Evaluation of oxygenation indices incorporating SpO₂ and PEEP for assessing ARDS severity: Evidence from the MIMIC-IV and eICU collaborative research database v2.0 databases

Abstract

Background The in-hospital mortality of acute respiratory distress syndrome can reach 35–45%, with patients requiring a more convenient and accurate way to assess the disease condition, which can change even more rapidly in patients undergoing mechanical ventilation in intensive care units. Methods Eligible patients in MIMIC-IV v3.0and eICU Collaborative Research Database v2.0were screened by the Berlin definition to examine the comparison of the diagnostic abilities of SpO 2 *10/FiO 2 *PEEP (S/F*P), PaO 2 *10/FiO 2 *PEEP (P/F*P), and SpO 2 / FiO 2 (S/F), with nine types of machine learning performed on S/F*P for 10 cross-validations, validating the diagnostic ability of the models for ARDS patients. Results ROC_AUC = 0.700(95 CI:0.624 ~ 0.777) for S/F, ROC_AUC = 0.720(95 CI:0.668 ~ 0.772) for P/F*P, and ROC_AUC = 0.761(95 CI:0.693 ~ 0.830) for S/F*P showed that S/F had a better fit in diagnosing ARDS with slightly inferior efficacy to P/F*P, had superior diagnostic efficacy after incorporating peep into S/F, and S/F*P showed good diagnostic efficacy in 9 machine learning and 10 cross-validations. In terms of predicting the prognosis of patients, the ability of S/F*P is not as good as S/F, but the grading of S/F*P has a more positive significance for the evaluation of the prognosis of patients. Conclusion S/F*P provides a more convenient judgment for mechanically ventilated patients, avoiding the phenomenon of clinical diagnosis of PEEP and oxygenation index separation as much as possible, minimizing invasive operation of patients and improving the selection of ARDS treatment modalities. Therefore, S/F*P provides a reference for the early treatment of ARDS in the clinic to improve the resource allocation in the ICU and reduce the mortality of patients, Given all patients had ARDS diagnosis, this study evaluated relative diagnostic performance among indices rather than disease vs. non-disease discrimination.