Machine learning for mortality prediction in community-acquired pneumonia emergency admissions

Community-acquired pneumonia (CAP) is the leading cause of mortality in UK hospitals. Severity scores are widely utilised to predict mortality but suffer from variable performance. This study aims to determine the potential of machine learning (ML) models for predicting mortality in admitted patients. Data was obtained from patients admitted to University Hospitals of Leicester from 2016 to 2018 with an admissions primary diagnosis of CAP. Mortality was defined as positive in cases where patients died in hospital or within 30 days of discharge. Initial features, including demographics, vital signs, and blood tests were taken from linked hospital records. Feature importance was evaluated among three different models: statistical significance, XGBoost (XGB) and Random Forest (RF). For classification, Support Vector Machines (SVM), XGB and RF models were fine-tuned using the entire data sets and divided by age tiers, with area under the curve (AUC) implemented as a metric to compare the performance of the models. A total of 9237 admissions were included of which 2138 (23.1%) were deaths relevant to this study, 4819 (52%) were females, and 7298 (79%) were over 65. Initial results showed XGB models produce an AUC of 0.74, with a reduction to 0.691 and 0.70 using SVM and RF, respectively. When data was stratified into three age groups [65 – 79] (31%), [80 – 89] (30%) and [> =90] (12%) the best model for each group produced AUCs of 0.81. Our use of ML models has produced promising results, relative to existing aggregate scores. Further optimisation using more sophisticated time-series based data models seems feasible and holds promise to improve CAP outcome by real time deployment.