¹Klinik und Poliklinik für Orthopädie und Sportorthopädie des Klinikums rechts der Isar, München, Deutschland
²Institut für KI und Informatik in der Medizin (AIIM) des Klinikums rechts der Isar der Technischen Universität München, München, Deutschland
³Sektion Sportorthopädie, Klinik und Poliklinik für Orthopädie und Sportorthopädie des Klinikums rechts der Isar, München, Deutschland

Text

Objectives and questions: Success of technically challenging hip arthroscopy for femoroacetabular impingement syndrome (FAIS) with demanding (steep) learning curve and increased risk of complications critically depends on optimal intraoperative visualization. For the first, time Machine Learning Algorithm (MLA) and statistical modelling approaches are been utilized to predict and intelligently identify key parameter patterns that influence visibility during arthroscopy.

Material and methods: A large-scale big data analysis powered by artificial intelligence was conducted on 11,403 measurements. In detail visualization conditions, categorized on a scale from 1–2 (good) to 3–5 (poor), were assessed at 10 key surgical steps (central/peripheral) during supine FAIS arthroscopy in a prospective consecutive monocentric single-surgeon cohort level 2 study involving 211 patients, and correlated with corresponding blood pressure readings, arthroscopy tower parameters, and patient- and surgery-related data. Preprocessing included handling of missing values, standardization, and categorical encoding. LASSO regression identified the most relevant features, which were then used to train an XGBoost classifier. The dataset was split into training (80%) and test (20%) subsets, and model performance was evaluated using accuracy, precision, recall, and F1-score.

Results: Across all timepoints, the model demonstrated strong overall performance, achieving an average accuracy of 83.6%, precision of 83.8%, recall of 89.6%, and F1-score of 86.5%. At beginning of arthroscopic procedure, the MLA model achieved even higher performance with 90.7% accuracy, 92.9% precision, 97.5% recall, and a 95.1% F1-score. LASSO regression identified age, prior hip surgery, arthrosis (Tönnis-grade), systolic blood pressure (SBP), diastolic blood pressure (DBP), and mean arterial pressure (MAP) as key predictors. Across all timepoints, most frequently selected variables were SBP, irrigation fluid flow rate (Flow), MAP, Tönnis-grade, and BMI. XGBoost feature importance ranked SBP as the strongest predictor, followed by Flow, MAP, and Tönnis-grade.

Discussion and conclusion: According to our MLA, the key predictors for intraoperative visualization are SBP > Flow > MAP > Tönnis-grade. These findings underscore the critical role of hemodynamic parameters and patient-specific factors in optimizing hip joint visibility during arthroscopy and highlighting the potential for real-time monitoring with intraoperative adjustments. Future advancements should focus on developing interactive, high-automation arthroscopy towers capable of dynamically adapting to intraoperative conditions. The integration of MLA-driven advanced analytics in arthroscopy holds promise for optimizing complex, non-linear parameter interactions, with the objective of shortening the learning curve, improving patient outcomes, and facilitating complication-free outpatient procedures.