A novel Garson Pruned ELM model with recursive feature selection for healthcare data analysis

Abstract In healthcare settings, selecting the optimal features for model implementation is a challenging task due to high dimensionality of dataset. Several methodologies have been devised to address this issue using medical records. Nevertheless, many physiological tests, such as lab test and vital test, are laborious and need the implementation of diverse machine learning models. In this paper, we offer a computational solution using recursive feature elimination coupled with extreme learning machine algorithm to predict the mortality of intensive care patients. The proposed work uses Garson algorithm to prune hidden neurons and achieved higher accuracy of 94.45% for WiDS Datathon data and 91.2% for MIMIC-III data. Empirical studies conducted on various widely used classification benchmark problems and datasets obtained from the physionet database demonstrate that the pruned method proposed in this study outperforms traditional algorithms in automatically identifying the optimal number of features and hidden nodes and exhibiting superior generalization performance.