ABSTRACT
Neurological surgery in the semi-sitting position is linked with a pronounced incidence of venous air embolism (VAE) which can be fatal and therefore requires continuous monitoring. Transesophageal echocardiography (TEE) provides a high sensitivity for the intraoperative detection of VAE; however, continuous monitoring with TEE requires constant vigilance by the anaesthesiologist, which cannot be ensured during the entire surgical procedure. We implemented a fully automatic VAE detection system for TEE based on a statistical model of the TEE images. In the sequence of images, the cyclic heart activity is regarded as a quasi-periodic process, and air bubbles are detected as statistical outliers. The VAE detection system was evaluated by means of receiver operating characteristic (ROC) curves using a data set consisting of 155.14 h of intraoperatively recorded TEE video and a manual classification of periods with visible VAE. Our automatic detection system accomplished an area under the curve (AUC) of 0.945 if all frames with visible VAE were considered as detection target, and an AUC of 0.990 if frames with the least severe optical grade of VAE were excluded from the analysis. Offline-review of the recorded TEE videos showed that short embolic events (≤ 2 min) may be overseen when monitoring TEE video manually. Automatic detection of VAE is feasible and could provide significant support to anaesthesiologists in clinical practice. Our proposed algorithm might possibly even offer a higher sensitivity compared to manual detection. The specificity, however, requires improvement to be acceptable for practical application. Trial Registration: German Clinical Trials Register (DRKS00011607).