Original title: A Systematic Analysis of Oversampling Intensity and Decision Thresholds in Imbalanced Medical Classification Under Leakage-Free Cross-Validation
Kuan-Chu Lu, Ting-Wei Wu · September 2026
Medical datasets often have a balance problem because healthy patients outnumber sick ones. Computers trained on this data struggle to spot the rare sick cases. Researchers usually fix this by making fake sick cases, called oversamplingadding extra examples of the rare class to balance the data, or by lowering the cutoff score needed to flag someone as sick. This study tested doing both together on three medical datasets. They found that changing the fake data amount and the cutoff score together matters a lot. For two diseases, lowering the cutoff below 0.50 helped most, while another worked best near 0.50. Picking between two common fake-data methods barely changed anything. However, creating fake data made the computer worse at predicting true probabilities. Doctors should adjust both settings together for their specific data.
Imagine you are packing for a weekend trip and the weather report keeps changing. You can stuff in extra layers, but your bag has limited space, and the right amount depends on where you are going. Pack too much and you cannot carry it; pack too little and you are cold. In this paper, the extra layers are the repeated examples of the rare condition, the changing forecast is the confidence cutoff for flagging a case, and the researchers show the right balance is different for each medical dataset.
Medical classifiers are often trained on data where one condition is rare, and how you handle that imbalance changes which cases get flagged. This work shows the right settings vary by dataset, so a one-size-fits-all fix can quietly hurt performance. Tuning the threshold and the amount of oversampling together, rather than picking a method, gives a more honest picture of what works.
The abstract doesn't say what the limitations are; it only reports results on three datasets and five classifiers, so the findings may not generalize. It also does not identify the specific oversampling intensities that were optimal, and it does not explain why lower thresholds helped some datasets or whether the Brier-score increase reflects a fundamental trade-off.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
A Systematic Analysis of Oversampling Intensity and Decision Thresholds in Imbalanced Medical Classification Under Leakage-Free Cross-Validation — Kuan-Chu Lu, Ting-Wei Wu — Mathematics — doi:10.3390/math14183421
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.