Surgical bleeding prediction using transformer: an application to laparoscopic cholecystectomy

Area of Science:

• Medical Informatics
• Surgical Technology
• Machine Learning in Healthcare

Background:

• Surgical bleeding is a significant risk during procedures, often linked to surgeon actions.
• Identifying high-risk behaviors in real-time can enhance surgical safety and prevent bleeding events.

Purpose of the Study:

• To develop a real-time tool for predicting intraoperative bleeding risk.
• To demonstrate the feasibility of an event-log-based framework for bleeding prediction in laparoscopic cholecystectomy.

Main Methods:

• Surgeon workflow represented as event logs (action, instrument, target, duration).
• Incorporated recent bleeding history and surgical phase information.
• Trained a Transformer-based model on real-time event logs from laparoscopic cholecystectomy datasets.

Main Results:

• The Transformer model outperformed LSTM and TCN approaches for short-term bleeding prediction.
• Achieved an F1-score of 68.6% for 30-60s prediction, an 8% improvement over TCN.
• Maintained competitive performance at 64.4% for 60-90s prediction.

Conclusions:

• Real-time event log analysis shows potential for intraoperative decision support to predict bleeding.
• Further data and detailed bleeding annotations are needed for enhanced model relevance and deployment.
• The developed tool could significantly improve patient safety by anticipating and mitigating surgical bleeding.