Machine Learning Pipeline for an Improved Medical Decision Support

Abstract

It has been projected that in recent years the amount of investment in automation is set to increase considerably while in the US alone it is set to reach $8 trillion. Given the rampant development in machine learning and artificial intelligence, there has been a gap between the advances in the field and the ethical and legal frameworks governing the field. This poses a problem which is set to grow in severity as more and more businesses and organization start adopting machine learning models. These models are essentially black boxes with even their developers having little knowledge as to how they came up with the results/decisions they did. In high stake decisions, this poses a serious problem as it leads to questioning of the trust ability of the model. There are a number of interpretability methods for different types of models and algorithms ranging from surrogate models to visual techniques for neural networks. However, there is often a trade-off between interpretability and model accuracy i.e., the more accurate a model tends to be, the less interpretable it becomes. Through this paper, we aim to demonstrate that interpretability of a model can actually be used to increase the model accuracy. We aim to achieve this by using the generated insights to update the model hyper parameters and comparing the accuracy scores before and after the update. We have used a breast cancer prediction dataset which we obtained from http://kaggle.com. Therefore, the aim of this paper is to create a machine learning pipeline which improves the model accuracy using insights from interpretability.

Keywords: neural networks, model accuracy, surrogate models, interpretability.