S2291
Physics - Machine learning and AI algorithms
ESTRO 2026
Surgery, IRCCS "Regina Elena" National Cancer Institute, Rome, Italy. 6 Radiotherapy Unit, A. Gemelli University Hospital Foundation IRCCS, Rome, Italy. 7 Institute of Pathology, A. Gemelli University Hospital Foundation IRCCS, Rome, Italy. 8 Medical Oncology, A. Gemelli University Hospital Foundation IRCCS, Rome, Italy. 9 Interventional Pulmonology Unit, A. Gemelli University Hospital Foundation IRCCS, Rome, Italy. 10 National Center for Tumor Diseases (NCT), NCT/UCC, Dresden, Germany. 11 HZDR, Helmholtz-Zentrum Dresden-Rossendorf, Dresden, Germany Purpose/Objective: Predicting clinical outcomes to effectively guide patient care in Non-Small Cell Lung Cancer (NSCLC) is a main challenge. These outcomes are driven by a complex interplay of clinical, genomic, radiomic, and physiological factors. This complexity necessitates the development of sophisticated, multimodal AI models that can move beyond single-variable analysis and provide holistic, patient-specific predictions instead. This study introduces an advanced, interpretable machine learning workflow designed to help clinicians make informed decisions in patient care1. We tested and validated this workflow by applying it to different clinical outcomes. Material/Methods: The core of this work is a novel, multi-stage XGBoost- based2 workflow that allows the integration of any multi-omic tabular dataset and is designed to build a stable and powerful predictive model (Figure 1). First, the dataset was split into cross-validation (CV) folds. Within each training fold, we performed parameter optimization and a nested CV for each omics modality independently, selecting only consistently important features for stability. Next, all stable features from all modalities were aggregated. A second feature selection round was performed on this merged set to identify crucial cross-modal interactions. Finally, the model was validated on the internal test set and, when available, an independent external cohort. SHapley Additive exPlanations (SHAP3) was then applied to provide full model interpretability.
Results: The workflow is tested on 271 NSCLC patients from LANTERN4 Consortium and on an independent external dataset of 232 patients. As a first validation, the workflow was applied to predict postoperative complications for NSCLC patients who underwent curative surgery for Stage I-III. The model demonstrated robust performance on an external cohort (AUC 0.68 [0.61-0.75]), and confirmed the high importance of pre-operative FEV1 ratio. The workflow was then extended to predict high PDL1 (>50%) expression, achieving an AUC of 0.69 [0.61-0.75] on the internal validation set. The SHAP analysis provided granular, patient-specific decision support. It generated explanation plots that show clinicians how each feature contributed to an individual patient's final prediction, allowing for a precise, interpretable basis for making clinical decisions (Figure 2).
Made with FlippingBook - Share PDF online