A temporal deep learning algorithm for prediction of extubation failures in critical care patients

Extubation failure in ICU patients is associated with poor outcomes. Existing prediction models often rely on static data, missing dynamic disease fluctuations. This study introduces TrAcE, a deep learning-based model integrating static and temporal data for improved extubation failure prediction with explainable results. The model was trained and validated using MIMIC-III (Medical Information Mart for Intensive Care-III) data and tested on the LOCAL-Ext (a local database for extubation) dataset. A Transformer-based neural network with temporal fusion was used to screen extubation records (Patients planned for post-extubation non-invasive ventilation or tracheostomy were excluded). Model performance was assessed using AUROC (area under the receiver operating curve) and AUPRC (area under the precision recall curve). Explainability was ensured via Captum's occlusion method, identifying feature attributions at both population and individual levels. TrAcE's extubation timing was compared to the spontaneous breathing test (SBT) in LOCAL-Ext. From MIMIC-III, 5,895 patients (4,126 for training, 1,729 for validation) were selected. LOCAL-Ext included 6,765 test patients. TrAcE outperformed other models, achieving AUROCs of 0.823 (95% CI: 0.808-0.838) in validation and 0.859 (95% CI: 0.836-0.881) in testing, with favorable AUPRCs (0.734 and 0.757, respectively). Compared to SBT in LOCAL-Ext, TrAcE identified suitable extubation days earlier (4.32 [IQR 0-20] vs. 5.65 [IQR 0-23], p < 0.01). In the present study, we developed a deep learning model (TrAcE) for predicting extubation failure in critical care patients. This model can process both static and dynamic data, provides risk prediction and explainability in real-time. Furthermore, it demonstrated an advantage over SBT procedures in promptly identifying patients ready for extubation.