Fusing Echocardiography Images and Medical Records for Continuous Patient Stratification.

Area of Science:

• Cardiovascular Medicine
• Artificial Intelligence
• Medical Informatics

Background:

• Cardiac function descriptors from echocardiography offer detailed patient insights.
• Hypertension presents a complex, continuous pathological spectrum.
• Integrating multimodal data (clinical records, echocardiograms) is crucial for comprehensive patient assessment.

Purpose of the Study:

• To develop a novel deep learning method for characterizing hypertension using multimodal data.
• To leverage Transformer models for learning patient representations from diverse data sources.
• To stratify hypertensive patients along a pathological continuum for improved understanding.

Main Methods:

• Utilized Transformer models adapted for tabular data (XTab foundation model).
• Employed modality-specific projections for echocardiographic and clinical data.
• Formulated patient stratification as an ordinal classification task to model pathological continuum.

Main Results:

• Achieved high performance (96.8% AUROC) with limited data (<200 samples).
• Demonstrated reproducible patient stratification across training runs (5.7% MAE).
• Identified novel patterns in cardiac descriptors related to hypertension, some aligning with existing knowledge.

Conclusions:

• The proposed method effectively integrates multimodal data for cardiovascular pathology characterization.
• Transformer models show promise in analyzing complex clinical data for disease stratification.
• Findings provide deeper insights into hypertension's impact on cardiac function and offer avenues for future research.