A stacked LSTM for atrial fibrillation prediction based on multivariate ECGs

Area of Science:

• Biomedical Engineering
• Cardiology
• Artificial Intelligence in Medicine

Background:

• Atrial fibrillation (AF) is a common arrhythmia increasing stroke and heart failure risk.
• Electrocardiography (ECG) is crucial for monitoring heart conditions.
• Accurate AF prediction from ECGs can improve patient outcomes.

Purpose of the Study:

• To develop and evaluate a novel model for predicting atrial fibrillation (AF) using ECG data.
• To address the gap in research focusing on AF prediction rather than just detection.
• To improve the accuracy and effectiveness of AF prediction systems.

Main Methods:

• Development of a recurrent neural network (RNN) named SLAP, utilizing stacked Long Short-Term Memory (LSTM) layers.
• Implementation of techniques to mitigate gradient explosion and vanishing gradient problems inherent in standard RNNs.
• Comprehensive experimental validation using two publicly available ECG datasets.

Main Results:

• The SLAP model achieved 92% accuracy in predicting atrial fibrillation.
• The model demonstrated a 92% f-score for AF prediction.
• Performance results surpassed those of existing state-of-the-art AF detection architectures, including standard RNN and LSTM models.

Conclusions:

• The developed SLAP model, based on stacked LSTMs, shows significant promise for accurate AF prediction.
• SLAP offers an effective approach to overcome limitations of traditional RNNs in learning complex temporal features.
• The high accuracy and f-score indicate SLAP's potential to enhance patient monitoring and management for AF.