Utility of the Simulated Outcomes Following Carotid Artery Laceration Video Data Set for Machine Learning...

Area of Science:

• Surgical data science
• Medical imaging
• Computer vision

Background:

• Adverse surgical events like hemorrhage are underrepresented in video datasets, limiting AI model generalizability and potentially introducing bias.
• Hemorrhage presents unique challenges for computer vision due to blood obscuring the surgical field.

Purpose of the Study:

• To evaluate the Simulated Outcomes Following Carotid Artery Laceration (SOCAL) dataset for benchmarking surgical data science techniques.
• To assess computer vision instrument detection, instrument use metrics, and outcome associations using the SOCAL dataset.
• To validate a neural network trained on SOCAL using real-world operative videos.

Main Methods:

• A dataset of 147 surgical hemorrhage control trials was created from 75 surgeons using cadaveric models.
• Videos were annotated for surgical instruments, and outcomes like time to hemostasis and blood loss were recorded.
• Deep neural networks were trained on the SOCAL dataset for instrument detection and validated on real surgical videos.

Main Results:

• The SOCAL dataset contains over 31,000 frames with 65,000 instrument annotations.
• Deep neural networks trained on SOCAL achieved a 0.67 mean average precision (mAP) for instrument detection, with 0.91 for suction.
• The model demonstrated 0.77 sensitivity and 0.96 positive predictive value on real surgical video, with instrument use metrics correlating with blood loss.

Conclusions:

• SOCAL is the first dataset to capture hemorrhage control, addressing a critical gap for surgical data science.
• It enables benchmarking of object detection, performance metrics, and outcome association identification.
• SOCAL provides a foundation for developing and validating future surgical data science models for adverse event management.