”A Neural Network Approach to Predicting Mortality in Pediatric Intensive Care”
Kaja Horvat presents her Master’s thesis
In this Master’s thesis, data from a pediatric intensive care unit for children and youth at the Skåne University Hospital in Lund, Sweden are used for predicting child mortality and investigating if the PIM2 model can be improved using neural networks and blood gas test results. In the analysis, 1155 patient admissions were included after removing patients without tests within the first hour of admission. Three different multilayer perceptron neural networks were created. To choose the variables and the parameters for the networks, a grid search and uni- and multivariate logistic regression were used; 5-fold cross validation was applied to ensure the networks’ adequate performance on new data. The results show that neural networks can predict child mortality in the ICU and are a feasible alternative to PIM2. Using blood gas variables improves the predictions, especially in addition to other physiological variables. Further research is needed to conclude if the improvement outweighs the extra computational effort, and should focus on a larger dataset with a wider grid search to find the optimal parameters for the networks.