Comparison of Chart Review and Administrative Data in Developing Predictive Models for Readmissions in Chronic Obstructive Pulmonary Disease

This study aimed to evaluate the performance of machine learning models for predicting readmission of patients with chronic obstructive pulmonary disease (COPD) based on administrative data and chart review data. The study analyzed 4327 patient encounters from the University of Chicago Medicine to assess the risk of readmission within 90 days after an acute exacerbation of COPD. Two random forest prediction models were compared. One was derived from chart review data, while the other was derived using administrative data. The data were randomly partitioned into training and internal validation sets using a 70% to 30% split. The 2 models had comparable accuracy (administrative data area under the curve [AUC]=0.67, chart review AUC=0.64). These results suggest that despite its limitations in precisely identifying COPD admissions, administrative data may be useful for developing effective predictive tools and offer a less labor-intensive alternative to chart reviews.