Optimal Triage for COVID-19 Patients Under Limited Health Care Resources With a Parsimonious Machine Learning...

The COVID-19 pandemic has placed an unprecedented burden on health care systems.

We aimed to effectively triage COVID-19 patients within situations of limited data availability and explore optimal thresholds to minimize mortality rates while maintaining health care system capacity.

A nationwide sample of 5601 patients confirmed with COVID-19 until April 2020 was retrospectively reviewed. Extreme gradient boosting (XGBoost) and logistic regression analysis were used to develop prediction models for the maximum clinical severity during hospitalization, classified according to the World Health Organization Ordinal Scale for Clinical Improvement (OSCI). The recursive feature elimination technique was used to evaluate the maintenance of model performance when clinical and laboratory variables were eliminated. Using populations based on hypothetical patient influx scenarios, discrete-event simulation was performed to find an optimal threshold within limited resource environments that minimizes mortality rates.

The cross-validated area under the receiver operating characteristic curve (AUROC) of the baseline XGBoost model that utilized all 37 variables was 0.965 for OSCI ≥6. Compared to the baseline model's performance, the AUROC of the feature-eliminated model that utilized 17 variables was maintained at 0.963 with statistical insignificance. Optimal thresholds were found to minimize mortality rates in a hypothetical patient influx scenario. The benefit of utilizing an optimal triage threshold was clear, reducing mortality up to 18.1%, compared with the conventional Youden index.

Our adaptive triage model and its threshold optimization capability revealed that COVID-19 management can be achieved via the cooperation of both the medical and health care management sectors for maximum treatment efficacy. The model is available online for clinical implementation.