Stratified Normalization Technique with Long Short-Term Memory-Based Autoencoder for Anomaly Detection in Heartbeat ECG Data

Abstract

Disruptions in heartbeat patterns have been identified as a critical health concern globally, often leading to serious health risks and death. Due to the subtle and infrequent nature of these irregularities, individuals may overlook early warning signs, highlighting the need for continuous monitoring and early detection systems. Traditional methods for detecting heartbeat abnormalities, while contributing significantly in the past, are often labour-intensive and lack the precision required for timely intervention. Recent advancements, particularly in wearable electrocardiogram (ECG) devices and machine learning (ML), have changed the narrative in how heartbeat data is collected and analysed. Despite the progress, existing ML models often fall short in accounting for inter-patient variability, which is essential for reliable anomaly detection across diverse populations. In addressing the limitation, this paper proposes Long Short-Term Memory implementation of Autoencoder (LSTM-AE) enhanced with Stratified Normalisation (SN), called LAESN. The LAESN model is designed to improve sensitivity to individual patient differences in ECG signals. The LAESN was trained and evaluated on the ECG5000 dataset using the ‘tanh’ activation function, a batch size of 32, and 50 epochs. It achieved an F1 score and AUC of 0.9660 and 0.9952, respectively, surpassing both the baseline AE and other state-of-the-art models. These results highlight the effectiveness of SN in strengthening the ECG anomaly detection model, enabling it to capture subtle variations in heartbeat signals and support a patient-centric anomaly detection system.