Enhancing patient stratification and interpretability through class-contrastive and feature attribution techniques

This work advances explainable machine learning for patient stratification in inflammatory bowel disease, particularly Crohn's disease subtypes. It employs Gaussian Mixture Modelling for interpretable probabilistic subtype modelling, identifies risk genes by subtype using a modified kernelSHAP approach accounting for gene correlations, pinpoints relevant gene modules per subtype, and develops class-contrastive techniques for visual explanation of patient subtype predictions. The model-agnostic approach has potential applications in other diseases and domains where explainability and incorporating feature correlations are crucial.