Predicting Diagnostic Success and Procedural Efficiency in Robotic Bronchoscopy Using Machine Learning

Background. Robotic-assisted bronchoscopy with the ION™ Endoluminal System facilitates precise access to peripheral pulmonary lesions. However, procedural duration and diagnostic performance remain influenced by patient and lesion-specific factors. To investigate the impact of lesion diameter, radiological appearance, and presence of bronchial signs on procedural duration and diagnostic yield using conventional regression and gradient boosting machine learning models. Methods. In this single-center retrospective cohort study, 189 ION™ Endoluminal System procedures (November 2024–June 2025) were analyzed. Procedural duration and diagnostic yield served as primary outcomes. Predictive modeling included multivariable regression and gradient boosting. Feature importance metrics were extracted. Results. The median lesion diameter was 12.3 mm, with a “strict” diagnostic yield of 87.3%. Gradient boosting regression identified lesion diameter as the primary predictor of procedural time (89.2% importance; test MSE = 865.6). Diagnostic classification achieved an ROC-AUC of 0.68, with lesion diameter (85.8%) and bronchial sign (14.2%) as key predictors. Conclusions. Lesion diameter emerged as the most consistent predictor of procedural efficiency and was associated with diagnostic performance, albeit within the limitations of the dataset. Broader datasets are needed for external validation and generalizability.