheld by the quality gate. All accuracy figures reported above are computed only on accepted readings — the SDK never returned an accepted reading that was later found to be materially wrong.

​

Why Finger Scan Outperforms Face Scan

RMSSD is built from tiny differences between consecutive heartbeat intervals. Any imprecision in detecting when each beat peaks contributes directly to RMSSD error, which means measurement SNR is the dominant factor — not post-processing. On a face, the PPG pulse is a faint color change in reflected ambient light across a large skin area. Room lighting variation and sub-centimeter head movement can swamp the pulse entirely, forcing heavy temporal filtering that smears the beat timing the SDK needs to be precise. A fingertip pressed against the rear-flash LED is a fundamentally different measurement geometry. The LED shines a bright, steady, point-source light directly into the skin with no air gap, producing:

• A pulse amplitude approximately 6× larger than a face signal
• Over 30× the signal power relative to the noise floor
• Immunity to ambient lighting variation, since the flash overwhelms any room light

The accuracy consequence is stark. The best published face-scan RMSSD result in the literature is 10.5 ms (Bioengineering, 2023, UBFC-rPPG dataset). The Vitals SDK achieves 3.4 ms against the same Polar H10 class of reference device — nearly three times more accurate.

​

Known Limitations

The following conditions and populations are either not yet validated or are expected to produce lower accuracy based on the study design. These limitations will be updated as the blind validation study and subsequent studies are completed.

Skin tones not yet validated

The study included participants with Monk skin tones 2, 3, 4, 5, 6, and 8. Monk skin tones 7, 9, and 10 have not been tested. Accuracy for these tones is not claimed. The validation study is actively being expanded to cover all tones.

Movement during the scan

All validation readings were taken with the participant seated and stationary. Accuracy during movement — including light walking, standing transitions, or hand tremor — has not been characterized. The quality gate will withhold readings when motion artifacts are severe, but mild motion-induced error is not quantified.

Breathing rate outside paced conditions

Breathing rate was only validated under a paced breathing protocol at 6 breaths per minute. Accuracy of the breathing rate metric under free, unpaced breathing has not been measured. The SDK applies confidence gating to breathing rate and will withhold it when the signal does not meet quality criteria.

Age distribution

The majority of participants (21 of 30) were in the 25–34 age range. Single-participant subgroups (18–24 and 45–54) describe one individual only. Accuracy for participants over 64 has not been characterized.

Irregular heartbeats

Three readings containing frequent irregular heartbeats were excluded from the reference dataset before analysis. The SDK’s behavior in users with persistent arrhythmias is not characterized and no claim is made for this population.

​

Full Validation Report

The complete validation report includes full Bland-Altman plots, scatter plots for all metrics, per-reading data appendices, and detailed methodology.