LOF-based Hybrid Sampling for Improved Prediction Performance in Imbalanced Medical Datasets

The imbalanced medical dataset often fails to promise satisfactory prediction performance in traditional learning models, potentially delaying patient treatments. The imbalanced learning problem often leads to biased predictive results towards majority class due to lack of representative examples in minority class. To address the imbalanced medical data problem, we develop a novel hybrid sampling technique based on Local Outlier Factor (LOF) algorithm for enhancing support vector machine (SVM) models' prediction performance for imbalanced datasets. For majority class data, we employ LOF algorithm to detect and eliminate outliers. For minority class data, we use LOF algorithm to select key samples and employ interpolation approach for creating synthetic minority samples in sparse regions. To evaluate the proposed method's efficacy, two medical datasets are deployed in this paper. Three evaluation indicators: G-mean, F1, and index of balanced accuracy (IBA) are used to measure prediction performance. The two medical datasets are set at imbalanced ratio (IR) as 9 to test classification performance using the proposed method. Additionally, this paper's developed method is compared to IMB (using imbalanced datasets) method and Synthetic Minority Over-sampling Technique (SMOTE) in terms of prediction performance. The experimental results demonstrate the proposed method obtains better prediction performance across two medical datasets.