SmartAlert: Machine learning-based patient-ventilator asynchrony detection system in intensive care units

Area of Science:

• Critical Care Medicine
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Patient-ventilator asynchronies (PVAs) are linked to increased mortality and ventilator-induced lung injury.
• Current PVA detection methods are limited by static thresholds, complex preprocessing, or proprietary data access.
• There is a need for real-time, automated PVA detection systems.

Purpose of the Study:

• To develop and validate SmartAlert, an online, real-time system for detecting and classifying PVAs.
• To classify PVA severity and alert clinicians promptly.
• To utilize readily available ventilator screen data for PVA analysis.

Main Methods:

• Developed the SmartAlert system using ICU patient ventilator screen recordings.
• Extracted pressure and flow waveforms from video, converting them to time-series data.
• Employed deep neural networks for PVA classification and alarm level assignment, validated against expert consensus.

Main Results:

• SmartAlert achieved high accuracy in alarm level prediction (83.8%) and PVA classification (89.3%).
• The system demonstrated strong performance with weighted AUC-ROC values of 0.943 for alarms and 0.951 for PVAs.
• High specificity was observed for urgent alarms (99.9%) and specific PVA types (e.g., ineffective triggering 98.5%).

Conclusions:

• SmartAlert is an automated, validated system for real-time PVA detection, classification, and clinician alerting.
• The system shows potential to reduce alarm fatigue and optimize ventilator management.
• Further clinical trials are needed to confirm its impact on patient outcomes.