GBC-AST: a cluster-based oversampling method for heart failure prediction in imbalanced medical data sets

The main aim behind this study is to assist medical practitioners and patients by providing a preemptive warning of the risk of potential heart failure. Researchers have applied machine learning for heart failure (HF) prediction. These studies found that HF is associated with multiple risk factors. A significant challenge in this domain is class imbalance, where the majority of instances belong to the majority class (e.g., healthy or non-critical cases), while only a small portion of instances represent the minority class (e.g., patients at risk or critical cases), making accurate prediction for the minority class particularly difficult. Traditional machine learning classifiers tend to favor the majority class, leading to poor classification of the minority class. In this paper, we have proposed a novel cluster-based oversampling technique (GBC-AST) for mixed types of data. The proposed GBC-AST first employs a Gaussian Mixture Model (GMM) to partition the data into sub-clusters. Oversampling is then performed in each sub-cluster by using Modified Adaptive Synthetic Sampling (M-ADASYN). M-ADASYN generates synthetic samples, particularly focusing on hard-to-learn (HTL) observations of the minority class. To further improve the classification of the minority class, GBC-AST incorporates a data cleaning step using Tomek links, which effectively expands the minority class region and reduces overlap with the majority class. The proposed GBC-AST is evaluated across eleven publicly available datasets using five different classifiers. Experimental results demonstrate that GBC-AST outperforms other oversampling techniques, achieving statistically significant improvements in terms of F-measure, G-mean, and AUC..