Integrative AI Model Combining Radiomics and Phenomics to Predict Survival in Non-Small Cell Lung Cancer Patients Treated with Immunotherapy Containing Regimen

Authors

Abstract

PurposeImmunotherapy has improved outcomes in non-small cell lung cancer (NSCLC), but only a subset of patients achieves durable survival benefit. Conventional biomarkers such as PD-L1 expressions have limited predictive power. This study aimed to develop an artificial intelligence (AI)-based radiomic framework integrating deep learning derived and traditional radiomic features with clinical variables to predict survival outcomes in advanced NSCLC treated with immune checkpoint inhibitor (ICI)-containing regimens. MethodsNinety-four patients with stage IV NSCLC treated with ICI-containing regimens between 2013 and 2022 were analyzed. Baseline CT scans were segmented and processed using LIFEx software, and deep radiomic features were extracted using a UNET convolutional encoder-decoder. Models were trained using nested five-fold cross-validation. Performance was evaluated using area under curve (AUC) and calibration assessment. ResultsThe ensemble model combining clinical, traditional radiomic, and deep radiomic features achieved AUCs of 0.70 and 0.84 for 12- and 36-month overall survival (OS), and 0.65 and 0.73 for 12- and 36-month progression-free survival (PFS). For 12-month OS, the model integrating clinical variables with deep radiomic features showed the highest performance (AUC 0.78). Beyond deep features, traditional radiomic features reflecting tumor heterogeneity and morphology were the most influential predictors. Patients classified as high risk by the ensemble model had significantly shorter survival (OS HR 2.10, p = 0.005; PFS HR 1.74, p = 0.015). ConclusionIntegrating deep learning-derived radiomic features with clinical data improved survival prediction in advanced NSCLC treated with ICI-containing regimens. These findings support the potential of an AI-based radiomic risk stratification model as a noninvasive tool for individualized outcome prediction, warranting prospective validation. Key PointsO_ST_ABSQuestionC_ST_ABSCan combining deep learning-derived radiomic features with clinical variables improve prediction of survival outcomes in patients with non-small cell lung cancer (NSCLC) treated with immune checkpoint inhibitor-containing regimens? FindingsIn a retrospective cohort of 94 patients with stage IV NSCLC, an ensemble model integrating clinical, traditional radiomic, and deep learning-derived radiomic features showed strong discrimination for 12- and 36-month overall survival with an area under curve of 0.70 and 0.84, respectively. Radiomic markers of tumor heterogeneity and morphology were among the most influential factors. MeaningDeep learning-based radiomic-clinical modeling can enhance individualized survival prediction in NSCLC.

Tags

Topics