Class Decomposition for Gastric Cancer Detection from Breath

Machine learning and data mining applications aimed at improving the accuracy of cancer detection that define it as a classification task usually define this as a binary class problem. However, in many cases, these classes are not homogenous: there are different subgroups of healthy participants with different diagnoses and different types of cancer, and this information can be useful when inducing classifiers (e.g., healthy individuals with pre-cancerous lesions can have different features than other healthy participants or even features that are similar to cancer patients). Therefore, in this study we propose a class decomposition approach that would allow taking advantage of this information and demonstrate this on a gastric cancer breath data example.