Modification and Application of the ProVent-14 Model to a Covid-19 Cohort to Predict Risk for In-Hospital Mortality

Abstract

Rationale: Deficiencies exist in the communication of prognosis for patients requiring prolonged mechanical ventilation (PMV) from COVID-19 pneumonia, in part because of clinician uncertainty about the natural history of disease and observational cohort studies with variable outcomes. In order to address this gap for PMV patients, we developed a modified clinical prediction model based on the ProVent-14 model to predict in-hospital mortality for patients receiving at least 14 days of mechanical ventilation for acute respiratory distress syndrome (ARDS) from COVID-19. Methods: We evaluated 107 patients with COVID-19 requiring PMV (at least 14 days of mechanical ventilation (MV)) at 2 tertiary care medical centers in the US in a retrospective observational cohort study. On day 14 of MV, we collected data for the original ProVent-14 variables (age, platelet count, requirement for vasopressors, non-trauma admission, and dialysis requirement). We also collected data for 2 other potential predictor variables (extra-corporeal membrane oxygenation (ECMO) on day 14 and neutrophil to lymphocyte ratio). Model Development: Logistic regression models were used to evaluate the performance of the ProVent-14 variables with the outcome inhospital mortality. We then assessed successive models adding variable combinations including requirement of ECMO and neutrophil to lymphocyte ratio on day 14 to predict inhospital mortality. We assessed discrimination of the models by measuring the area under the receiver operating characteristic curve (AUC). We assessed calibration by the Hosmer-Lemeshow goodness of fit statistic. Results: The AUC for the model using original Provent-14 variables was 0.78 (trauma omitted for N=1). The most parsimonious model using the additional variables includes risk factors age 50-64 and ≥65;platelet count <100, and requirement for vasopressors, renal replacement or ECMO on day 14 of MV. The area under the curve for this model is 0.83. Calibration for the modified parsimonious model is provided in the table below (Goodness-of-fit statistic p=0.80). Dichotomized neutrophil to lymphocyte ratio on day 14 (N:L>15) improves the model slightly AUC=0.83, Goodness-of-fit p=0.61, though this variable was available for only 60% of the cohort. Conclusion: A modified clinical prediction model based on the previously validated ProVent-14 model is a simple method to accurately identify patients with ARDS from COVID-19 requiring PMV who are at high risk of in-hospital mortality. Further validation of model performance in a larger population and including long-term survival is warranted.