BACKGROUND AND OBJECTIVE: Patient-ventilator asynchronies (PVA) are associated with ventilator-induced lung injury and increased mortality. Current detection methods rely on static thresholds, extensive preprocessing, or proprietary ventilator data. This study aimed to develop and validate a fully online, real-time system that detects and classifies PVAs directly from ventilator screen data while alerting clinicians based on severity. METHODS: The SmartAlert system was developed using ventilator screen recordings from ICU patients. It extracts pressure and flow waveforms from video recordings, converts them into time-series data, and employs deep neural networks to classify asynchronies and assign alarm levels from no urgency to most urgent. A dataset of 381,280 double-breath units was independently annotated by two expert intensivists. Two deep learning models were trained: one for alarm prediction and another for asynchrony classification (ineffective triggering, double cycling, high inspiratory effort, no asynchrony). Performance was evaluated using accuracy, sensitivity, specificity, and AUC-ROC, compared to expert consensus. RESULTS: SmartAlert demonstrated strong performance for alarm level prediction (overall accuracy: 83.8 %, weighted AUC-ROC: 0.943 [95 % CI: 0.941-0.945]) and PVA classification (weighted accuracy: 89.3 %, weighted AUC-ROC: 0.951 [95 % CI: 0.950-0.953]). It showed high specificity for urgent alarms (99.9 % for level 3) and PVA types (98.5 % for ineffective triggering, 96.9 % for double cycling, 94.8 % for high inspiratory effort). CONCLUSIONS: We developed and internally validated SmartAlert, an automated system that detects PVAs, classifies severity, and alerts clinicians in real time. Its potential to reduce alarm fatigue, optimize ventilator settings, and improve patient outcomes remains to be tested in clinical trials.

Cognitive Robotics Faculty of 3mE Delft University of Technology Mekelweg 2 Delft CD 2628 The Netherlands

Department of Anesthesiology and Intensive Care 3rd Faculty of Medicine Charles University and Kralovske Vinohrady University Hospital Prague Šrobárova 50 Prague 100 34 Czech Republic

Department of Artificial Intelligence Czech Institute of Informatics Robotics and Cybernetics Czech Technical University Prague Jugoslávských partyzánů 1580 3 Prague 160 00 Czech Republic

Department of Cognitive Systems and Neurosciences Czech Institute of Informatics Robotics and Cybernetics Czech Technical University Prague Jugoslávských partyzánů 1580 3 Prague 160 00 Czech Republic

Department of Computer Graphics and Interaction Faculty of Electrical Engineering Czech Technical University Prague Karlovo náměstí 13 Prague 121 35 Czech Republic

Division of Clinical Geriatrics Center for Alzheimer Research Department of Neurobiology Care Sciences and Society Karolinska Institute Blickagången 16 Huddinge 14183 Sweden

Citace poskytuje Crossref.org