Machine learning methods for predicting days to discharge in Intensive Care Units patients

Identificador:  TDX:4318

Handle: https://hdl.handle.net/20.500.11797/TDX4318

Autors:  Cuadrado Gómez, David

Resum:
ICUs provide specialized care for critically ill patients, with constant monitoring and dedicated healthcare teams until stability for safe transfer or discharge. Inadequate planning can lead to reduced care quality, delays, and increased morbidity and mortality. To address this issue, reliable computer tools are needed to predict ICU stays, aiding in bed, medication, and personnel planning. Two predictive approaches exist: static tools estimating length of stay (LOS) and dynamic tools predicting days to discharge (DTD) as treatment progresses. This thesis aims to predict ICU patient discharge accurately using Machine Learning and private data from Joan XXIII Hospital and the eICU public dataset for diverse patient representation. To enhance accuracy, it is critical to understand the heterogeneity of patients. The thesis proposes four measures to quantify heterogeneity and two approaches to analyze LOS and DTD: biomarker identification and phenotype recognition. Three algorithms (Random forest, XGBoost, and lightGBM) were used to create predictive models for DTD and LOS, and combining them into a hybrid model improved accuracy. The LOS model performed well initially, while the DTD model excelled later in the stay. The results achieved a root mean square error (RMSE) and mean average error (MAE) within acceptable ranges