PO-04-214 VIDEO-BASED FACIAL DETECTION OF ATRIAL FIBRILLATION ACROSS HUMAN COMPLEXIONS

Abstract

It has been posited that the Black vs. White disparity in stroke risk may be partly attributable to the relative underdetection of AF in Black compared with White patients. We propose a novel monitoring solution that could minimize the impact of the socio-economic disparities in access to remote monitoring solutions. We developed a video-based monitoring technology to detect AF without the need for a patient to adopt a dedicated wearable device. It is a low-cost software-only solution running on smartphones. It eliminates requirements for patient compliance with recording procedures since it passively monitors patients while they are using their smartphones for other purposes. The technology captures a pulsatile signal from patients' faces. Remote video-based cardiac monitoring measures the modulation of ambient light reflected from one’s face due to changes in blood volume (hemoglobin) in the upper layers of the skin. In addition to hemoglobin, melanin is also a chromophore of the human skin. Hence, we aim to measure the accuracy of the technology to measure heart rate (HR) and its performance to detect the presence of AF across the whole spectrum of human complexions. We enrolled patients with paroxysmal and persistent AF. We defined 3 complexion groups using the Fitzpatrick score: white (A), tanned and olive-color (B), brown and black (C). Measurements were conducted under 4 indoor illumination levels: 50, 100, 200, and 500 lux. For reference, we used a single-lead EKG. We used 25% of the cohort to train the technology. HR accuracy was assessed using the Bland-Altman method with paired replicate measures based on one-way repeated ANOVA. The detection performance was measured as average sensitivity (SENS) and specificity (SPEC) adjusted for repeated measures. We collected 7,042 recordings from 60 AF patients (67±10 yrs, 44 men). We enrolled 23, 26, and 11 patients in the complexion groups A, B, and C, respectively. In the validation, 87% of the recordings were conclusive. Overall SENS and SPEC were 96% and 92%. For an illumination >50 lux, the SENS and SPEC for the complexion groups were: [A: 99%,94%], [B: 91%,93%], [C: 96%,90%]. The mean HR measurement errors in reference to the EKGs were: [A: -2.0 bpm], [B: -0.2 bpm], and [C: 0.3 bpm] with limits of agreements <5 bpm (upper to lower bound intervals). The proposed monitoring technology enables an accurate measure of HR and the detection of AF across all human skin complexions and under indoor illumination.