Ensemble machine learning as an approach for improvement of classification on very small datasets

Abstract

Machine learning is a branch of artificial intelligence most commonly used for solving problems related to classification and regression analysis through the use of supervised learning approaches. Machine learning models require high quality and a sufficient quantity of data to produce good results. This paper investigates an approach which incorporates ensemble learning to increase prediction capabilities in cases in which a very limited amount of data is available for training. The ensemble model was trained on a patient fractional flow reserve biomarker dataset and had a goal of classifying patients into risk classes based on their risk of suffering an acute myocardial infarction. The ensemble model was built by aggregating multiple random forest classification models which were trained with different combinations of training and test data to improve the prediction accuracy over the use of a single random forest model. Final ensemble achieved a prediction accuracy of approximately 71.3% which was an immense improvement over the 36% prediction accuracy of a single random forest classification model.