Artificial Intelligence in Surgery: predicting outcomes and measuring performance
Simon Baltus is a PhD student in the department Robotics and Mechatronics. (Co) Promotors are prof.dr. I.A.M.J. Broeders and dr. C.O. Tan from the faculty Electrical Engineerings, Mathematics and Computer Science (EEMCS), University of Twente.
Artificial Intelligence (AI) holds promise for surgery by supporting decision-making, personalizing treatment strategies, and enhancing performance evaluation. As data is collected before, during, and after surgical procedures, AI can be applied across the entire perioperative spectrum. This thesis investigated how AI can perform preoperative prediction of surgical outcomes and postoperative assessment of surgical performance. Six studies form the foundation of this work.
The first part focused on predicting anastomotic leakage after rectal cancer surgery and the duration of gallbladder surgery. In the first study, a deep learning approach was developed to automate the measurement of pelvic dimensions in magnetic resonance imaging scans. The model performed at a human level, enabling an objective and standardized assessment of the dimensions relevant to surgical difficulty. In the second study, multimodal machine learning models were developed to predict anastomotic leakage after rectal cancer surgery preoperatively. Pelvic inlet width, interspinous distance, and distance to the anorectal junction were identified as independent predictors. While the predictive performance of the models remained moderate, the study highlighted the relevance of pelvimetry in clinical decision-making. In the third study, an adaptive framework was created to predict the procedure duration of laparoscopic cholecystectomy based on preoperative factors and intraoperative difficulty. The procedure duration was independently associated with the level of expertise, history of cholecystitis, hydropic gallbladder, and surgical difficulty. While the model's predictive value was modest, the study highlighted the importance of intraoperative difficulty for dynamic, personalized surgical planning.
The second part addressed postoperative performance assessment through the analysis of electrosurgical device usage. An energy dashboard was introduced to visualize information extracted from energy generator data, providing a basis for feedback and benchmarking. The pilot study demonstrated significant differences between surgeons in total usage, turn-on count, and amount of applied energy. Building on this, a machine-learning-based algorithm was developed to automatically detect device-induced bleeding in laparoscopic diaphragmatic hernia repair. Although precise detection proved challenging, the algorithm provided a novel performance marker for suboptimal energy application. Finally, surgical phase recognition in diaphragmatic hernia repair videos was implemented to contextualize electrosurgical events within procedural steps. The deep learning methodology can accurately recognize the course surgical phases, but requires further development to distinguish the different dissection parts.
Collectively, this thesis laid the groundwork for preoperative and postoperative tools to support surgery through standardized analysis of clinical, imaging, and intraoperative data. By quantifying anatomical constraints, risk factors, and device usage, these studies facilitate objective decision-making and performance evaluation. AI played a crucial role in converting multimodal data into interpretable metrics for risk assessment, surgical planning, and skill development. While current predictive models and assessments need further refinement for clinical use, they demonstrate a scalable framework that enhances surgical expertise with context-aware insights, ultimately improving patient treatment and continuous development.”
