OBJECTIVES

Adult Community-Acquired Pneumonia (CAP) is one of the major causes of mortality and morbidity worldwide. The purpose of this study was to develop the risk of in-hospital mortality prediction model for adult community-acquired pneumonia patients.

METHODS

We performed a retrospective observation study for nationally representative samples of hospitalized patients with adult CAP using the administrative database from the Diagnosis Procedure Combination Per-Diem Payment System (DPC/PDPS) of 159,176 patients over 65 in 2016. The population was divided into a derivation cohort (70%, 94,693) and a validation cohort (30%, 40,582). The prediction model was developed using logistic regression (LR) for reference. LR model was compared with random forest (RF) and XGBoost(XGB). All administrative data up to day 1 of hospitalization were used in the model to predict in-hospital mortality. Overall model performance was assessed using AUC.

RESULTS

The baseline characteristics of the derivation and validation cohorts were similar. The XGB model showed the best performance (Derivation cohort: AUC 0.8800). AUCs for prediction models were 0.8476 for LS and 0.8321 for RF. In the XGB, LS and RF models, the relatively important features were disease severity (Fletcher-Hugh- Jones scale and Japan Respiratory Society community associated pneumonia severity index) and nursing need (motor functions).

CONCLUSIONS

The XGBoost model performed better than other models in developing for predicting the risk of in-hospital mortality, but are not good to surpass existing models because the data structures were sparse and categorical.