| dc.description.abstract | Critical congenital heart disease (cCHD) is present in two to three of every 1,000 newborns. Children diagnosed with cCHD are admitted to the pediatric intensive care unit (PICU) and closely monitored to ensure the highest possible quality of healthcare. During this period large quantities of continuous data streams are collected. The aim of this study is to analyze these large quantities of data using machine learning techniques such as random forest and boosting to provide insights in detecting deterioration by classifying periods as stable and unstable. The data consisted of 86 patients with information on five vital signs: heart rate, respiratory rate, invasive mean blood pressure, oxygen saturation and regional cerebral oxygen saturation. Pre-processing steps were necessary to transform the data and generate artificial labels for model training. Using the pre-processed data, hyperparameter tuning was performed, and final models were created. Based on these models, classifications were made on a left-out dataset consisting of nine patients. These model classifications are compared with a clinical classification established by a medical expert. The findings revealed an accuracy range of 64.4% to 87.1%, a sensitivity range of 66.0% to 97.3%, and a specificity range of 23.1% to 94.3%. These numbers demonstrate generally favorable accuracy and sensitivity scores. However, some models had very low specificity scores, indicating large amounts of true unstable periods were not classified as unstable. Relying on such a model would result in missing many critical situations at the PICU. Nevertheless, some models showed great potential. This study highlights that machine learning techniques such as random forest and boosting can be used to provide insights in detecting deterioration in patients. Consequently, it is recommended to explore the best performing models further and assess if further improvements are possible. Additionally, it is important to analyze the reasons behind certain incorrect classifications to enhance the understanding of the model. Finally, choosing a best model depends on the task at hand and should be considered carefully. | |
