Machine learning-based prediction for 30-day unplanned readmission in all-types cancer patients

Unplanned readmission is a key indicator of quality of medical care and effectiveness of healthcare resource management. We aimed to develop the prediction model of 30-day unplanned readmission of cancer patients that can be applied to cancer hospitals. We obtained 15,877 patients and 4,437 variables obtained from the Clinical Data Warehouse (CDW) in the National Cancer Center in Korea and used conventional machine learning (ML) models to predict unplanned readmission. XGBoost showed the best performance with F1-Score (0.74), Accuracy (0.84), Precision (0.85), AUROC (0.90), AUPRC (0.86) respectively. With XGBoost, we conducted additional experiments on the test set by reducing the number of top features to develop a general model with common features. Using the top 30 features, the performance of F1-score (0.72) is equivalent to using all variables. The 30 variables are generally obtainable from EHR of cancer hospitals. Our model can be expected to apply to EHR of other cancer hospitals. Our result may help minimize the risk of unplanned readmission and improve the quality of care for cancer patients.