Task-level kinematics improve outcome prediction in robotic cholecystectomy

Researchers retrospectively analyzed 234 Calot’s triangle dissections during robotic cholecystectomy: 198 performed by attending surgeons and 36 by trainees. For each prediction target—surgeon experience, case complexity, and outcomes—they compared models using 10 preoperative clinical variables, 10 task-level kinematic objective performance indicators (OPIs), or both. Five algorithms were evaluated using nested cross-validation, with performance measured by area under the receiver operating characteristic curve (AUROC).

For surgeon experience, OPI-only and combined models had AUROCs of 0.967 and 0.964, respectively; for case complexity, the corresponding values were 0.791 and 0.792. Neither difference was statistically significant, and OPI-only models outperformed clinical-only models for both targets. For outcomes, the combined model reached an AUROC of 0.753, compared with 0.720 for either single-modality model; these differences were statistically significant.

The findings suggest task-level kinematic data contain predictive information beyond routine preoperative variables, with incremental value depending on the target. The improvement for outcomes was modest. This retrospective analysis evaluated model discrimination in a limited set of procedures; it does not establish that using these predictions improves surgical decisions or patient outcomes.