AI-enhanced prediction of diagnostic failure in early-stage lung cancer: Preprocedural score for ION robotic bronchoscopy

Background Diagnosis of early-stage NSCLC often depends on bronchoscopic sampling of peripheral small nodules. Although ION bronchoscopy improves access, non-diagnostic outcomes persist. We developed and validated an AI model to predict diagnostic failure, deriving a clinically applicable risk score. Methods Retrospective cohort of ION analyzed. Diagnostic failure was defined as non-diagnostic histology in lesions later confirmed or presumed malignant. Gradient boosting model was trained (5-fold cross-validation), SHAP analysis guided variable selection for logistic regression–based clinical score. Results The dataset comprised 189 IONs performed for small peripheral nodules suspected for early-stage NSCLC. The machine learning model based on gradient boosting (XGBoost) was trained using 80/20 split with stratified 5-fold cross-validation and Bayesian hyperparameter optimization. On external test fold, model achieved ROC-AUC = 0.90, precision = 0.76, recall = 0.78, specificity = 0.83 for predicting diagnostic failure. Model outperformed both logistic regression and random forest baselines (AUC 0.81 and 0.85, respectively). Feature importance analysis using SHAP highlighted five predictors of diagnostic failure: inadequate ultrasound visualization, no bronchial sign, apical/posterior segment location, GGO morphology, alone forceps use. These five features were combined into a simplified 10-point clinical score, derived through penalized logistic regression and calibrated using isotonic regression. Score performance mirrored the AI model’s stratification ability: • Low-risk (0–2 points) = 8.3% failure rate • Intermediate-risk (3–5 points) = 20.9% failure rate • High-risk (≥6 points) = 42.3% failure rate Score demonstrated a strong correlation with the model’s predicted probability (Pearson's ρ = 0.87, p <0.001). A publicly accessible web-based calculator and visual interface were developed to support planning workflows. Conclusions AI-based modeling enables accurate prediction of non-diagnostic outcomes in ION. The derived score provides a clinically interpretable decision-support tool that may enhance diagnostic yield and workflow efficiency. Prospective validation is underway. Legal entity responsible for the study The authors. Funding Has not received any funding. Disclosure All authors have declared no conflicts of interest. [163eP]