Smartphone Camera Measurement of Breathing-Related Heart Rate Variability Compared With ECG in Healthy Adults

Single group. Each participant completes free breathing (5 min), then paced breathing at 4.0, 4.5, 5.0, 5.5, 6.5 and 7.0 breaths/min (3 min each) and 6.0 breaths/min (5 min) in randomized order, and finally paced breathing at 15 breaths/min (3 min), with 2-minute rests. Throughout each condition the face and fingertip are measured simultaneously by one smartphone, while a chest ECG and a respiratory belt record continuously as references.

Research-prototype iOS data-collection app on a single iPhone 17 Pro. The front camera measures the face for contact-free remote photoplethysmography; the rear main camera, with the flashlight on, measures a fingertip lightly placed over the lens for contact photoplethysmography. No video is stored; the app tracks both cheek regions and saves region-averaged signals with frame timestamps for later analysis. The app plays the breathing guide. It is not used for diagnosis, and readings shown during recording are not used as study results.

Fixed-rate audio cues at 15.0 and at 4.0, 4.5, 5.0, 5.5, 6.0, 6.5 and 7.0 breaths/min, with inhalation taking 40% and exhalation 60% of each cycle, at a comfortable depth without breath holding. A free-breathing condition without cues is included. Participants may rest or stop at any time.

ln(A_pp face / A_pp ECG) for each slow paced-breathing condition (4.0, 4.5, 5.0, 5.5, 6.0, 6.5 and 7.0 breaths/min). A_pp is the peak-to-peak amplitude (ms) of the first harmonic of inter-beat intervals at the guided breathing frequency, estimated separately for the face and the ECG within the same window (60-180 s after the start of each 3-minute condition and 60-300 s for the 5-minute 6.0 breaths/min condition). Unit: natural-log ratio (unitless; 0 means no difference); +0.1 means a facial amplitude about 10% larger than the ECG amplitude. The reported value is the bias, the mean log ratio pooled across participants and the seven rates, with repeated-measures 95% limits of agreement and participant-level confidence intervals. Conditions without a valid facial estimate are counted in the facial measurement failure outcome.

ln(A_pp fingertip / A_pp ECG), with A_pp, analysis windows and summary defined as in the primary outcome: the bias pooled across participants and the seven slow rates, with repeated-measures 95% limits of agreement and participant-level confidence intervals. Unit: natural-log ratio (unitless; 0 means no difference). Conditions without a valid fingertip estimate are counted in the fingertip measurement failure outcome.

For each participant, |ln(A_pp face / A_pp ECG)| minus |ln(A_pp fingertip / A_pp ECG)| in the same window, averaged over the slow paced-breathing conditions in which both sensors give a valid estimate. Unit: natural-log ratio (unitless); positive values mean a larger facial error. The reported value is the mean across participants with a participant-level 95% confidence interval. Each sensor is also compared with the ECG separately, because agreement between face and fingertip does not imply agreement with the ECG.

Percentage of all originally attempted breathing conditions, pooled across participants, in which the facial pathway yields no valid estimate under the quality rules fixed before evaluation data are analyzed. The denominator is every attempted condition, including conditions whose ECG is unsuitable as a reference. Failed conditions are not removed, and the reason for each failure is reported.

Percentage of all originally attempted breathing conditions, pooled across participants, in which the fingertip pathway yields no valid estimate under the quality rules fixed before evaluation data are analyzed. The denominator is every attempted condition, including conditions whose ECG is unsuitable as a reference. Failed conditions are not removed, and the reason for each failure is reported.

Percentage of ECG beats (R peaks) in the analysis windows of all originally attempted conditions, including conditions counted as measurement failures, for which no facial pulse beat is detected within 150 ms of the R peak after the median delay between the ECG and the pulse signal in that condition is subtracted, with each beat matched at most once; a pulse beat detected but displaced beyond this tolerance leaves its ECG beat unmatched. Results at tolerances of 100 ms and 200 ms are reported as a sensitivity analysis. ECG segments unsuitable as a reference are excluded from the denominator and reported separately. The reported value is the mean of participant-level percentages with a 95% confidence interval. Errors computed only from matched beats do not show missed beats.

