deep neural network quantifies individual cardiovascular disease risk

Chao, Hanqing
Thumbnail Image
Other Contributors
Kalra, Mannudeep
Hahn, Juergen
Yan, Pingkun
Wang, Ge
Issue Date
Biomedical engineering
Terms of Use
This electronic version is a licensed copy owned by Rensselaer Polytechnic Institute (RPI), Troy, NY. Copyright of original work retained by author.
Full Citation
Cardiovascular diseases (CVDs) are the leading cause of global mortality, emphasizing the urgent need for accurate, accessible, and interpretable risk assessment methodologies. This dissertation introduces a deep learning pipeline aimed at individualized CVD risk quantification. The proposed pipeline integrates risk factors derived from low-dose computed tomography (LDCT) with conventional tabular risk factors. It seeks to harness the rich, albeit noisy, data provided by LDCT images effectively. This pipeline is designed to contribute to the clinical field by offering CVD risk assessments that are accessible owing to their reliance on widely available LDCT imaging and readily available patient information, and that are also interpretable for medical professionals, thus facilitating timely interventions. The venture of applying deep learning to CVD risk assessment is relatively unexplored, and in the process of crafting this pipeline, three key questions emerged: 1) Can LDCT provide sufficient information for deep learning models to effectively quantify CVD risk? 2) How can risk factors be embedded in a high-dimensional space to maximize deep learning capabilities? 3) How can diverse risk factors be integrated in an interpretable manner? These questions mark critical barriers in the path of developing a viable deep learning pipeline for CVD risk assessment. The dissertation is structured around three specific aims, each corresponding to one of these questions. Aim 1 introduces Tri2D-Net, a deep learning model that efficiently extracts CVD-related features from 3D LDCT images using three orthogonal 2D views, demonstrating impressive performance in CVD screening and risk quantification. Aim 2 presents the Regression Metric Loss (RM-Loss), a novel loss function that guides a deep learning model to generate interpretable high-dimensional embeddings of risk factors, thereby bridging the gap between the low-dimensional space of risk factors and the high-dimensional space where deep learning thrives. Aim 3 brings forward MIX-CVD, a model based on the Mixture of Experts (MoE) approach. MIX-CVD adeptly analyses the complex relationships among risk factors, generating highly accurate predictions of CVD risk while maintaining interpretability through an adaptive assignment of explicit weights to each risk factor based on an individual's condition. Together, these components form a pipeline that successfully addresses the key questions, providing an effective, accessible, and interpretable tool for individualized CVD risk quantification. By offering accurate predictions that are easy to interpret and rely on widely accessible LDCT imaging, this work could potentially advance the current state of CVD risk assessment, underscoring the potential of deep learning in enhancing clinical decision-making and patient care.
School of Engineering
Dept. of Biomedical Engineering
Rensselaer Polytechnic Institute, Troy, NY
Rensselaer Theses and Dissertations Online Collection
Restricted to current Rensselaer faculty, staff and students in accordance with the Rensselaer Standard license. Access inquiries may be directed to the Rensselaer Libraries.