A Hybrid Feature Selection for Improving Prediction Performance with a Brain Stroke Case Study

Abstract

In the contemporary era, artificial intelligence (AI) is making strides into every conceivable field. With advancements in place, there have been applications of machine learning (ML) in healthcare domain. Particularly for diagnosis of diseases with data-driven approach, ML algorithms are capable of learning from training data and make predictions. Many supervised ML algorithms came into existence with varied capabilities. However, they do rely on quality of training data. Unless quality of training data is ensured, they tend to result in mediocre performance. To overcome this problem, feature engineering or feature selection methods came into existence. From the literature, it is understood that feature selection plays crucial role in improving performance of prediction models. In this paper, a hybrid feature selection algorithm is proposed to leverage performance of machine learning models in brain stroke detection. The algorithm is named as Hybrid Measures Approach for Feature Engineering (HMA-FE). It returns best features that could contribute toward prediction of class labels. A prototype application is built to demonstrate the utility of the proposed framework and the underlying algorithms. The performance of prediction models are evaluated without and with feature engineering. Its empirical results showed the significant impact of proposed feature engineering on various brain stroke prediction models. The proposed framework adds value to Clinical Decision Support System (CDSS) used in healthcare units by supporting brain stroke diagnosis.