CardioPRIME: Cardiovascular Physiological Representation Integration With Multimodal Embeddings
Levine, Z Rossman, H Segal, E
Levine, Z
Rossman, H
Segal, E
Author
Supervisor
Department
Computational Biology
Embargo End Date
Type
Workshop
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
We introduce CardioPRIME, a hybrid mechanistic–deep learning framework for electrocardiogram (ECG) analysis, which integrates clinical phenotypes along-side classical ordinary differential equation-based cardiac models. Despite the efficacy of purely data-driven neural networks, real-world medical data are rooted in decades of physiological research. By enforcing consistency between ECG- derived latent representations and multiple clinical modalities, CardioPRIME produces more discriminative and physiologically grounded embeddings than a non- integrated baseline. Our results, reflected by higher clustering metrics (NMI, AMI, Homogeneity, and Completeness) across top prevalent diseases, underscore the significance of clinical integration for improved disease separation and enhanced interpretability. This indicates that bridging physiological modelling with data-driven techniques can substantially advance representation learning in the cardiovascular domain for general disease diagnosis.
Citation
Z. Levine, H. Rossman, P. Ai, and E. Segal, “CardioPRIME: Cardiovascular Physiological Representation Integration With Multimodal Embeddings,” Mar. 06, 2025.
Source
Learning Meaningful Representations of Life (LMRL) Workshop at ICLR 2025
Conference
Keywords
Interpretable representation learning, multimodal representation learning, neural ordinary differential equations, electrocardiogram
Subjects
Source
Publisher
OpenReview