Applying machine learning algorithms to electronic health records to predict pneumonia after respiratory tract infection

Objectives: To predict community acquired pneumonia after respiratory tract infection (RTI) consultations in primary care by applying machine learning to electronic health records. Study design and Setting: A population-based cohort study was conducted using primary care electronic health records between 2002 to 2017. Sixteen thousand two hundred eighty-nine patients who consulted with RTIs then subsequently diagnosed with pneumonia within 30 days were compared with a random sample of eligible RTI patients. Variable selection compared logistic regression, random forest and penalized regression models. Prediction models were developed using classification and regression trees (CART) and logistic regression. Model performance was assessed through internal and temporal validations. Results: Older age, comorbidity, and initial presentation with lower respiratory tract infection (LRTIs) were identified as the main predictors of pneumonia diagnosis. Developed models achieved good discrimination accuracy with AUROC for the logistic regression model being 0.81 (0.80, 0.84) and 0.70 (0.69, 0.71) for CART during internal validation, and 0.80 (0.79, 0.81) vs. 0.68 (0.67, 0.69) for temporal validation. Conclusion: From a large number of candidate variables, a small number of predictors of pneumonia were consistently identified through machine learning variable selection procedures. Logistic regression generally provided better model performance than CART models. (c) 2022 Elsevier Inc. All rights reserved.

Automated summary

This study aimed to predict community acquired pneumonia after respiratory tract infection (RTI) consultations in primary care using machine learning applied to electronic health records. The main predictors of pneumonia diagnosis were identified as older age, comorbidity, and initial presentation with lower respiratory tract infection (LRTIs). The developed models achieved good accuracy in internal and temporal validations.