Improved Prolonged Length of ICU Stay Prediction using Stacked Machine Learning Models

Abstract

The most important component of the modern healthcare industry and the hospital are Intensive Care Units (ICUs). The optimal management of the resources with respect to the prolonged length of stay (pLOS) of patients in the ICUs is equally important. The objective of this study was to develop improved prediction models using stacked double layered machine learning models for the prolonged length of stay (pLOS) of ICU patients. The analysis of the eICU-CRD dataset used in the study suggests that Gradient Boosted Decision Tree (GBDT), RF (Random Forest), and DL (Deep Learning) demonstrate great potential and are recommended machine learning models for the intended application. The stacked layers approach utilizes a total combination of nine stacked models using GBDT, RF and DL. These models were compared based on area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), Brier score and calibration curve. Out of all the proposed models, the GBDT-GBDT layered model (AUPRC = 0.783, AUROC = 0.595, Brier score = 0.140) achieved the best performance. According to the calibration curve, RF-GBDT and DL-GBDT model performance was better among the nine models. All the proposed stacked models performed better than all the base layer models (GBDT, RF and DL), which implies the significance of using multi-layered models rather than single layer models for better predictions.