Using decision tree algorithms for estimating ICU admission of COVID-19 patients

Abstract

Introduction: Coronavirus disease 2019 (COVID-19) outbreak has overwhelmed many healthcare systems worldwide and put them at the edge of collapsing. As intensive care unit (ICU) capacities are limited, deciding on the proper allocation of required resources is crucial. This study aimed to develop and compare models for early predicting ICU admission in COVID-19 patients at the point of hospital admission. Materials and methods: Using a single-center registry, we studied the records of 512 COVID-19 patients. First, the most important variables were identified using Chi-square test (at p < 0.01) and logistic regression (with odds ratio at P < 0.05). Second, we trained seven decision tree (DT) algorithms (decision stump (DS), Hoeffding tree (HT), LMT, J-48, random forest (RF), random tree (RT) and REP-Tree) using the selected variables. Finally, the models’ performance was evaluated. Furthermore, we used an external dataset to validate the prediction models. Results: Using the Chi-square test, 20 important variables were identified. Then, 12 variables were selected for model construction using logistic regression. Comparing the DT methods demonstrated that J-48 (F-score of 0.816 and AUC of 0.845) had the best performance. Also, the J-48 (F-score = 80.9% and AUC = 0.822) gained the best performance in generalizability using the external dataset. Conclusions: The study results demonstrated that DT algorithms can be used to predict ICU admission requirements in COVID-19 patients based on the first time of admission data. Implementing such models has the potential to inform clinicians and managers to adopt the best policy and get prepare during the COVID-19 time-sensitive and resource-constrained situation.

Keywords

• COVID-19
• Coronavirus
• Machine learning
• Intensive care unit
• Decision tree