Predicting Post-Covid Pulmonary Fibrosis with Explainable Deep Learning and Optimal Biomarker Discovery

PROJECT SUMMARY Since the beginning of COVID-19 (SARS-CoV-2) pandemic, we have seen a number of prominent variants, and two of them are very recent. COVID-19 continues to affect millions of people: over 771 million people were affected and over 6.9 million deaths occurred (as of October 2023). A significant portion of patients who suffer from severe COVID-19 disease are at risk for developing long-term Post-Acute Sequelae (PASC) with pulmonary fibrosis. Identifying patients who are most at risk for post COVID-19 pulmonary fibrosis could allow early intervention with anti-fibrotic drugs and other potential treatment strategies. Given the extent of people infected with COVID-19, post COVID-19 pulmonary fibrosis is a significant cause of mortality and morbidity, yet there is no strategy developed to address this unmet critical clinical need. Hence, the overall goal of this proposal is to address this need for predicting post COVID-19 pulmonary fibrosis at an early stage by developing novel explainable deep learning (DL) algorithms on multimodal data (combined imaging and electronic health record (EHR) data) from multiple centers. Main Hypothesis: The proposed explainable DL algorithms will identify patients at risk for development of post COVID-19 related pulmonary fibrosis at early phases by multimodal data with high accuracy. In Aim 1, we will identify imaging and EHR characteristics associated with post COVID-19 fibrosis from patient data gathered from Northwestern, Columbia, NIH, and UPenn. We will automatically extract imaging (CT scan) features in two ways: (i) pulmonary analysis (PA) and (ii) body composition analysis (BCA). For each patient, we will collect EHR data consisting of demographic, clinical (including vital, medication, vaccination, and comorbidity) and laboratory information. We expect to retrospectively collect a balanced dataset of 1,150 initial CT scans and associated EHR data. In Aim 2, we will develop Transformer embedded explainable capsule networks (X-TCaps) for prediction of post COVID-19 pulmonary fibrosis. We will integrate our newly established visual explanation algorithm (called IBA) into the machine-generated results in addition to radiographical explanations, and PA and BCA features captured by X-TCaps. In Aim 3, we will employ our established optimal biomarker (OBM) method to determine the most potent features (biomarkers) from EHR and imaging data to predict post COVID-19 pulmonary fibrosis at the highest accuracy possible. PASC is a massive emergency and very little is known about it. Once accomplished, our proposed framework will provide early prediction of post COVID-19 pulmonary fibrosis and determine biomarkers to understand pulmonary fibrosis better. Our study is innovative as no previous study has investigated post COVID-19 pulmonary fibrosis, which is paramount to developing a robust knowledge database and informing clinical practice in this area. With this project, we will provide mechanistic understanding of post COVID-19 pulmonary fibrosis, which is central in determining therapeutic options and will ultimately play a role in lung transplant considerations in the long term.