Ensemble of Autoencoders for Anomaly Detection in Biomedical Data: A Narrative Review

Abstract

In the context of biomedical data, an anomaly could refer to a rare or new type of disease, an aberration from normal behavior, or an unexpected observation requiring immediate attention. The detection of anomalies in biomedical data has a direct impact on the health and safety of individuals. However, anomalous events are rare, diverse, and infrequent. Often, the collection of anomalous data may involve significant loss of human life and healthcare costs. Therefore, traditional supervised machine and deep learning algorithms may not be directly applicable to such problems. Biomedical data are often collected in the form of images, electronic health records, and time series. Typically, an autoencoder (AE) or its corresponding variant is trained on normal data, and an anomaly is identified as a significant deviation from these data based on reconstruction error or other metrics. An Ensemble of AEs (EoAEs) can serve as a robust approach for anomaly detection in biomedical data by combining diverse and accurate views of normal data. An EoAE can provide superior detection to a single encoder; however, its performance can depend on various factors, including the diversity of the created data, the accuracy of the individual AEs, and the combination of their outcomes. Herein, we perform a comprehensive narrative literature review on the use of EoAEs when using different types of biomedical data. Such an ensemble provides a promising approach for anomaly detection in biomedical data, offering the potential for performance improvement by leveraging the strengths of diverse AEs. However, several challenges remain, such as the need for data specification and determination of the optimal number of AEs in the ensemble. By addressing these challenges, researchers can enhance the effectiveness of EoAEs for anomaly detection in various types of biomedical data. Furthermore, through this review, we highlight the significance of evaluating and comparing the performance of an EoAE with that of single AEs by establishing agreed-upon evaluation metrics and investigating normalization techniques for anomaly scores. We conclude the review by presenting challenges and open questions in the field with for future research.