This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Fusing Echocardiography Images and Medical Records for Continuous Patient Stratification
Created by
Haebom
Author
Nathan Painchaud, J er emie Stym-Popper, Pierre-Yves Courand, Nicolas Thome, Pierre-Marc Jodoin, Nicolas Duchateau, Olivier Bernard
Outline
This paper presents a deep learning method for automatically and robustly extracting cardiac function descriptors (e.g., ejection fraction or strain) from cardiac ultrasound image sequences. By mimicking the process of a physician assessing a patient's condition by considering cardiac function descriptors alongside global variables from the medical record, we propose a novel Transformer model applied to tabular data to learn representations of challenging-to-characterize continuous cardiovascular diseases such as hypertension. After projecting each variable into a unique representation space using a modality-specific approach, the normalized representation of this multimodal data is fed into a Transformer encoder to learn a comprehensive patient representation using a clinical grade prediction task. This stratification is formulated as an ordinal classification to reinforce the pathological continuum in the representation space. We observe key trends along this continuum in a cohort of 239 hypertensive patients, providing unprecedented detail on the impact of hypertension on various cardiac function descriptors. Our analysis shows that i) the XTab baseline model architecture achieves excellent performance (96.8% AUROC) even with limited data (less than 200 training samples); ii) stratification across populations is reproducible across training runs (within 5.7% mean absolute error); and iii) patterns emerge in the descriptors, some of which are consistent with existing physiological knowledge about hypertension and others that may pave the way to a more comprehensive understanding of this pathology. The code is available at https://github.com/creatis-myriad/didactic .
Development of a hypertension diagnosis model that achieves high accuracy (96.8% AUROC) even with limited data.
◦
Provides details on the impact of hypertension on various cardiac function descriptors.
◦
Discovering new patterns consistent with existing physiological knowledge and suggesting the potential to advance understanding of hypertension.
◦
Effectively leveraging Transformer models for multimodal data integration.
◦
Reproducibility of patient stratification confirmed (mean absolute error within 5.7%).
•
Limitations:
◦
Further validation of generalizability is needed using a relatively small cohort (239 participants).
◦
Further research is needed to understand the interpretability of the model. (The physiological interpretation of the effects of specific descriptors is limited.)
◦
Further studies are needed to determine generalizability to other cardiovascular diseases.
