Achieving 90% Recall with a Deep Neural Network: How SMOTE and ModelCheckpoint Callbacks Improved Results on a Skewed Classification Problem