Leveraging Machine Learning and Real-World Data to Predict Chronic Obstructive Pulmonary Disease Exacerbations

Reynold A Panettieri Jr,1 Jason Roy,2 Natalia Gontarczyk Uczkowski,1 Allison Tyler,3 Jason Attanucci,3 Thomas G O’Riordan,4 Kristin Kahle-Wrobleski4 1Rutgers Institute for Translational Medicine and Science, New Brunswick, NJ, USA; 2Rutgers School of Public Health, New Brunswick, NJ, USA; 3Deep 6 AI, Inc, Pasadena, CA, USA; 4US Value Evidence and Outcomes, GSK, Philadelphia, PA, USACorrespondence: Reynold A Panettieri Jr, Rutgers Institute for Translational Medicine and Science, 89 French Street, Suite 4211, New Brunswick, NJ, 08901, USA, Email [email protected]: Previously, we reported that applying artificial intelligence and natural language processing to electronic health record (EHR) data can identify patients at risk of chronic obstructive pulmonary disease (COPD) exacerbations, based on clinical attributes identified in COPDGene.Purpose: Building on these data and using real-world data, we established a predictive model for identifying patients at risk of COPD exacerbations within 24 months of their initial COPD diagnosis.Methods: Structured and unstructured data were obtained from Epic EHR data. Summary statistics for independent variables, including age, gastroesophageal reflux disease, coronary artery disease, congestive heart failure, cor pulmonale, asthma, dyspnea, smoking status, number of comorbidities, and blood eosinophil counts, were calculated. Bivariate associations with COPD exacerbations were calculated using odds ratios and 95% confidence intervals. A multivariable prediction model using the flexible machine-learning approach, Bayesian Additive Regression Trees (BART), was then developed. Model performance was assessed using receiver operating characteristic (ROC) curves and area under the ROC curve (AUC).Results: Of the 3007 patients with COPD as a primary diagnosis, 886 had a COPD exacerbation within 24 months. In the bivariate logistic regression analyses, strong associations (odds ratio > 1.5; P < 0.05) existed between COPD exacerbation and cor pulmonale, moderate and severe dyspnea, and number of comorbidities (≥ 4 vs 0). In the BART model, the predictors that were selected most for the branching-tree analyses were eosinophil count, pack years, and moderate dyspnea (in order of most selected). The AUC derived from our BART model was 0.69.Conclusion: Eosinophil count and dyspnea were identified as important predictors of exacerbations. Our data suggest that active monitoring of eosinophil counts and selected patient-reported experiences of dyspnea may identify patients at risk of exacerbations, enabling clinicians to tailor therapies to improve health outcomes among patients with COPD.Plain Language Summary: For patients with chronic obstructive pulmonary disease (COPD), exacerbations (symptom flare-ups) represent a significant health burden. The ability to accurately predict exacerbations would help to identify patients at risk of future exacerbations, so they can receive appropriate treatment. Previously, we applied artificial intelligence and natural language processing (a form of artificial intelligence that helps computers to process human language) to electronic health record (EHR) data to identify patients at risk of COPD exacerbations. Building on this foundation and using data from real-life studies, we developed a model to help identify patients at risk of COPD exacerbations within 24 months of their initial COPD diagnosis. We used patient data from the EPIC EHR database. Factors included in the model to predict exacerbation occurrence were age, gastroesophageal reflux disease, coronary artery disease, congestive heart failure, pulmonary heart disease, asthma, shortness of breath (SOB), smoking status, number of medical conditions aside from COPD, pulmonary function testing, pack years (a unit used to measure how much a person has smoked over time), and blood eosinophil counts. We reported the percentage of patients with each factor. We included 3007 patients with an average age of 70.5 years. Using this model, we found that blood eosinophil count and SOB were important predictors of exacerbations for patients with COPD. These results suggest that active monitoring of blood eosinophil counts and patient-reported experiences of SOB may help to identify patients at risk of exacerbations, which could help doctors to tailor patients’ treatment to improve their health.Keywords: eosinophilia, lung diseases, morbidity, natural language processing, symptom flare-up