Performance of Noncontact Video-based Detection of Pulse Rate and Atrial Fibrillation on the iOS Platform

Gill Tsouri, Alex Page, Margot Lutz, Jean-Philippe Couderc
VPG Medical


Abstract

Background: We developed a video-based monitoring technology to detect AF and Pulse Rate (PR) without the need for a patient to adopt a dedicated wearable device. It minimizes requirements for patient compliance with recording procedures since it passively monitors patients while they use their personal computer for other purposes. Similar to PPG, the technology captures a pulsatile signal from a patient's face. In this work, we investigate performance on the iOS platform using a personal iPhone X smartphone device.

Methods: We validated our algorithms that have already been trained on the Samsung Galaxy S10 device on an iPhone X smartphone. We collected validation data from 22 paroxysmal and persistent AF patients who did not participate in the training of the algorithms. Subjects were spread across the entire skin-tone Fitzpatrick scale. Measurements per subject were collected under 4 indoor illumination levels (50, 100, 200, and 500 lux) for both LED and incandescent light sources. For reference, we used a single-lead EKG. PR accuracy is assessed using Bland-Altman (BA) analysis, and AF detection performance is assessed using RoC performance curve of sensitivity vs. specificity.

Results: We collected 958 recordings from 22 AF patients (age: 67±10 years). 148 recordings were automatically discarded by our technology quality filtering algorithms due to subject motion and face obstruction. The remaining 810 recordings (85% of all recordings) were used as the validation data set. The BA plot shows unbiased estimation with strong agreement to ECG. 97% of estimated PR values have no more than 5 BPM deviation from the ECG HR measure. The AF detection ROC curve has an Area Under the Curve (AUC) of 0.93.

Conclusion: These results demonstrate how HealthKam Afib trained on the Android Galaxy S10 platform ports well to the iOS platform, when considering PR estimation and AF detection.