Interpreting Deep Patient Stratification Models with Topological Data Analysis

Patient stratification is a crucial task aimed at categorizing individuals with a specific disease into more homogeneous subgroups based on critical disease-related characteristics. This process enables personalized interventions, optimized care management, and tailored treatments. Patient stratification plays a significant role in drug development and clinical practice for many diseases. However, with the increasing availability of biomedical data, such as gene expression data, clinical records, and lifestyle/environmental factors, the analysis of this vast and multimodal data becomes highly challenging. Machine learning offers methods that can help address the challenges of transforming this extensive and diverse data into usable decision-support tools. Deep learning methods, in particular, have shown impressive results in tasks such as risk stratification and treatment response prediction. However, their impact on data-driven medicine remains limited due to their 'black-box' nature and their inability to provide human-interpretable outputs. In this study, we propose applying topological data analysis to enhance the interpretability of deep learning patient stratification models. Specifically, we suggest using the Mapper algorithm to visualize the latent space learned by the models through the lens of its predictions. We apply the Mapper algorithm to various architectures of recently developed deep patient stratification models and demonstrate how it helps reveal relationships among different patient subgroups. Furthermore, we adapt the Normalized Mutual Information measure to identify the Mapper's parameters that yield the most optimal graph-based representation of the latent space. This approach aims to enrich the power of deep learning with interpretable results in the field of patient stratification.