Cardiovascular disease (CVD) is a killer among noncommunicable diseases, making early detection and management of its risk factors crucial for improving patient outcomes. However, the continuous monitoring of these risk factors, both in and out of the clinic, presents significant challenges. In-clinic monitoring requires costly equipment and intensive labor, while out-of-clinic monitoring is often impractical due to device limitations.
Therefore, we propose to provide a cost-effective and automatic solution using representation learning methods and tackling the challenges in three folds:
* Creating representation learning methods of patient health status with health data to create robust methods of risk prediction for multiple short and long-term health outcomes.
* Creating continuous signal translation models that map remote data to health data, using questionnaires as labels for self-supervised learning.
* Developing time-varying risk prediction modeling by updating learned representations from Aim 1 with data in Aim 2 for improved risk prediction.
Multimodal contrastive learning is a representation learning method dedicated to aligning diverse data coupled with individual physiological differences into unified health embedding to downstream risk predictions. Translating remotely collected data into clinical quality health vitals can enable more flexible, time-varying representations of health status which would be used for predictive modeling of future adverse events. Therefore, we construct a signal translation framework that cardiac features from various remote data, consolidates them into cardiac health embeddings, and translates them into associated cardiac health data to enable both short- and long-term risk prediction and improve health outcomes