Clinical models for predicting 30-day mortality in ARDS: A focus on ventilatory ratio-defined subgroups

Abstract

Background

It is supposed that acute respiratory distress syndrome (ARDS) patients with increased dead space, indicated by elevated ventilatory ratio (VR), had higher mortality. The difference in mortality predictors among ARDS patients categorized by VR remains unclear, so we aimed to investigate the risk factors for mortality prediction in subgroups defined by VR and develop risk models to predict the 30-day mortality in distinct ARDS subgroups.

Methods

This study performed a retrospective analysis using the Medical Information Mart for Intensive Care IV (MIMIC-IV) and eICU Collaborative Research Database (eICU-CRD) databases, as well as data from NHLBI ARDS Clinical Trials Network (ARDSnet). Patients were divided into high VR (VR ≥2) and low VR (VR <2) subgroups based on baseline VR. In ARDSnet cohort, two 30-day mortality risk prediction models were constructed using logistic regression, internally validated using the bootstrap method, and externally validated in the MIMIC-IV and eICU-CRD cohorts. The performance of models was evaluated using receiver operating characteristic curves, Brier scores, calibration curves, and decision curve analysis (DCA) curves, and DeLong test was used to compare the predictive efficacy of different models in each subgroup.

Results

This study included a total of 2977 ARDS patients: 1031 in the high VR training cohort, 1506 in the low VR training cohort, 159 in the high VR external validation cohort, and 281 in the low VR external validation cohort. Through stepwise regression analysis, 11 predictors were finally selected to construct high VR prediction model including age, body mass index (BMI), heart rate, mean arterial pressure, body temperature, blood urea nitrogen (BUN), bilirubin, minute ventilation, peak inspiratory pressure, inspired oxygen fraction (FiO₂), and peripheral oxygen saturation (SpO₂), and 13 predictors to construct low VR prediction model including age, heart rate, body temperature, platelet count, BUN, bilirubin, respiratory rate, peak inspiratory pressure, FiO₂, hypercapnia, hypocapnia, acidemia, and alkalemia. The area under the curve (AUC) of high VR model in the training cohort was 0.76 (95 % CI: 0.73 to 0.79), and the AUC of low VR model in training cohort was 0.76 (95 % CI: 0.74 to 0.79). The Brier scores for high VR model and low VR model in training cohort were 0.171 and 0.139, respectively. Decision curve analysis (DCA) showed that the DCA curve for high VR model waspositioned away from two extreme curves across a threshold probability range of 0.2 to 0.8, and the curve for low VR model was positioned away from two extreme curves across a threshold probability range of 0.1 to 0.6. The AUC of high VR model in low VR subgroup training set was 0.74, significantly lower than the low VR model (DeLong test, P=0.024). The AUC of low VR model in high VR subgroup training set was 0.73, significantly lower than high VR model (DeLong test, P=0.001).

Conclusion

This study developed and validated prognostic prediction models for patients of high and low VR subgroups, respectively. The models demonstrated good discrimination, calibration, and clinical utility. Prognostic risk factors differed between high and low VR subgroups, and prediction models developed for specific VR subgroups exhibited better predictive performance within their respective subgroup populations.