چكيده به لاتين
These days the high cost of intensive care units and lack of equipment in the hospitals, especially in developing countries such as Iran is an important challenge. So to allocating intensive care unit facilities to qualified patients it is essential to predicting the mortality of the patient in the entering time. In this research, we investigated this issue with the help of machine learning methods. Instead of focusing on complicated algorithms such as deep learning, we were trying to use preprocessing methods to clean the dataset and then feed this data to a two-level stacking method which consists of state of the art algorithms like extreme gradient boosting. By decreasing model complexity this suggested procedure could yiled superior results in some metrics. This metrics include accuracy=0.81±0.041, AUPRC=0.59±0.051, F1=0.78±0.048, matthews correlation coefficient=0.41±0.051. According to importance of the last 3 metrics in the case of imbalanced datasets,compared to the latest researches, suggested model performs well. Also in this research we ranked different features of the ICU dataset. GCS which relates to patients awareness factor, is the most important factor that relates to patients outcome status.
