Surgical bleeding prediction using transformer: an application to laparoscopic cholecystectomy

Surgical bleeding is often associated with the surgeon's actions during procedures and may be preventable if high-risk behaviors are identified in advance. Here, we aim to develop a real-time tool for predicting intraoperative bleeding risk using an event-log-based framework, and we demonstrate its feasibility in laparoscopic cholecystectomy with potential for broader generalization.

We represent the surgeon's workflow as an event log composed of triplets (action, instrument, target), their durations, recent bleeding history, and surgical phase information extracted from the preceding minutes of the procedure. Using real-time event logs from laparoscopic cholecystectomy procedures in the CholecTrack20 and CholecT50 datasets, we preprocess these data and train a Transformer-based model to estimate the probability of bleeding in the immediate future.

Our Transformer-based model outperforms LSTM- and TCN-based approaches, achieving its best performance for short-term bleeding prediction. For 30-60 s ahead, the F1-score reaches 68.6%, surpassing TCN (60.1%) by 8%. For 60-90 s, it remains competitive at 64.4%, about 6% higher than TCN, with performance gradually declining over longer horizons.

Predicting bleeding events from real-time event logs based on recent surgical activity shows strong potential for effective intraoperative decision support. To further enhance model relevance and enable reliable real-time deployment, additional data (i.e., more surgical procedures) and more detailed bleeding annotations (e.g., bleeding intensity, the location of the bleeding, and whether the bleeding is active) will be beneficial.