Integrating Multi-Modal Data and Biomechanics in COPD: Toward Robust and Interpretable Biomarkers for Disease Subtyping and Progression

Abstract Chronic Obstructive pulmonary disease (COPD) is a devastating, progressive disease affecting 13 million Americans and 300 million globally. The disease is highly heterogeneous, with both behavioral risk (e.g., smoking) and genetic predisposition. Patient-specific biomechanics also play a crucial role in determining the progression of the disease. While the global lung function test is used for clinical diagnosis, it is insufficient to understand the underlying mechanism of the disease. High-resolution CT imaging is also used to quantify the two major phenotypes of the disease (e.g., emphysema and airway inflammation). However, current classical radiomics measures lack the sensitivity to quantify the progression. While Deep Learning radiomics have shown promising results, they are unimodal, lack interpretability, and are sensitive to data acquisition protocols such as scanner type. We aim to extend our prior project, which demonstrated that deep learning radiomics features trained via a self- supervised approach extract features more generalizable than traditional supervised methods. Our novel radiomics features predict disease progression effectively and when combined with blood transcriptomics, facilitate the identification of distinct disease subtypes. We hypothesize that our multi-modal deep learning approach, which integrates imaging, genomic data, and lung biomechanics, will enhance understanding of COPD progression with the capability to provide robust, interpretable, and accurate clinical predictions. In this proposal, we seek to develop new radiomic features that derive insights from CT images and are informed by genetic and genomic data, enhancing their robustness by integrating biological insights into imaging patterns (Aim 1). We train the model on a vast dataset of over 65,000 CT images to enhance its robustness across various imaging protocols, including low-dose and high-dose CT scans. We will develop a novel high- dimensional mediation analysis technique to discover the causal pathways of the disease subtypes. In Aim 2, we will develop a method to estimate the biomechanical properties of the parenchyma, more specifically tissue stiffness, from CT images using a novel Biomechanically-Informed Neural Network, which is highly interpretable. We will validate our method ex-vivo with lung Cristal ribcage, a transparent ribcage model allowing for multiscale optical imaging from the whole organ down to single cells. The technology developed in this proposal can potentially impact beyond COPD to other pulmonary diseases that impact the biomechanical properties of the extracellular matrix.