Document Type : Technical Paper
Authors
Shahid Beheshti University
Abstract
Anomaly detection in healthcare datasets is vital for identifying unusual and potentially critical patterns that can influence clinical decisions. However, because many medical datasets are small and have high dimensionality, this task is still a challenge. While deep neural networks have achieved significant success in large-scale data environments, their performance often diminishes when applied to small datasets. Recently, ensemble approaches have gained considerable attention because of their capacity to combine the results of multiple detectors for more robust anomaly detection. In this study, we present an Anomaly Detection approach with Bagging-Principal Component Analysis (ADBP) specifically designed for small, high-dimensional healthcare datasets. Our approach integrates bagging sampling with Principal Component Analysis to create multiple diverse feature spaces, effectively addressing the challenges of limited data and high dimensionality. Specifically, several subsets of the training data are randomly sampled to learn separate Principal Component Analysis subspaces. Within each subspace, the K Nearest Neighbors algorithm is employed to detect anomalies. Final anomaly detection is obtained through a voting mechanism that aggregates the predictions from all detectors. Our ADBP method was evaluated on eight public UCI datasets-Arrhythmia, Parkinson, Annthyroid, Cardiotocography, Hepatitis, Lymphography, Heart Disease, and Pima. Our experimental results show that ADBP algorithm outperforms several state-of-the-art methods, especially in scenarios involving small datasets. The results confirm that integrating bagging, Principal Component Analysis, and ensemble learning techniques significantly improve detection performance in high-dimensional, small datasets. These findings highlight the potential of the ADBP approach as a reliable anomaly detection in healthcare applications.
Keywords
Main Subjects
)