Percentage of ECG beats (R peaks) in the analysis windows of all originally attempted conditions, including conditions counted as measurement failures, for which no fingertip pulse beat is detected within 150 ms of the R peak after the median delay between the ECG and the pulse signal in that condition is subtracted, with each beat matched at most once; a pulse beat detected but displaced beyond this tolerance leaves its ECG beat unmatched. Results at tolerances of 100 ms and 200 ms are reported as a sensitivity analysis. ECG segments unsuitable as a reference are excluded from the denominator and reported separately. The reported value is the mean of participant-level percentages with a 95% confidence interval. Errors computed only from matched beats do not show missed beats.

Percentage of facial pulse beats detected in the analysis windows of all originally attempted conditions that have no matching ECG beat (spurious beats), using the same one-to-one matching as for missed beats: within 150 ms of the R peak after the median delay between the ECG and the pulse signal in that condition is subtracted, with each beat matched at most once. A beat displaced beyond this tolerance from its ECG beat also counts as unmatched. Results at tolerances of 100 ms and 200 ms are reported as a sensitivity analysis. Beats in ECG segments unsuitable as a reference are excluded and reported separately. The reported value is the mean of participant-level percentages with a 95% confidence interval.

Percentage of fingertip pulse beats detected in the analysis windows of all originally attempted conditions that have no matching ECG beat (spurious beats), using the same one-to-one matching as for missed beats: within 150 ms of the R peak after the median delay between the ECG and the pulse signal in that condition is subtracted, with each beat matched at most once. A beat displaced beyond this tolerance from its ECG beat also counts as unmatched. Results at tolerances of 100 ms and 200 ms are reported as a sensitivity analysis. Beats in ECG segments unsuitable as a reference are excluded and reported separately. The reported value is the mean of participant-level percentages with a 95% confidence interval.

ln(RMSSD face / RMSSD ECG). RMSSD (ms) is computed from facial pulse intervals and from ECG RR intervals in the same window, using only consecutive valid interval pairs. Unit: natural-log ratio (unitless; 0 means no difference). The reported value is the bias (mean across participants) with 95% limits of agreement, for the primary 60-300 s window and, to indicate whether free breathing could be shortened to 3 minutes, for the 60-180 s window of the same recording.

ln(RMSSD fingertip / RMSSD ECG). RMSSD (ms) is computed from fingertip pulse intervals and from ECG RR intervals in the same window, using only consecutive valid interval pairs. Unit: natural-log ratio (unitless; 0 means no difference). The reported value is the bias (mean across participants) with 95% limits of agreement, for the primary 60-300 s window and, to indicate whether free breathing could be shortened to 3 minutes, for the 60-180 s window of the same recording.

ln(SDNN face / SDNN ECG). SDNN (ms) is computed from facial pulse intervals and from ECG RR intervals in the same window, using only valid intervals. Unit: natural-log ratio (unitless; 0 means no difference). The reported value is the bias (mean across participants) with 95% limits of agreement, for the primary 60-300 s window and, to indicate whether free breathing could be shortened to 3 minutes, for the 60-180 s window of the same recording.

ln(SDNN fingertip / SDNN ECG). SDNN (ms) is computed from fingertip pulse intervals and from ECG RR intervals in the same window, using only valid intervals. Unit: natural-log ratio (unitless; 0 means no difference). The reported value is the bias (mean across participants) with 95% limits of agreement, for the primary 60-300 s window and, to indicate whether free breathing could be shortened to 3 minutes, for the 60-180 s window of the same recording.

ln(A_pp face / A_pp ECG) in the 60-180 s window of the 15 breaths/min condition, with A_pp defined as in the primary outcome. Unit: natural-log ratio (unitless; 0 means no difference). The reported value is the bias (mean across participants) with 95% limits of agreement.

ln(A_pp fingertip / A_pp ECG) in the 60-180 s window of the 15 breaths/min condition, with A_pp defined as in the primary outcome. Unit: natural-log ratio (unitless; 0 means no difference). The reported value is the bias (mean across participants) with 95% limits of agreement.

