Training Prediction Models
An ensemble approach was employed in order to create a more stable and
reliable model resistant to outliers. Four different types of algorithms
were used: deep neural network, logistic regression, adaboost, and
random forest. We employed over-sampling (SVM Smote) and under-sampling
(Repeated Edited Nearest Neighbors) of the training data in order to
better balance the primary outcome within the dataset. For each type of
algorithm, 100 different models were trained using varying degrees of
data re-sampling to produce variability in each model’s underlying
training data. The resulting 400 algorithms were subsequently combined
into the final ensemble prognostic model.