
Research Article
Machine Learning-Based Predictors for ICU Admission of COVID-19 Patients
@INPROCEEDINGS{10.1007/978-3-031-06371-8_38, author={Nagham Alhawas and Serkan Kartal}, title={Machine Learning-Based Predictors for ICU Admission of COVID-19 Patients}, proceedings={Science and Technologies for Smart Cities. 7th EAI International Conference, SmartCity360°, Virtual Event, December 2-4, 2021, Proceedings}, proceedings_a={SMARTCITY}, year={2022}, month={6}, keywords={Covid-19 Vital information EHR Intensive care unit}, doi={10.1007/978-3-031-06371-8_38} }
- Nagham Alhawas
Serkan Kartal
Year: 2022
Machine Learning-Based Predictors for ICU Admission of COVID-19 Patients
SMARTCITY
Springer
DOI: 10.1007/978-3-031-06371-8_38
Abstract
The burden on the health sector has increased when covid-19 was declared as a critical pandemic, making the decision-taking more crucial. This study aimed mainly to build predictors to aid in making decisions for severe patients to predict whether a patient has to be admitted to the intensive care unit (ICU) based only on the vital records. Statistical techniques were used on the electrical health records (EHR) that were accessible for the covid-19 patients. Samples were processed and then extracted based on criteria that support data imputation. Then, several feature selection techniques were utilized based on the field knowledge, Pearson correlation coefficient, and finally by taking the permutation importance of a hypothetical model to retain features that have the highest relationship with the target variable. Then two versions of data were obtained as stateless and grouped data with and without feature selection which were used to build models with various machine learning algorithms; logistic regression, linear support vector machine SVM, SVM with radial basis function RBF, and artificial neural network ANN. In this respect, the models reached an accuracy of more than 95% in most of the used classifiers and the best one scored is RBF-SVM with accuracy up to 98% and achieve 0.95 areas under curve (AUC) performance. These results indicate that trustworthy models were built to fulfill the high demand for accuracy that is more or less commensurate with the cost of accuracy in the health sector relying only on vital information.