Pediatric Photoplethysmography (PPG) In Vivo Data collection

This in vivo dataset was used to benchmark performance in August 2025 of a custom respiratory rate (RR) estimation algorithm, available with its results at: Pulse ROxi Github repo. The dataset contains a 3-min recording for each of 509 participants. Data were collected at three secondary institutions within Nigerian states of Kano, Abuja, and Lagos. The study protocol and all associated forms were approved by Institutional Review Boards at the Nigerian Institute of Medical Research and the Nigerian National Code for Health Research Ethics. Photoplethysmograms (PPG) from pulse oximetry were collected from ~34 participants per study site within five age cohorts: < 3 months, 3–6 months, 7–12 months, 13–36 months, and 37–59 months. Participants were enrolled according to the following: Inclusion criteria Male and female participants < 5 years old Presenting to facility Able to receive PPG sensors on appendage (foot or hand) and forehead Able to receive respiration motion sensor affixed to chest Exclusion criteria Blunt force or penetrating trauma Requiring immediate medical attention Cognitive impairment Wards of the State A designated researcher at each site reviewed the study protocol with a parent/guardian of prospective participants. The parent / guardian would then sign an informed consent form to enroll a participant. At one study site, for 169 participants, skin pigmentation was assessed objectively using individual typology angle (ITA) from the finger dorsum. Objective skin pigmentation protocol was similar as that described in Leeb et al., “The performance of 11 fingertip pulse oximeters during hypoxemia in healthy human participants with varied, quantified skin pigment”, eBioMedicine 2024, Volume 102, 105051. Video of each participant’s torso was taken from two camera angles during the 3-minute data collection. Reference respiration rate for each participant was determined by a panel of three expert raters, all emergency medicine physicians practicing in the United States. Using the video editing software Adobe Premiere Pro, each panelist would timestamp observed breaths while blinded to the other panelists’ annotations. When a panelist could not confidently mark breaths because of camera problems, participant motion, or other complications, the panelist could mark that portion for exclusion and optionally record the reason, with the agreement that a portion excluded by any panelist would be excluded from use as reference data. The panel was also instructed that portions would be excluded when the panelists’ 30-second-average respiratory rates disagreed by more than 3 breaths per minute (bpm). After all panelists annotated the sessions, the authors informed each panelist of any exclusions made by only that panelist and of any short breaths (< 0.33 s). The panelists reviewed these quality checks and optionally revised their annotations. These final annotations are included in the dataset. For non-excluded portions, the authors calculated each panelist’s instantaneous respiratory rates according to Berger et al., “An efficient algorithm for spectral analysis of heart rate variability.” IEEE Trans Biomed Eng. 1986 Sep;33(9):900. The authors then calculated the reference respiratory rate as the 30-second average of the three panelists’ rates, each calculated in overlapping frames taken every 5 seconds. Among all 15,779 such frames, 2,423 frames (15.4%) were excluded for containing a panelist’s mark of uncertainty. The empirical cumulative distribution of disagreement in respiratory rate reference (RR ref) is shown below. After also excluding frames with disagreement > 3 bpm, a total of 9,742 frames (61.7%) were excluded. Of the 509 sessions collected, 388 sessions (76.2%) were represented in part or whole in the 6,037 frames of reference data. Information that could possibly identify the source of the data or participant protected health information has been removed. Acknowledgements The study was sponsored by Global Health Labs, Seattle, WA, US. The study was carried out in collaboration with eHealth Africa, headquarters in Kano, Nigeria. We also are grateful to all study personnel and rater panelists.