A Neural Network for Atrial Fibrillation Detection via PPG

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Atrial fibrillation (AF) is a common arrhythmia leading to serious health issues like stroke.
• Current AF detection methods face challenges in accessibility and scalability.
• Early detection of AF is crucial for effective patient management.

Purpose of the Study:

• To develop a predictive neural network model for atrial fibrillation (AF) detection using smartphone-derived photoplethysmography (PPG) signals.
• To explore the feasibility of using readily available technology for scalable cardiac arrhythmia screening.
• To establish a foundation for digital health solutions in cardiovascular disease monitoring.

Main Methods:

• Utilized photoplethysmography (PPG) signals captured via smartphone applications.
• Collected data from student participants and augmented with open-source datasets.
• Developed and trained a multilayer perceptron (MLP) neural network using TensorFlow.
• Evaluated model performance using accuracy, sensitivity, specificity, and F1-score.

Main Results:

• The developed multilayer perceptron (MLP) model achieved an initial accuracy of approximately 75% in detecting atrial fibrillation.
• The results demonstrate the potential of PPG signals for non-invasive AF screening.
• Further optimization is needed to enhance predictive accuracy and reduce false positive rates.

Conclusions:

• Smartphone-based PPG signals offer a viable, accessible tool for preliminary atrial fibrillation detection.
• The developed neural network model shows promise for scalable digital health applications.
• Continued research and validation are essential for clinical implementation and integration into programs like the Austrian Digital Heart Program.