Incremental SMOTE with Control Coefficient for Classifiers in Data Starved Medical Applications

Prediction models for data-starved medical applications lag behind general machine learning solutions, despite their potential to improve early interventions. This is largely due to the assumption that optimization approaches are applied on a balanced distribution of events, yet medical data often has an imbalanced distribution within classes. The curse of dimensionality is further exacerbated by small samples and a high number of features in individual-based risk prediction models. In this paper, we propose a data augmentation system to gradually create synthetic minority samples with a control coefficient, which improves the quality of generated data over time and consequently boosts prediction model performance. This system incrementally adjusts to the data distribution, avoiding overfitting. We evaluate our approach using four synthetic oversampling techniques on real asthma patient data. Our results show that this system enhances classifiers' overall performance across all four techniques. Specifically, applying the incremental data augmentation approach to three oversampling methods led to an increase in sensitivity of 4.01% to 7.79% in deep transfer learning-based classifiers.