Two step Outlier Detection to enhance XGBoost Accuracy in Predicting Length of Stay for Type 2 Diabetic Patients

The complexity of type 2 diabetes mellitus length of stay data increases the risk of outliers. Outliers negatively impact model performance in both classification and regression models. However, there is limited research combining Isolation Forest and Boxplot outlier detection with ensemble methods. This study aims to evaluate the combination of isolation Forest and Boxplot as outlier detection methods for improving extreme gradient boosting performance in predicting T2DM length of stay. The data used for the analysis consisted of 587 T2DM patients. Using two-step outlier detection with both isolation forest and boxplot, MAPE has decreased by 1.4%. XGBoost with Isolation Forest and Boxplot outlier detection outperformed the standalone XGBoost model and XGBoost with only isolation forest. In predicting complex data, two-step outlier detection provides better result than using a single outlier detection methods