Multimodal representation learning for personalized cardiovascular care : predicting disease progression and treatment from clinical data

Cardiovascular disease remains a leading cause of global morbidity and mortality, with timely diagnosis and management often limited by resource constraints and heterogeneous disease trajectories. This thesis presents a series of deep learning-based multimodal representation learning frameworks for personalized cardiovascular disease management, using routinely collected and non-invasive clinical data. The thesis first introduces TREAT-Net, an acute coronary syndrome treatment prediction framework that provides echocardiogram analysis through electronic medical record-guided cross-attention. By contextualizing cardiac imaging with complementary patient information, the model enables personalized risk stratification at the point of initial presentation. Building on this work, TREAT-Netv2 incorporates regional wall motion-informed representations and transformer-based multiple instance learning for enhanced clinical alignment, strengthening decision support for particularly challenging presentations such as non-ST-elevation myocardial infarction and unstable angina. Extending beyond acute treatment prediction, AS-TIME presents a time-conditioned framework for predicting aortic stenosis progression using a single baseline echocardiogram study, cardiologist report, and elapsed time to follow-up. The model learns time-aware multimodal representations through hierarchical aggregation, adaptive modality gating, and mixture-of-experts architecture to capture heterogeneous progression trajectories. Across datasets comprised of nearly 20,000 cardiovascular cases, these frameworks consistently outperform conventional baselines, demonstrating the value of clinically aligned multimodal representation learning for cardiovascular prediction. Together, this thesis strives to enable earlier, non-invasive, and personalized clinical decision support with the potential to reduce healthcare disparities in resource-limited settings.