Objectives
Unplanned pneumonia readmissions increase patient morbidity, mortality and healthcare costs. Among pneumonia patients, the middle-aged and elderly (≥45 years old) have a significantly higher risk of readmission compared with the young. Given that the 14-day readmission rate is considered a healthcare quality indicator, this study is the first to develop survival machine learning (ML) models using emergency department (ED) data to predict 14-day readmission risk following pneumonia-related admissions.
Design
A retrospective multicentre cohort study.
Setting
This study used the Taipei Medical University Clinical Research Database, including data from patients at three affiliated hospitals.
Participants
11 989 hospital admissions for pneumonia among patients aged ≥45 years admitted from 2014 to 2021.
Primary and secondary outcome measures
The dataset was randomly split into training (80%), validation (10%) and independent test (10%) sets. Input features included demographics, comorbidities, clinical events, vital signs, laboratory results and medical interventions. Four survival ML models—CoxNet, Survival Tree, Gradient Boosting Survival Analysis and Random Survival Forest—were developed and compared on the validation set. The best performance model was tested on the independent test set.
Results
The RSF model outperformed the other models. Validation on an independent test set confirmed the model’s robustness (C-index=0.710; AUC=0.693). The most important predictive features included creatinine levels, age, haematocrit levels, Charlson Comorbidity Index scores, and haemoglobin levels, with their predictive value changing over time.
Conclusions
The RSF model effectively predicts 14-day readmission risk among pneumonia patients. The ED data-based model allows clinicians to estimate readmission risk before ward admission or discharge from the ED, enabling timely interventions. Accurately predicting short-term readmission risk might also further support physicians in designing the optimal healthcare programme and controlling individual medical status to prevent readmissions.