A Graph-Based Density-Variation Clustering Framework for Outlier Removal in Chronic Disease Prediction with Smote-Enn And Random Forest

Chronic diseases such as diabetes, hypertension, chronic kidney disease (CKD), and heart disease impose a major global health burden. Machine-learning prediction systems offer valuable early-diagnostic support, but their performance is often degraded by outliers and class imbalance in real-world clinical tabular data. The standard remedy, DBSCAN, requires manual ε-tuning and assumes globally uniform density — assumptions that rarely hold on heterogeneous medical data. The recent AutoSCAN method removes the manual ε but has been validated only on synthetic clustering benchmarks, not on downstream medical classification. We propose YadroSeg, an extension of the graph-based coring framework of Le et al. [ 1 ], originally introduced for image segmentation, and adapted here to tabular medical data. YadroSeg builds a k-nearest-neighbour weighted graph and analyses a heap-based density variation sequence to identify noisy samples in an adaptive, parameter-free manner, using a novel geometric elbow criterion for automatic threshold selection. The cleaned data are then balanced with SMOTE-ENN and classified by a Random Forest. On five public chronic-disease datasets — diabetes, hypertension, CKD, Cleveland and Statlog heart disease — under 10-fold cross-validation, YadroSeg attains the best F1 on three datasets, reaching 98.87%. On the remaining two, YadroSeg-Grid and AutoSCAN are best, the latter benefiting from full data retention combined with SMOTE-ENN. These results confirm that graph-theoretic density analysis provides a mathematically principled and practically effective approach to noise filtering in imbalanced clinical tabular data.

Hali tarjima qilinmagan