Endoscopic sleeve gastroplasty video assessment: do technical features influence ESG integrity and weight loss at 6...

Endoscopic Sleeve Gastroplasty (ESG) allows for gastric volume reduction and shortening leading to weight loss and resolution of obesity-related comorbidities. While position statements and recommendations are being developed, limited studies have explored how technique influences outcomes. Video-based assessment (VBA) of endoscopic and surgical procedures are increasingly being adopted to achieve a deeper understanding of procedures technical aspects. This study explores how ESG technical features and anatomical landmarks relate to outcomes, and to develop predictive models from them.

Videos of ESG were collected, and an annotation manual was developed by an ESG expert, outlining technical and anatomical landmarks possibly influencing 6 and 12 month outcomes. Evaluated outcomes included technical (i.e., intact gastroplasty) and clinical (i.e., total and excess weight loss percentages) success. Two independent surgeons annotated the videos. Analysis of annotated features and of an engineered feature was performed; predictive regression models were developed. Pearson correlation and inter-rater reliability (IRR) of the annotations were evaluated.

Forty videos were annotated. The features "Bleeding," "Hemicircumferential" suture pattern encircling 180° of the stomach's incisura, "Suture number" and "Regular tubular ESG" configuration were analyzed. The engineered feature "Suture_Regular" defined by a combination of Regular tubular ESG" and "Suture number" was developed and analyzed. Predictive models were developed using progressively expanded feature sets in various combinations. XGBoost models demonstrated best performances across outcomes and timepoints, with R² values of 0,74-0,87, depending on the outcome and timepoint evaluated. Single features exhibited limited predictive power. IRR varied by feature, with Cohen's kappa ranging from slight to substantial.

VBA can enhance the understanding of endoscopic techniques. Combinations of events occurring during ESG and technique-specific features can be used to predict technical and clinical success. Future studies should include multicentric data to enhance models generalizability. Automating feature recognition could enable real-time guidance and improve patient management.