New Autoencoder-Based Method for Efficient Anomaly Detection in ECG Bio-Signals

Abstract

Anomaly detection is crucial in many fields, including finance, healthcare, and cybersecurity. It can identify irregular patterns, ensure system integrity, and prevent significant losses. In this paper, we explore using an autoencoder for anomaly detection. Autoencoder is a type of neural network that is suitable for unsupervised learning. This work introduces a new autoencoder-based method and examines the architecture and training process of the autoencoder, evaluates its performance on the ECG dataset, and compares its effectiveness with baseline anomaly detection methods. The results indicate that the autoencoder significantly outperforms conventional machine learning models, achieving an accuracy of 96.96%, a high precision of 98.85%, and a balanced recall of 95.88%. Additionally, it attains an F1-score of 97.34% and a ROC-AUC of 97.17%, demonstrating superior detection ability and minimal false positives compared to PCA, MCD, and Isolation Forest models. Despite its strengths, the paper identifies critical drawbacks of autoencoders, including their sensitivity to hyperparameter selection and the need for extensive training datasets to achieve adequate performance. These challenges underscore the importance of fine-tuning and large-scale data, which can be resource-intensive. Finally, it recommends enhancing the reliability of anomaly detection systems, emphasizing the need for robust methodologies to address these limitations. It also identifies areas for future research, suggesting that further investigation could lead to more flexible and efficient anomaly detection methods.