¹Institute for Computational Biomedicine, Universitätsklinikum Aachen, Aachen, Germany
²Klinik für Operative Intensivmedizin und Intermediate Care, Universitätsklinikum RWTH, Aachen, Germany

Text

Introduction: Acute Respiratory Distress Syndrome (ARDS) is a critical complication in ICU patients, associated with high morbidity and mortality [1]. Early detection is critical [2], but ARDS is frequently underdiagnosed [3]. We hypothesized that rapid declines in oxygenation, measured by the Horowitz index (PaO₂/FiO₂), can serve as an early surrogate marker for ARDS, enabling timely intervention and potentially improving outcomes.

Methods: We retrospectively analyzed data from 3,676 mechanically ventilated ICU patients (including 296 COVID-19 cases) at University Hospital RWTH Aachen from 2019 to 2021 [4]. A gradient boosted tree machine learning model was trained on the data of the past 24 hours to predict significant drops in oxygenation based on more than 80 clinical parameters. These parameters included routinely measured Intensive Care Unit (ICU) data, including ventilator settings, lab values, and demographic data. The data was split into 90% training and a 10% hold-out test set, with a 5-fold cross-validation then run on the training data to arrive at optimal hyperparameters.

Results: At a prediction horizon of up to 72 hours before onset the model achieved a receiver operating characteristic area under the curve (ROC AUC) of 0.90 and a precision-recall AUC of 0.54. At a high sensitivity (0.99), the specificity was 0.47. The model then missed only 1.6% of rapid oxygenation loss events. Performance was consistent across COVID and non-COVID cohorts, with successful transfer learning from the non-COVID to COVID group.

Conclusion: This study introduces a machine learning model that predicts rapid oxygenation loss as a surrogate marker for ARDS in ventilated ICU patients. Trained on 3,676 ICU encounters, it achieved a ROC AUC of 0.90 with a 72-hour prediction horizon. The model outperformed existing approaches [5] in specificity and generalized well across COVID-19 and non-COVID-19 cohorts. By focusing on dynamic changes in oxygenation, it offers a clinically relevant tool for early ARDS risk detection.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.

Literatur

[1] Fan E, Brodie D, Slutsky AS. Acute Respiratory Distress Syndrome: Advances in Diagnosis and Treatment. JAMA. 2018;319(7):698–710. DOI: 10.1001/jama.2017.21907
[2] Zahar JR, Azoulay E, Klement E, De Lassence A, Lucet JC, Regnier B, et al. Delayed treatment contributes to mortality in ICU patients with severe active pulmonary tuberculosis and acute respiratory failure. Intensive Care Med. 2001;27(3):513–20. DOI: 10.1007/s001340000849
[3] Bellani G, Laffey JG, Pham T, Fan E, Brochard L, Esteban A, et al. Epidemiology, Patterns of Care, and Mortality for Patients With Acute Respiratory Distress Syndrome in Intensive Care Units in 50 Countries. JAMA. 2016;315(8):788–800. DOI: 10.1001/jama.2016.0291
[4] Marx G, Bickenbach J, Fritsch SJ, Kunze JB, Maassen O, Deffge S, et al. Algorithmic surveillance of ICU patients with acute respiratory distress syndrome (ASIC): Protocol for a multicentre stepped-wedge cluster randomised quality improvement strategy. BMJ Open. 2021;11(4):e045589. DOI: 10.1136/bmjopen-2020-045589
[5] Le S, Pellegrini E, Green-Saxena A, Summers C, Hoffman J, Calvert J, et al. Supervised machine learning for the early prediction of acute respiratory distress syndrome (ARDS). J Crit Care. 2020;60:96–102. DOI: 10.1016/j.jcrc.2020.07.019