Machine Learning Characterization of Important Tool-tissue Interaction Forces Using Bipolar Forceps

Introduction: During surgical training, knowledge of tool-tissue interaction forces may be difficult to impart to trainees. Furthermore, quantifying ideal force characteristics using surgical tools remains to be determined within the context of surgical performance and education. This study aimed to identify predictive features during the use of bipolar forceps in neurosurgery and how they may help differentiate surgeons as ‘expert’ or ‘novice.’

Methods: Four surgeons used the SmartForceps System, a force-sensing bipolar forceps developed to quantify surgical tool-tissue forces. An ‘expert’ surgeon (n=1) with 30+ years of experience and ‘novice’ surgeons (n=3; three PGY 6 resident surgeons) performed tumor resection on 51 adult patients between October 2021 and June 2022. Five pre-determined surgical tasks were recorded: dissection, coagulation, retracting, pulling, and manipulating. The importance of features of each surgical task was quantified using decision trees, AdaBoost, and XGBoost. The results of each algorithm were averaged to obtain a rank-order list based on the importance of each force component.

Results: The accuracy of the decision tree was 81.3% (F1-score: 0.808), AdaBoost was 81.2% (F1-score: 0.828), and XGBoost was 81.3% (F1-score: 0.865) in predicting whether a force segment was performed by an expert or novice neurosurgeon. The most predictive features were maximum force (17.6%), force range (10.9%), mean force (9.4%), force variance (7.0%), minimum force (6.8%), median force (5.6%), and the first time derivative of force (5.4%). The least predictive features were force duration (1.7%), force stability (1.7%), frequency of force application (1.4%), and force spikiness (0.0%).

Conclusion : The algorithms quantify the relative importance of tool-tissue interaction force components and accurately predict expert or novice use of bipolar forceps. Future work will expand our analysis to investigate progress tracking and feedback for neurosurgical trainees, providing a quantitative surgical training paradigm to accelerate skill acquisition for trainees.