Artificial Intelligence Identifies Factors Associated with Blood Loss and Surgical Experience in Cholecystectomy

Laparoscopic surgery videos offer valuable insights into the intraoperative skills of surgeons. Traditionally, skill assessment has focused on trainees, but analyzing the operative techniques of established surgeons can reveal behaviors that are associated with surgical expertise. Computer vision (CV), a domain of artificial intelligence (AI), facilitates scalable, video-based assessment, enabling the discovery of novel associations between surgical skill and clinical outcomes. For this study, we developed an AI-powered CV model capable of autonomously recognizing fine-grained surgical actions in laparoscopic videos and uncovering associations between these actions and operative blood loss and surgical experience.

We utilized a dataset of laparoscopic surgical videos from 243 patients who underwent cholecystectomy. We used a subset of these videos to train an AI-powered CV model to recognize 150 fine-grained surgical action triplets (SATs) comprising unique combinations of three components: surgical instruments (16 total), motions (13), and anatomical structures (19). We then used the trained AI model to recognize these SATs in all 243 case videos. We considered estimated blood loss, as reported postoperatively by the performing surgeon, and refined this measure using retrospective video review by experienced surgeons, yielding operative blood loss. We also considered surgeon experience, defined as the number of postresidency years of the operating surgeon. We used a logistic regression model to infer blood loss and surgical experience on the basis of AI-identified surgical actions in the laparoscopic videos. We subsequently analyzed the relationships among surgical actions, operative blood loss, and surgical experience.

The operating surgeons in the video dataset had 8 to 31 years of surgical experience. Estimated operative blood loss among patients ranged from 0 to 175 ml. Our model predicted binary blood loss (low vs. moderate) with an area under the receiver operator characteristic (AUROC) of 0.81 and binary surgical experience (low vs. high) with an AUROC of 0.78. Higher blood loss was significantly associated with increased duration of use of a laparoscopic suction irrigator to dissect the cystic pedicle (P=0.04) and with use of the irrigator to aspirate blood (P=0.03) or irrigate the cystic pedicle (P=0.04). High surgical experience was moderately associated with longer duration of dissection of connective tissue with L-hook electrocautery (P=0.07) and with total duration of the case (P=0.07). High surgical experience was strongly associated with elective cases (P<0.001).

This study demonstrates the capability of AI CV models to analyze intricate surgical activity in large volumes of video data. By training the CV model on a set of laparoscopic cholecystectomy videos and then deploying it to recognize surgical actions in a larger cohort, we obtained novel and scalable insights without labor-intensive manual review. We specifically demonstrate the capability of AI-powered CV models to correlate surgical experience and technique with intraoperative outcomes (blood loss). (Funded by the Stanford Clinical Excellence Research Center and others.).