Abstract
- We propose a shallow AI model for reliable detection of breast cancer subtypes based on gene expression data determining the survival probability. This prototype-based model provides easy interpretability of classification decisions and allows to integrate domain knowledge available from bio-medical databases. Decision robustness is obtained by the inherent large classification margin property of the model, which also can be used to establish a reject option for declining model suggestions in case of uncertain decision. Model interpretability is confirmed by the vector quantization approach for the prototypes together with the relevance learning extension, which optimizes the classification decision exploiting those correlation between data features contributing to better class separation. Additionally, the model can be used to mitigate or to remove suspected bias information – here the menopause information. Finally, this prototype-based model is easy to combine with deep learning approaches such that it is working in the resulting embedding space determined by the deep backbone. We demonstrate the model abilities for breast cancer subtypes identification.