Using artificial intelligence to model expert panel diagnosis of cholecystitis severity

Determining cholecystitis severity via the clinically validated Parkland Grading Scale (PGS) is useful for predicting case difficulty and likelihood of postoperative complications. A panel assessment by multiple surgeons can reduce variation in PGS due to subjectivity, but is time-consuming. An artificial intelligence (AI) model trained on the assessments of an expert clinician panel may improve efficiency and reduce variability in diagnosis in image-based assessments. Laparoscopic cholecystectomy videos were obtained from one public and two private data sources. Representative frames were chosen for PGS grading and manually labeled. Three surgical experts independently assigned PGS scores to the selected frames. They then convened as a panel to decide on the score if those were discrepant at individual scoring. Weighted Cohen's kappa statistic was measured for inter-rater variability. Two AI models were developed for automated PGS grading and their accuracy and interpretability evaluated. 319 videos were compiled. Three surgical experts independently assigned identical PGS grades for 51% of cases, and weighted Cohen's kappa statistics ranged between 0.76 and 0.83. The accuracy of Model A using absolute agreement with the expert panel's consensus was 69%, and weighted Cohen's kappa statistic was 0.62. The accuracy of Model B using absolute agreement with the panel's consensus was 72%, and weighted Cohen's kappa statistic was 0.77. Interpretability analysis was conducted. Three anatomical structures played a key role in Model B's grading of cholecystitis severity: the appearance of the gallbladder, liver, and omentum had notable impact on performance. A transformer-based AI model can be trained on consensus from an expert panel to predict ratings of cholecystitis severity (Parking Grading Scale), performing competitively with some individual experts at predicting PGS when compared to the panel-based ground truth. However, variance and subjectivity of PGS remain, thus presenting its limitations as a ground truth for computer vision-based models.