Non-Contact Atrial Fibrillation Detection From Face Videos by Learning Systolic Peaks
Key result
A deep learning-based non-contact atrial fibrillation detection method using 30-second face videos achieved 96.00% accuracy, 95.36% sensitivity, and 96.12% specificity for classifying healthy subjects versus AF patients.
Why the study?
Does a non-contact face video-based approach accurately detect atrial fibrillation compared to standard ECG?
Population
100 healthy subjects and 100 AF patients
Comparison
Non-contact video-based PPG vs contact PPG and ECG
Authors
Discussion
Loading...
Member takes
Implication
Supports video-based AF screening feasibility; leaves open prospective ECG validation before clinical adoption.
Study Design
Type
Cross-Sectional (n=200)
Multicenter
No
Structured PICO
Does a non-contact face video-based approach accurately detect atrial fibrillation compared to standard ECG?
P
Population
200 participants, comprising 100 healthy subjects and 100 patients with atrial fibrillation, evaluated to validate a non-contact AF detection algorithm using face videos.
I
Intervention
Non-contact atrial fibrillation detection from face videos using a 3D convolutional neural network and support vector machine classifier
C
Comparator
Electrocardiography (ECG) and contact photoplethysmography (PPG)
O
Outcome
Diagnostic accuracy, sensitivity, and specificity for classifying healthy versus AF and sinus rhythm versus AFsurrogate
Non-contact AF detection from face videos using deep learning is highly accurate and could enable home-based self-screening and monitoring.
Limitations
- Limited number of atrial flutter (AFL) patients in the dataset, leading to unbalanced classes for AFL detection.
- Spontaneous facial motions can lead to noises in remote PPG signals.
- Some PPG signals from patients contain systolic peaks with very low amplitudes due to small blood stroke volume, which may cause wrongly detected peaks.
Cite This Study
Sun et al. (2022) conducted a cross-sectional in Atrial Fibrillation (n=200). Non-contact AF detection from face videos (3DCNN-PEAK) vs. Electrocardiography (ECG) and contact photoplethysmography (PPG) was evaluated on Classification accuracy of healthy versus AF. A deep learning-based non-contact atrial fibrillation detection method using 30-second face videos achieved 96.00% accuracy, 95.36% sensitivity, and 96.12% specificity for classifying healthy subjects versus AF patients.