Abstract Rationale A subset of patients with Acute Respiratory Distress Syndrome (ARDS) experience rapid resolution of hypoxemia and are therefore less likely to benefit from therapies targeting respiratory failure. Early identification of patients with rapidly improving acute hypoxemic respiratory failure (riAHRF) may improve clinical management and optimize trial enrollment. We hypothesize that a machine learning approach could be used to predict riAHRF using clinical data available prior to intubation. Methods We performed a single-center retrospective study of adults (≥18 years) presenting to the Emergency Department (03/01/2022-02/01/2024) who were intubated for AHRF (defined by S:F 315) and admitted to the medical/surgical ICU. Resolution of AHRF was defined as the first instance of PEEP ≤5 cm H2O and FiO2 ≤ 0.4 maintained for 6 hours; riAHRF was defined as resolution within 24 hours of intubation. An XGBoost binary classifier predicting riAHRF was trained using clinical data from the six hours preceding intubation. Longitudinal vital signs and ventilator metrics were summarized at the patient level. Patients were temporally split by medical record number into training (75%), validation (15%), and testing (10%) sets. Five-fold forward time-series cross-validation guided hyperparameter selection using AUROC, AUPRC, sensitivity, and specificity. To address class imbalance, we applied a scale factor to increase the penalty for misclassified positive samples and reducing bias toward the majority class. The final model was retrained on the full training set, with an optimal decision threshold selected on the validation set by maximizing Youden’s J. Model performance was assessed on the independent test set. Feature importance was summarized by average gain (i.e. the mean improvement in model loss attributed to splits on each feature across all trees) reflecting both frequency of use and discriminative power in partitioning the data. Results Among 551 patients, 392 (71%) experienced riAHRF. Mortality was lower among patients with riAHRF (27%) compared with those without (38%; OR 1.6 1.01-2.60, p = 0.04). The final model achieved an AUROC of 0.87 in the test set, with sensitivity 0.90 and specificity 0.60. FiO2 and SpO2 were the most influential predictors, but model performance improved when incorporating age, sex, and additional vital signs. Conclusion A substantial portion of patients with AHRF reach minimal ventilator settings within 24h of intubation and have better outcomes. It is likely that riAHRF can be predicted at the time of intubation using tree based binary classifiers like XGBoost. This abstract is funded by: NA
Chiacchia et al. (Fri,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: