A Practical Clinical Score Predicting Respiratory Failure in COVID-19 Patients.

ABSTRACT

BACKGROUND: The coronavirus disease 2019 (COVID-19) pandemic resulted in repeated surges of patients, sometimes challenging triage protocols and appropriate control of patient flow. Available models, such as the National Early Warning Score (NEWS), have shown significant limitations. Still, they are used by some centers to triage COVID-19 patients due to the lack of better tools. OBJECTIVES: To establish a practical and automated triage tool based on readily available clinical data to rapidly determine a distinction between patients who are prone to respiratory failure. METHODS: The electronic medical records of COVID-19 patients admitted to the Sheba Medical Center March-April 2020 were analyzed. Population data extraction and exploration were conducted using a MDClone (Israel) big data platform. Patients were divided into three groups non-intubated, intubated within 24 hours, and intubated after 24 hours. The NEWS and our model where applied to all three groups and a best fit prediction model for the prediction of respiratory failure was established. RESULTS: The cohort included 385 patients, 42 of whom were eventually intubated, 15 within 24 hours or less. The NEWS score was significantly lower for the non-intubated patients compared to the two other groups. Our improved model, which included NEWS elements combined with other clinical data elements, showed significantly better performance. The model's receiver operating characteristic curve had area under curve (AUC) of 0.92 with of sensitivity 0.81, specificity 0.89, and negative predictive value (NPV) 98.4% compared to AUC of 0.63 with NEWS. As patients deteriorate and require further support with supplemental O2, the need for re-triage emerges. Our model was able to identify those patients on supplementary O2 prone to respiratory failure with an AUC of 0.86 sensitivity 0.95, and specificity 0.7 NPV 98.9%, whereas NEWS had an AUC of 0.76. For both groups positive predictive value was approximately 35. CONCLUSIONS: Our model, based on readily available and simple clinical parameters, showed an excellent ability to predict negative outcome among patients with COVID-19 and therefore might be used as an initial screening tool for patient triage in emergency departments and other COVID-19 specific areas of the hospital.