Machine learning in anesthesiology: Detecting adverse events in clinical practice

Area of Science:

• Anesthesiology
• Artificial Intelligence
• Machine Learning

Background:

• Threshold-based alarms in anesthesia monitoring lack credibility and often provide non-informative warnings.
• Existing alarm systems may not be suitable for all types of surgical procedures.

Purpose of the Study:

• To demonstrate the potential of Machine Learning (ML) techniques for generating meaningful anesthesia alarms.
• To explore different ML approaches for improved intraoperative patient monitoring.
• To develop alarm systems without procedural constraints.

Main Methods:

• Two ML approaches were evaluated: Complication Detection (supervised learning) and Anomaly Detection.
• A simple feed-forward Neural Network was used for Complication Detection.
• An Encoder-Decoder Long Short-Term Memory (LSTM) architecture was employed for Anomaly Detection, minimizing the need for extensive manual labeling.

Main Results:

• The feed-forward Neural Network performed optimally for the Complication Detection task.
• The LSTM-based Anomaly Detection approach proved more flexible.
• The Anomaly Detection method aligns with Explainable Artificial Intelligence (XAI) principles.

Conclusions:

• Machine Learning offers a viable path to enhance the informativeness and credibility of anesthesia alarms.
• Anomaly Detection using LSTM presents a flexible and promising method for real-time patient monitoring.
• The developed ML approaches have the potential for future improvements in anesthesia safety and explainability.