Abstract
This study was carried out by the New York State Department of Health, between 2012 and 2016. This experiment relates to six supervised machine learning methods: Support Vector Machine (SVM), Logistic Regression (LR), Gradient Boosting (GB), Random Forest (RF), Deep Learning (DL) and the Ensemble Model, all of which are used in the prediction of infant mortality. This experiment applied ensemble model that concentrated on assigning different weights to different models per output class in order to obtain a better predictive performance for infant mortality. Efforts were made to measure the performance and compare the classifier accuracy of each model. Several criteria, including the area under ROC curve, were considered when comparing the ensemble model (GB, RF and DL) with the other five models (SVM, LR, DL, GB and RF). In terms of these different criteria, the ensemble model outperformed the others in predicting survival rates among infant patients given a balanced data set (the areas under the ROC curve for minor, moderate, major and extreme were 98%, 95%, 92% and 97% respectively, giving a total accuracy of 80.65%). For the imbalanced dataset, (the areas under the ROC curve for minor, moderate, major and extreme were 98%, 98%, 99% and 99% respectively, giving total accuracy increased to 97.44%). The results of the experiments used in this dissertation showed that using the ensemble model provided a better level of prediction for infant mortality than the other five models, based on the relative prediction accuracy for each model for each output class. Therefore, the ensemble model provides and extremely promises classifier in terms of predicting infant mortality.