The primary measure, the bias of ln(A_pp face / A_pp ECG) pooled across participants and the seven slow rates, recomputed with less of the start excluded (from 0, 15 or 30 seconds instead of 60 seconds to the end of the third minute) and with fixed 120-second windows starting at 0, 15, 30 and 60 seconds, which correspond to conditions of 2, 2.25, 2.5 and 3 minutes (60-180 s is the primary window). For 6.0 breaths/min the first 3 minutes of the 5-minute recording are used. This separates the effect of the excluded settling period from that of analysis length. Unit: natural-log ratio (unitless; 0 means no difference). Overlapping windows are not treated as independent samples, and a shorter recording is not called sufficient without pre-specified agreement limits.

Bias of ln(A_pp face / A_pp ECG) (mean across participants, with 95% limits of agreement) at 6.0 breaths/min from the 60-180 s window, as in a 3-minute condition, and from the 60-300 s primary window of the same recording, to indicate whether this condition could be shortened to 3 minutes. Unit: natural-log ratio (unitless; 0 means no difference). The 3-minute comparison is limited to this condition and to free breathing (RMSSD and SDNN outcomes) and is not extended to the other rates.

For each participant, the rate between 4.0 and 7.0 breaths/min with the largest A_pp is identified separately from the face and from the ECG. The outcome is the percentage of participants for whom the two rates are the same, among participants with valid estimates at all seven rates from both sources. A maximum at 4.0 or 7.0 breaths/min is flagged as a grid boundary. This rate is not treated as a physiological resonance frequency or an optimal treatment rate.

For each participant, the rate between 4.0 and 7.0 breaths/min with the largest A_pp is identified separately from the fingertip and from the ECG. The outcome is the percentage of participants for whom the two rates are the same, among participants with valid estimates at all seven rates from both sources. A maximum at 4.0 or 7.0 breaths/min is flagged as a grid boundary. This rate is not treated as a physiological resonance frequency or an optimal treatment rate.

Absolute difference (breaths/min) between the breathing rate measured by the respiratory belt and the target rate, averaged over the eight paced conditions for each participant and summarized as the mean across participants. Values for each target rate are supplementary.

Symptoms such as dizziness, tingling, chest discomfort, breathlessness, or warmth or pain at the fingertip that lead to pausing or stopping a recording.

Single study-specific questionnaire item (not a validated scale), rated once for the average experience across all measurements of the visit: difficulty keeping the face within the camera view, on a 0 to 10 numeric rating scale (0 = not at all difficult, 10 = very difficult). Higher scores mean greater difficulty, a worse outcome. Reported as the mean with a 95% bootstrap confidence interval, and the median and interquartile range.

Single study-specific questionnaire item (not a validated scale), rated once for the average experience across all measurements of the visit: difficulty keeping the fingertip on the rear camera, on a 0 to 10 numeric rating scale (0 = not at all difficult, 10 = very difficult). Higher scores mean greater difficulty, a worse outcome. Reported as the mean with a 95% bootstrap confidence interval, and the median and interquartile range.

For each participant, the fingertip-contact difficulty rating minus the face-position difficulty rating (both 0 to 10, as above). Possible range -10 to 10; positive values mean that keeping the fingertip in place was rated more difficult. This is the only pre-specified comparison of the questionnaire. Reported as the mean difference with a 95% percentile bootstrap confidence interval (2,000 participant-level resamples).

Single study-specific questionnaire item asked once at the end of the visit, after both methods have been experienced and rated: which method the participant would choose to measure for 5 minutes every night at home before bed (facial measurement, fingertip measurement, or no preference). Before this item, participants are handed the smartphone used in the session and may handle it as they like. The face and fingertip options appear in the participant's assigned order (face first or fingertip first), with 'no preference' always last. Reported as the number and percentage of participants choosing each option, with 95% Wilson confidence intervals. This is a stated hypothetical choice after simultaneous measurement, not a comparison of methods used independently.