Background

Assessment of surgical competency by review of surgical videos is becoming an accepted standard. We sought to design an artificial intellegence (AI) algorithm capable of recognizing the surgical steps of a gastric bypass surgery to aide in video assesment.

Methods

A total of 53 gastric bypass videos from a single instutition were divided into training (n = 32), testing (n = 17), and validation (n = 4) datasets. Each procedure was separated into 13 steps: angle of His dissection (HIS), perigastric dissection (PD), pouch formation (PC), omental transection (OT), ligament of Treitz identification (LT), counting biliopancreatic limb (BP Limb), gastro-jejunal anastomosis (G-J A), jejunal transection (DSI), leak test (LT), Counting of Roux n Y limb (RNY), jejuno-jejunal anastomosis (J-J), closure mesenteric defect (MD), and flow test and hemostasis (FT). The AI model learned to recognize the steps using the training dataset and evaluated for initial learning in the testing dataset. The validation dataset was used for final model evaluation.

Results

In the validation dataset, the model had a microaveraged receiver operator charateristic curve value of 0.97 and overall accuracy of 79% for predicting surgical steps. Accuracies of individual surgical steps were 100% HIS, 77% PD, 73% PC, 100% OT, 69% LT, 85% BP Limb, 86% G-J A, 86% DSI, 60% LT, 71% RNY, 79% J-J, 79% MD, and 69% FT.

Conclusions

AI can recognize the surgical steps of a gastric bypass surgery and may be used in the future to aide in surgical video assesment for training or quality.