Machine learning for predicting extubation success: secondary analysis of the COBRE-US study

<bold>Introduction:</bold> The performance of a machine learning system trained on clinical variables to estimate the success of extubation in critically ill patients under ventilatory support remains unassessed. <bold>Methods:</bold> Prospective multicenter study in a prospective cohort with patients recruited to participate in the COBRE-US study from February 2019 to November 2021. Unsupervised and supervised machine learning methods were trained with the training samples, evaluating the diagnostic accuracy internally and externally of the success of the spontaneous breathing trial (SBT) based on diaphragmatic contraction velocity and Cough, and for successful extubation (SE) based on diaphragmatic contraction velocity, Cough, and SBT outcomes. <bold>Results:</bold> 367 patients were included, with a median age of 61 years (SD:49-72), and 57.9% (219/367) were men. For SE, decision trees showed a predictive capacity of 95.7% (sensitivity:99%-specificity:87.5%) and 93.5% (sensitivity:99%-specificity:81.3%) for neural networks and support vector machines in the test group (Table 1). Similar data were observed in SBT success, which showed a predictive capacity of 77.3% (sensitivity:99%-specificity:1%) in the training group for the same models. <bold>Conclusion:</bold> The proposed artificial intelligence system can estimate the predictive capacity for the success of SBT and SE through training with variables such as diaphragmatic contraction velocity, cough, and SBT. <fig><object-id>erj;64/suppl_68/OA4565/F1</object-id><object-id>F1</object-id><object-id>F1</object-id><graphic></graphic></fig>