Advantage and Core Benefit
- Predicts ventilatory difficulty events over one minute before onset using time-series data from routinely monitored intraoperative physiological parameters.
- Machine learning model developed using data from approximately 460 pediatric patients.
- Addresses a significant clinical need for safe airway management in pediatric general anesthesia with supraglottic airway devices (SGAs).
Background and Technology
Supraglottic airway devices (SGAs) are commonly used for airway management during general anesthesia. Compared to tracheal intubation, SGAs are less invasive and are associated with fewer hemodynamic responses and postoperative complications. They are frequently used for minor surgeries involving the limbs or superficial tissues.
However, inadequate anesthesia during SGA use can lead to laryngospasm—a protective airway reflex triggered by pain or secretions—especially in children. Laryngospasm may result in complete airway obstruction, causing ventilation difficulty, hypoxemia (SpO₂ < 90%), and potentially cardiac arrest. The incidence of laryngospasm is higher in children, particularly those with upper respiratory infections or when managed by less experienced anesthesiologists.
Although several technologies aim to detect hypoxemia, there is limited reaction time once it occurs. Therefore, early prediction and prevention of the preceding event—ventilatory difficulty—is highly desirable.
The research team hypothesized that changes in vital signs and airway compliance could signal impending ventilatory difficulty. They developed a machine learning model that uses intraoperative time-series physiological data—including indicators of sympathetic activity and respiratory status—to predict such events. Two board-certified anesthesiologists reviewed surgical data, ventilation settings, and clinical notes to identify events of ventilatory difficulty. The model was trained in approximately 460 pediatric cases and validated on around 120 additional cases. It achieved a sensitivity of approximately 60% for detecting ventilatory difficulty within the window from 5 to 1 minute before onset, with a low false positive rate.
Data
- The prediction model was evaluated using archived time-series monitoring data. The model using clinically important features identified by experienced anesthesiologists—heart rate (HR), systolic blood pressure (SBP), end-tidal CO₂ (EtCO₂), peak inspiratory pressure (PIP), tidal volume (TV), and minute ventilation (MV)—achieved a true positive rate (TPR) of 57% and a false positive rate (FPR) of 0.65 events per hour.
Patent & Publication
Toshiyuki Nakanishi et al., The 38th Annual Conference of the Japanese Society for Artificial Intelligence (2024)
DOI: https://doi.org/10.11517/pjsai.JSAI2024.0_1K3GS1002
Researcher
Dr. Toshiyuki Nakanishi (Nagoya City University)
Expectations
We are seeking collaboration with companies involved in anesthesia-related equipment and monitoring systems to co-develop this program. We envision its integration into intraoperative monitoring systems with a dedicated mode for supraglottic airway device use. Meetings with the principal investigator can be arranged upon request.
Project ID: WL-04937