Detection of intra-beat waves on ambulatory ECG using manifolds: An explainable deep learning approach

The development of wearable technologies for acquiring bioelectrical signals is expanding, increasing the demand for algorithms to analyze large clinical datasets. Accurate intra-beat waveform detection in ECG analysis is critical for assisting cardiologists in diagnosing cardiac diseases. We present a novel deep-learning detector capable of detecting specific intra-beat waves without the need for heartbeat identification, addressing a key limitation in current approaches. The model is trained to directly detect one of three key waveforms: P, QRS, or T. This approach is well-suited for applications such as identifying the T wave for repolarization alternans or ischemia. We employ a rigorous patient separation methodology and use a gold standard with expert manual labels from two public databases: QTDB and LUDB. The model utilizes a simple autoencoder (AE) architecture, offering interpretability insights by visualizing decision-making in a latent space. During the model design stage, the system achieves F1-scores of 0.93, 0.97, and 0.93 for P, QRS, and T wave identification. In real-world ambulatory environments, the in-line detector performs at 0.94, 0.98, and 0.96 for each wave. This work uses manifold learning for ECG intra-wave detection with simple, explainable models based on the applied machine learning principles. It outperforms state-of-the-art methods and provides valuable insights into the decision-making process, making it particularly well-suited for real-world applications, offering both accuracy and interpretability in ambulatory ECG analysis. • Intra-beat ECG detector identifies P, QRS, or T waves without prior heartbeat detection. • The method suits real-time and offline use in ambulatory recordings under real conditions. • Deep Learning model based on an autoencoder tackles black-box issues, offering decision-making insights. • Latent space interpretation boosts physician confidence and system transferability. • Simplified architecture enhances model interpretability and applicability in practice.