Interpreting deep patient stratification models with topological data analysis

Research output: Chapter in Book/Report/Conference proceedingConference contribution

17 Downloads (Pure)

Abstract

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.

Original languageEnglish
Title of host publicationInternational Conference on e-Health and Bioengineering (EHB 2023 ): proceedings
PublisherSpringer Cham
Pages563–574
ISBN (Electronic)9783031625022
ISBN (Print)9783031625015
DOIs
Publication statusPublished - 30 Aug 2024
Event11th International Conference on e-Health and Bioengineering - Romania, Bucharest, Romania
Duration: 09 Nov 202310 Nov 2023
http://www.ehbconference.ro/Home.aspx

Publication series

NameIFMBE Proceedings
PublisherSpringer, Cham
ISSN (Print)1680-0737
ISSN (Electronic)1433-9277

Conference

Conference11th International Conference on e-Health and Bioengineering
Abbreviated titleEHB 2023
Country/TerritoryRomania
CityBucharest
Period09/11/202310/11/2023
Internet address

Keywords

  • interpreting
  • deep patient stratification models
  • topological data analysis

Fingerprint

Dive into the research topics of 'Interpreting deep patient stratification models with topological data analysis'. Together they form a unique fingerprint.

Cite this