Identifying longitudinal disease trajectories and their clinical associations in 146,000 individuals

Pankhuri Singhal1 (BS), Lindsay Guare1 (BS), Anastasia Lucas1,2 (BS), Colleen Morse1 (DPT), Marta Byrska-Bishop3 (PhD), Marie A. Guerraty1 (MD), Dokyoon Kim1 (PhD), Anurag Verma1,2 (PhD) , and Marylyn D. Ritchie1 (PhD), 1 University of Pennsylvania Perelman School of Medicine, Philadelphia, PA 2Corporal Michael Crescenz VA Medical Center, Philadelphia, PA 3New York Genome Center, New York, NY

Many chronic diseases are associated with hypertension, yet we are unable to determine which future disease states are likely in patients with hypertension. Interrogation of the comorbidity architecture of hypertension using individual-level electronic health record (EHR) data is needed to improve precision medicine strategies for predicting disease trajectories after a hypertension diagnosis. Currently, there is no optimal way to model temporal patterns in longitudinal data by parsing meaningful clinical signals in patient charts. In order to extract "signpost" events (labs, medications, ICD codes) that are most predictive of an individual's disease trajectory, determining correlation structure of clinical variables and identifying significant directionality and order of events is essential. We developed a framework to identify disease trajectories in Penn Medicine EHR data for 146,654 individuals with hypertension (I10*). A disease co-occurrence network was generated using ICD codes for each individual appearing after I10* (3,642 codes). The Ising Model, a probabilistic graphical model, estimated pairwise co-occurrence between each condition considering all other conditions. This generated an undirected comorbidity network of significant ICD pairs. Then, directionality was determined by calculating relative risk (RR) in both directions per pair, yielding 23,729 significant pairs (RR > 1). Each individual was assigned a trajectory, which was assembled using an iterative pairwise RR framework starting from I10* to subsequent ICD neighbors occurring within 5 years. This approach uniquely identified conditions that functioned as markers of disease path. Thus far, our results have yielded various clinical trajectories; for example, some individuals start with type 2 diabetes, then develop chronic kidney disease complications while others develop rheumatic or glaucoma complications. Further simulation experiments are underway to benchmark and clinically validate our approach to train a model to predict comorbid conditions and disease trajectories that may present in the future for individuals with hypertension.