Towards Dynamic EHR Phenotyping: A Generative Clustering Model
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Clustering, Healthcare, Time-Series, Generative
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: In healthcare, identifying clinical phenotypes—subgroups defined by specific
clinical traits—is essential for optimizing patient care. The wealth of Electronic
Health Record (EHR) information has fueled data-driven approaches to tackle
this challenge. Unfortunately, the heterogeneity, multi-modality, and dynamic
nature of EHR data pose significant hurdles. We propose DeepGC, a novel
generative, clustering, outcome-sensitive end-to-end deep learning (DL) model for
uncovering dynamic phenotypes within temporal EHR data. DeepGC leverages
patient trajectories and outcomes to identify clinically meaningful phenotypes that
evolve over time. Our generative model employs a dynamic sequential approach
based on a Markovian Dirichlet distribution and Variational Auto-Encoders (VAEs),
which is capable of providing insights into the evolution of patient phenotypes
and health status. Preliminary evaluation indicates that DeepGC shows promise
in identifying distinct and interpretable phenotypes, and outperforming existing
benchmarks, particularly with regard to outcome sensitivity (3 % increase in F1).
We also showcase the model’s potential to yield valuable insights into the future
evolution of patients’ health status
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7811
Loading