OBJECTIVES: Hospital readmissions for pneumonia are a growing concern in the US. This study evaluated the prediction models for hospital readmission in pneumonia using machine learning (ML) methods.

METHODS: A retrospective study using the 2016 Healthcare Cost and Utilization Project- National Readmission Database (HCUP-NRD) was conducted. The study included patients aged >=18 years with index admissions for pneumonia identified using ICD-10-CM codes as the principal diagnosis. The ML algorithms, including the Least Absolute Shrinkage and Selection Operator (LASSO), random forest, and RuleFit were used to develop prediction models for 30-day all-cause readmissions. Models were trained on randomly partitioned 50% of the data and evaluated using the remaining dataset. Model hyperparameters were tuned using the 10-fold cross-validation on the resampled training dataset. The model performances were evaluated using Area Under the Curve (AUC), accuracy, precision, recall, sensitivity, specificity, and F1 score.

RESULTS: There were 400,690 hospital admissions from 372,293 unique patients with a pneumonia diagnosis. The 30-day readmissions rate was 12.97%. Most of the re-admitted patients were >=65 years old (65%), female (52%), and with Medicare (70%). Among the ML methods, the random forest model achieved the best model performance with an AUC of 0.7139 (accuracy 65%, precision 21%, recall 80%, sensitivity 62%, specificity 80% and F1 score 37%), followed by the RuleFit model with an AUC of 0.6119 (accuracy 60%, precision 19%, recall 63%, sensitivity 59%, specificity 63% and F1 score 29%) and the LASSO model with an AUC of 0.6068 (accuracy 62%, precision 19%, recall 59%, sensitivity 63%, specificity 59% and F1 score 29%).

CONCLUSIONS: The model performance of various ML methods varied, with the random forest model providing the best performance for predicting the risk of 30-day readmission in pneumonia. More research is needed to evaluate ML methods for other diseases of interest.