JMIR Publications

ABSTRACT

Background:

Predicting respiratory failure in COVID-19 patients based on an early clinical profile can help triaging, resource allocation, and morbidity reduction by appropriately monitoring patients at risk. Given the complexity of the disease, this effort can benefit from machine learning (ML) approaches.

Objective:

Our objective is to establish a machine learning model that predicts respiratory failure within 48 hours of admission based on data from the emergency department (ED).

Methods:

Data was collected from patients with COVID-19 who were admitted to Northwell Health acute care hospitals and discharged, died, or spent a minimum of 48 hours in the hospital between March 1, 2020 and May 11, 2020. Of 11,525 patients, 933 (8.1%) were placed on invasive mechanical ventilation within 48 hours of admission. The variables used by the models included clinical and laboratory data commonly collected in the ED. We trained and validated an XgBoost model alongside two other prediction models using cross hospitals validation. We compared model performance and a Modified Early Warning Score (MEWS) using receiver operating characteristic (ROC) curves, precision-recall (PR) curves, and other metrics.

Results:

The XgBoost model had the highest mean accuracy of 0.908 (AUC = 0.83), outperforming MEWS and the other models. Important variables included the type of oxygen delivery used in the ED, patient age, emergency severity score, respiratory rate, serum lactate, heart rate, and serum glucose values.

Conclusions:

XgBoost has high predictive accuracy, outperforming other early warning scores. The clinical plausibility and predictive ability of XgBoost suggest that the model could be used to predict 48-hour respiratory failure in admitted patients with COVID-19.