Validation of the three-level hepatectomy complexity classification and its AI application in robotic liver surgery

Robotic liver surgery (RLS) is expanding in recent years. Complication prediction is crucial for postoperative outcomes. Traditional MIS scores are poorly studied in RLS, and conventional statistics often oversimplify the multifactorial and interrelated nature of these complications. This study aimed to evaluate the three-level complexity Institut Mutualiste Montsouris (IMM) classification in RLS and assess its integration into an AI algorithm to predict major complications. We retrospectively analyzed data of patients underwent RLS. Surgical complexity was stratified into grades I (low complexity), II (intermediate), and III (high). The cumulative incidence rate and conditional probability of postoperative complication and risk factors for complication ≥ Clavien-Dindo grade II were assessed. The prediction model was developed by training/testing a machine learning (ML) algorithm after feature selection with uni-multivariate analysis. We calculated the receiver operating characteristic (ROC) curve and model accuracy. We analyzed 1,045 patients who underwent RLS, classifying them into three complexity levels: Grade I (n = 581), Grade II (n = 267), and Grade III (n = 109). Significant differences were observed in intra- and postoperative outcomes across the three grades. Multivariate analysis identified ASA score (HR 2.1, p = 0.02), number of lesions (HR 1.8, p = 0.001), and operative time (OR 1, p = 0.004) as key predictors of complications. Associated with the three-level complexity classification, the Neural Network showed the best performance with AUC (0.653) and a precision of 0.996. Three-level complexity IMM classification is a useful tool in RLS for predicting intra-postoperative outcomes. It can be integrated into the Neural Network algorithm to predict major complications.