TLDR: A new lightweight deep learning model, combining Temporal Convolutional Networks (TCN) and Mamba, can accurately predict Atrial Fibrillation (AF) up to two hours in advance using only 30 minutes of heart rate interval data (RRIs). This model is computationally efficient, making it suitable for wearable devices, and outperforms traditional methods in both accuracy and compactness, offering a promising tool for early intervention against AF progression.
Atrial Fibrillation (AF) is the most common type of irregular heartbeat, significantly increasing the risk of serious health issues like stroke and heart failure. While current technologies are good at detecting persistent AF, early stages, particularly paroxysmal AF (PAF) which has sudden and short episodes, often go unnoticed. This is a critical problem because undetected PAF can progress into more severe, sustained AF, leading to higher mortality rates and complications. Early prediction of AF offers a crucial window for preventive treatments, such as specific medications, to slow down or stop disease progression.
Researchers Yongbin Lee and Ki H. Chon have introduced a novel, lightweight deep learning model designed for the early prediction of AF. Their model utilizes only RR Intervals (RRIs), which are the time differences between consecutive heartbeats, making it ideal for integration into wearable devices. The innovative architecture combines a Temporal Convolutional Network (TCN) for encoding the position of data points with Mamba, a selective state space model, to efficiently process sequences of heart rate data in parallel.
The proposed model, named TCN–Mamba–FC, works by first using the TCN block to process the RRI signals, capturing gradual changes that occur before AF onset. This is followed by the Mamba layer, which excels at modeling temporal dependencies within the encoded features. Finally, fully connected (FC) layers are used to predict the probability of AF. This design allows for efficient sequence modeling with linear time complexity, a significant improvement over more computationally intensive methods like Transformers or traditional Recurrent Neural Networks (RNNs).
Key Advantages and Performance
One of the most significant breakthroughs of this model is its ability to predict AF up to two hours in advance using just 30 minutes of input data. This extended lead time is vital for implementing preventive interventions. In rigorous subject-wise testing, the model demonstrated impressive performance metrics, achieving a sensitivity of 0.908, specificity of 0.933, an F1-score of 0.930, an AUROC of 0.972, and an AUPRC of 0.932.
Beyond its accuracy, the model stands out for its computational efficiency. With only 73.5 thousand parameters and 38.3 MFLOPs (million floating-point operations), it is significantly more compact than traditional Convolutional Neural Network – Recurrent Neural Network (CNN–RNN) approaches. For instance, other models tested had parameters ranging from 200K to 10.7M and FLOPs from 54.4M to 517M. This low complexity makes the TCN-Mamba model highly suitable for deployment on resource-constrained devices like smartwatches and other outpatient monitoring tools.
The study utilized two outpatient datasets, IRIDIA-AF and NSR RR, for training and evaluation, ensuring the model’s performance was assessed in realistic ambulatory scenarios. The researchers conducted five separate train/validation/test splits to ensure the robustness and generalizability of their findings.
Also Read:
- Advancing Wearable Healthcare with Mixed-Signal Flexible Systems
- Our Bodies’ Emotional Language Changes Over Time: A Study on Physiological Variability
Future Directions
While the TCN-Mamba model shows great promise, the researchers acknowledge that further validation across larger and more diverse outpatient datasets is necessary to fully assess its generalizability. Additionally, future work aims to incorporate other biosignals, such as ECG (electrocardiogram), PPG (photoplethysmography), and EDA (electrodermal activity), to potentially enhance model robustness and clinical utility. This research represents a significant step towards proactive AF monitoring and prevention, offering a new tool for clinicians and patients alike. You can read the full research paper here.
