Feature Representation Matters: Preprocessing Effects on One- Class Classification for Class-Imbalanced Medical Data

Abstract Class-imbalanced medical datasets present significant challenges for traditional binary classifiers, particularly when minority-class instances represent clinically critical cases. One-class classification (OCC) has emerged as an effective alternative for addressing such imbalance by learning decision boundaries from only the target class. However, while previous studies have primarily focused on OCC algorithm design, the impact of data preprocessing on OCC performance remains insufficiently explored. In particular, how feature representation changes induced by feature selection and instance selection affect OCC performance has not been systematically investigated. Therefore, this study examines the effects of different preprocessing strategies on OCC-based medical classification using seven hard-to-classify medical datasets selected from the KEEL repository. Three OCC algorithms, namely one-class support vector machine (OCSVM), isolation forest (IF), and local outlier factor (LOF), are evaluated under multiple preprocessing procedures, including feature selection (FS), instance selection (IS), and their combinations in two different orders: FS followed by IS (FSIS) and IS followed by FS (ISFS). Experimental results show that preprocessing significantly improves OCC performance in terms of both AUC and recall compared with baseline models. Among individual preprocessing methods, CNN-based instance selection yields the best overall performance, particularly when combined with LOF, while CART-based feature selection performs better for highly imbalanced datasets. For combined preprocessing strategies, CART-ENN in the FSIS framework and ENN-CART in the ISFS framework achieve the best results, demonstrating that combining feature and instance selection further enhances detection performance. The findings also reveal that preprocessing effectiveness is influenced by class imbalance severity, indicating that adaptive preprocessing strategies are essential for robust OCC-based medical prediction systems. These results provide practical insights into designing effective preprocessing pipelines for one-class classification in class-imbalanced medical applications.