(K) Predicting Early Readmission and Death After Hospitalization for Chronic Obstructive Pulmonary Disease Exacerbation Using Machine Learning

Publication Number: 661

Background: While early readmission and death after hospitalization for chronic obstructive pulmonary disease exacerbation (ECOPD) are important adverse outcomes for both patients and society, prior studies found inconsistent risk factors and moderate prediction ability.

Objectives: To examine the ability of machine learning approaches on predicting early readmission and death after ECOPD hospitalization.

Methods: We used Taiwan’s National Health Insurance Research Database to identify ECOPD hospitalizations between 2015 and 2019. All hospitalizations were split into the development dataset (discharged between 2015 and 2018) and the validation dataset (discharged in 2019). Predictors were collected during the previous year and the index hospitalization, including patient characteristics, administrative health data, individuals’ linked laboratory tests, and environmental data. Prediction on 30-day readmission and death were derived in the development dataset using lasso regression, random forest, extreme gradient boosting (XGBoost), and neural network, with LACE index as the reference. Calibration plots and receiver operating characteristic (ROC) curves were evaluated based on the validation dataset. The top predictors were identified using in-built methods.

Results: There were 101,011 hospitalizations included in the development dataset and 17,565 in the validation dataset. The 30-day readmission rate was 29.1% and the 30-day mortality rate was 4.3% after ECOPD hospitalization. In the validation dataset, the best machine learning model (XGBoost) had an area under ROC curve (AUROC) of 0.721 (95% CI, 0.713-0.729) for 30-day readmission and an AUROC of 0.809 (95% CI, 0.794-0.824) for 30-day death, which were both significantly higher than the corresponding value for LACE index (0.651 and 0.641, respectively). All machine learning models had good calibration. The number of hospitalizations in the previous year and the lowest hemoglobin level during hospitalization were the top predictors in the XGBoost model for readmission and death, respectively.

Conclusions: Early readmission or death after ECOPD hospitalization can be predicted with acceptable to good accuracy using machine-learning algorithms and big data. Identification of important predictors can facilitate effective post-discharge care for these patients.