<![CDATA[Class imbalance is a common problem in health datasets and often leads to poor recognition of the minority (disease) class. NR-Clustering-SMOTE is a cluster-based oversampling method that combines noise reduction, K-Means clustering, and SMOTE with a modified distance metric to improve classification performance on imbalanced health data. However, the original method only performs noise reduction before SMOTE, so noisy synthetic samples generated around borderline or highly overlapped regions may still degrade classifier performance and introduce epistemological bias in the learned decision rules. This paper proposes a lightweight extension called NR-CluSMOTE-KNC (NR-Clustering-SMOTE with Post-SMOTE k-NN Cleaning). After the standard NR-Clustering-SMOTE pipeline, a k-nearest neighbour filter is applied solely to synthetic minority samples; synthetic points that are surrounded predominantly by majority neighbours are identified as extreme noise and removed. On the Pima Indians Diabetes dataset using Random Forest, the proposed method improves accuracy from 0.8481 (baseline NR-Clustering-SMOTE) to 0.8589 with NR-CluSMOTE-KNC, accompanied by consistent gains in G-Mean and AUC. These results indicate that a simple post-SMOTE cleaning step can epistemologically refine the representation of minority concepts in the data, producing more reliable and fair predictive models for health decision support.]]>
Copyrights © 2026
