Can Better Sampling Fix Biased Cancer Predictions?

This study investigates the effect of sampling techniques on cancer type classification across five categories: breast, kidney, colon, lung, and prostate. We evaluate the impact of three sampling strategies: (i) Simple Random Oversampling/Undersampling, (ii) Stratified Random Sampling, and (iii) Synthetic Minority Oversampling Technique (SMOTE) on the performance of seven different machine learning models. Results show that incorporating sampling noticeably improves model performance, with SMOTE achieving the highest gains. On average, sampling improved F1 scores by 11.44%, highlighting its critical role in addressing class imbalance in medical datasets. These findings provide a practical and reproducible approach to mitigate data imbalance in biomedical machine learning workflows.