Counterfactual synthetic minority oversampling technique: solving healthcare's imbalanced learning challenge

The application of machine learning in the healthcare domain has groundbreaking potential across a wide range of scenarios. However, this potential is often stalled by data-related challenges, such as the imbalanced nature of the domain's data, where critical outcomes tend to be inherently rare. To address this challenge, we propose a novel oversampling approach, the counterfactual synthetic minority oversampling technique (Counterfactual SMOTE), which combines SMOTE with a counterfactual generation framework. Our method intrinsically performs an oversampling process near the decision boundary within a safe region of space, allowing for the generation of informative but non-noisy minority samples. To validate the proposed framework, a rigorous experimental procedure was conducted across a set of highly imbalanced binary classification challenges in healthcare. The results demonstrate the superiority of the proposed method over several of the most commonly used oversampling alternatives presented in the literature. Notably, Counterfactual SMOTE was the only method to present a convincingly superior performance when compared with the original SMOTE. Although the proposed method was specifically validated in the healthcare domain, owing to its relevance and frequently imbalanced nature, we expect the findings of this study to be generalizable to any imbalanced scenario.