239P Dual AI models for perioperative decision support in resectable NSCLC: A real-world cohort analysis

Background Accurate perioperative assessment is crucial in early-stage NSCLC, yet current staging and prognostic tools remain limited. We developed and internally validated two machine learning models using the IBM XGBoost algorithm: (1) to predict the likelihood of pathological upstaging (≥ stage II) at surgery based on perioperative data; (2) to estimate the probability of achieving a favorable postoperative outcome, defined as no evidence of disease (NED) at follow-up, versus non-NED. These AI-driven tools aim to enhance decision-making and personalize treatment strategies. Methods A retrospective analysis of 197 NSCLC patients undergoing curative resection (2023–2025) was performed. Models incorporated demographic, clinical, pathological, and treatment data, using an 80/20 stratified train-test split. Performance was evaluated using ROC-AUC, precision, recall, F1-score, precision-recall AUC (AUC-PR), which was used for imbalanced data. SHAP (Shapley Additive Explanations) was applied post hoc to support clinical interpretation of feature contributions. Results Model 1 predicted advanced pathological stage (≥ II) in 197 NSCLC patients (158 stage I, 39 ≥ II), achieving 86.7% accuracy, ROC-AUC 0.95, and AUC-PR 0.71. For the positive class: recall 75%, specificity 88.5%, precision 50%, and F1-score 0.60. SHAP analysis confirmed the relevance of key features, consistent with known prognostic markers. Model 2 targeted the binary outcome NED vs non-NED (172 vs 25 patients), with 93.9% accuracy, ROC-AUC 0.90, and AUC-PR 0.70. For the non-NED class: recall 33%, specificity 84%, precision 100%, and F1-score 0.50. This dual-model approach highlights the potential of AI to study the likelihood of pathological upstaging and postoperative disease status, with implications for staging and treatment strategies. Conclusions This dual-model XGBoost-based framework offers interpretable and scalable AI-based decision-driven tools for decision-making in NSCLC surgery, enabling more accurate preoperative staging and individualized postoperative prognostication. Ongoing external validation will determine its generalizability and readiness for clinical integration, including potential application in decision-support software. Legal entity responsible for the study C. Bardoni. Funding Has not received any funding. Disclosure All authors have declared no conflicts of interest. [239P]