Early Detection and Classification of Patient-Ventilator Asynchrony Using Machine Learning

• Published in: Artificial Intelligence in Medicine pp. 238 - 248
• Main Authors: Gao, Erdi, Ristanoski, Goce, Aickelin, Uwe, Berlowitz, David, Howard, Mark
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing 2022
• Series: Lecture Notes in Computer Science
• Subjects: Classification; Data mining; Patient-ventilator asynchrony
• ISBN: 9783031093418; 3031093410
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-09342-5_23

During mechanical ventilation, a common problem known as patient-ventilator asynchrony (PVA) occurs when there is a mismatch between the needs of the patient’s breathing and the breath cycle delivered by the ventilator. PVA is problematic because it can be associated with adverse effects such as discomfort for the patient, increased work of breathing, longer mechanical ventilation duration and ventilator-induced lung injury. An automated means of early PVA detection and classification could lead to improved health outcomes and help reduce the impact of PVA on hospital resources. This paper presents a machine learning framework to detect PVA events using only the first half second of data after the start of a PVA event. When trained on more than 5000 PVA events sampled from 25 subjects, our logistic classifier achieves a sensitivity (specificity) of 99.81% (99.72%) for detecting PVA events. We then present a system capable of early classification of Ineffective Effort (IE) and Double Trigger (DT) events, which achieves a sensitivity (specificity) of 63.73% (92.88%). By demonstrating the feasibility of early PVA event detection and classification, our findings suggest that more effective intervention processes could be possible, including automated interventions with different response strategies for different PVA event types.