| Summary: | Introduction: Identifying and analyzing mortality risk factors will lead to more accurate planning and prevention in health platforms. This research provides models for predicting mortality in the intensive care unit with machine-learning techniques. Method: We extracted data from 1400 patients' medical records admitted to the intensive care unit of Imam Hossein Hospital in Tehran. Period of inclusion was January 1, 2018, to February 31, 2020. In this study, we used twelve predictor algorithms for data analysis, including: SVM, k-nearest neighbor, decision tree, logistic regression, random forest, MLP, RBF, Gradient Boosting, Fast Large Margin, Rule Model, Naive Bayesian, and Deep learning. We reported the performance of algorithms based on the precision, accuracy, sensitivity, specificity, and AUC. Results: The findings showed that algorithms according to the percentage of AUC were 0.80 in Gradient Boosting, 0.781 in deep learning, 0.746 in logistic regression, 0.740 in SVM and 0.735 in MLP. Algorithm with the highest level of AUC was Gradient Boosting. Gradient Boosting's AUC = (0.8) was better at predicting mortality in intensive care units. 1) bilirubin 2) INR 3) Age 4) WBC had the strongest relationship with mortality in Gradient Boosting. Conclusion: Data analysis in ICU patients can be a valuable tool for predicting mortality and morbidity associated with mortality, but the results vary according to the data quality. However, the processes and methods mentioned in this study suggested that the rules extracted from the Gradient Boosting can be used as a model to predict mortality in intensive care units.
|
|---|