PASC Pulmonary Fibrosis Prediction with Deep Learning and Multi-Modal Data

The overall goal of this proposal addresses this urgent unmet need by developing novel explainable deep learning (DL) algorithms to identify patients at risk for development of PASC pulmonary fibrosus(PF) using multimodal data from multiple centers by combining imaging (initial CT scan) and electronic health record (EHR). Proposed solution: We hypothesize that deep learning (DL) algorithms to analyze multimodal data (initial CT scans and EHR data) can identify patients at the highest risk for long-term sequelae for development of PF, who would benefit most from early intervention [3-6]. To test this, we will develop DL algorithms to identify predictive biomarkers of PASC PF from imaging and EHR data. To overcome barriers to adoption of DL, we will use explainable algorithms that enable human understanding of their output. Our advanced image analysis techniques will automatically conduct (1) pulmonary analysis (PA): lung and lung lobe volumetry, parenchymal properties from CT scans (intensity & texture), identification of abnormal imaging patters, and airway wall thickness measurements, and (2) body composition analysis (BCA): volume, intensity, and texture properties of subcutaneous fat, visceral fat, skeletal muscle, and osseous (bony/skeletal) structures. Our explainable DL algorithm will associate its output (PASC PF or not) with findings from BCA, PA, and radiographical findings from CT scans. Our optimal biomarker (OBM) selection method will combine imaging and EHR data to further enhance prediction accuracy through complementary strengths of imaging and non-imaging features. Our study will use a multi-center setting, addressing heterogeneity among institutions and offering opportunities for model (prediction) generalizability.