Tree Mean Shift Clustering-Assisted Mean K-Dimensional Near Miss Undersampling with Gated Recurrent Units for Imbalanced Medical Data Classification

Abstract

Medical data classification is inherently challenging, especially when it comes to minority class recognition, and is enhanced by class imbalance. A Tree Mean Shift Clustering-assisted Mean K-Dimensional Near Miss (TMSC-MKDNM) undersampling strategy is proposed in this study and is combined with a Gated Recurrent Unit (GRU) classifier. TMSC detects structural distributions in the majority class, and MKDNM filters out "noisy" instances by sparsely discarding redundant instances based on multiple neighborhood relationships prior to classification by GRU. The framework takes into account a dataset of 10,000 records having 20 features initially with a ratio of 9:1 and reaches a 1.5:1 ratio after controlled undersampling. Based on computational assessment, the highest accuracy reached by TMSC-MKDNM-GRU is 96.4%, whereas the highest precision, recall, and F1 values are 95.1%, 96.2%, and 95.6%, respectively. G-Mean value is 96.3% while the highest AUC value is 0.982 with ROC score. Ablation and comparative analyses show that our structure-aware undersampling method achieves better recognition of the minority classes in comparison to GRU, random undersampling-GRU, NearMiss-GRU, and MKDNM-GRU, supporting the effectiveness of structure-aware undersampling for robust imbalanced medical classification.