Experimental Evaluation of Resampling Techniques for Imbalanced Stroke Data

Imbalance data, which are one class that dominates in number compared to the other class, is a common problem with medical predictions. This issue frequently causes model bias towards the majority class, which results in terribly low detection rates for important yet infrequent conditions like strokes. The ratio of stroke to nonstroke cases reflects real-world problems in clinical practice; this bias is a risk since the missed stroke predictions lead to delayed intervention and increased mortality rates. Accurate medical diagnosis through machine learning depends on solving class imbalance issues in order to enhance both reliability and efficiency. The performance of five oversampling and four undersampling methods is examined. The results indicate that the size of the dataset is crucial in determining the optimal resampling strategy. Gradient Boosting with Random Oversampling for the smaller dataset achieved the highest F1 score (0.22). In general, the undersampling methods consistently outperformed the oversampling approaches, producing more stable outcomes and higher recall scores. These findings provide valuable insight into the handling of class imbalance in medical datasets, suggesting that straightforward undersampling techniques may be preferable to complex oversampling methods for stroke prediction.