CT radiomics machine learning for predicting treatment response in unresectable stage III-IV NSCLC: insights from dual chemo-immunotherapy and immunotherapy cohorts.
Authors
Affiliations (5)
Affiliations (5)
- Department of Diagnostic Imaging, Sandro Pertini Hospital, Via Dei Monti Tiburtini 385, 00157, Rome, Italy.
- Department of Radiological, Oncological and Pathological Sciences, Policlinico Umberto I Hospital, Sapienza University of Rome, Viale Regina Elena 324, 00161, Rome, Italy.
- Department of Emergency Radiology, Policlinico Umberto I Hospital, Sapienza University of Rome, Viale del Policlinico 155, 00161, Rome, Italy.
- Department of Radiological, Oncological and Pathological Sciences, Policlinico Umberto I Hospital, Sapienza University of Rome, Viale Regina Elena 324, 00161, Rome, Italy. [email protected].
- Department of Emergency Radiology, Policlinico Umberto I Hospital, Sapienza University of Rome, Viale del Policlinico 155, 00161, Rome, Italy. [email protected].
Abstract
The purpose of this study was to develop a machine learning model based on radiomic features extracted from baseline contrast-enhanced CT scans for predicting early treatment responses in two cohorts of patients with stage III-IV NSCLC treated with distinct therapeutic regimens (first-line chemo-immunotherapy or immunotherapy alone). In this retrospective bicentric study including two cohorts, patients with a confirmed diagnosis of advanced NSCLC (stage III-IV) were retrospectively collected. At Center 1, after application of the exclusion criteria, 90 eligible patients (52 treated with CHT/IT and 38 with IT) were included for model development and internal validation. An independent external cohort from Center 2 comprised 40 additional patients, including 20 treated with CHT/IT and 20 with IT, and was used for external validation. Radiomic features were extracted from segmented tumor volumes using the Trace4Research™ platform, and multiple machine learning models, including random forest, support vector machine, K-nearest neighbors, multilayer perceptron, and logistic regression, were trained and compared. In the CHT/IT cohort, seven radiomic features were retained after outlier removal. Among the tested models, the support vector machine (SVM) classifier showed the best performance, with a ROC-AUC of 0.90 (95% CI: 0.82-0.97), accuracy of 82%, sensitivity of 82%, specificity of 83%, positive predictive value of 67%, and negative predictive value of 94% (p < 0.005). In the IT cohort, 29 radiomic features were retained, and the SVM again achieved the best performance, with a ROC-AUC of 0.84 (95% CI: 0.73-0.95), accuracy of 77%, sensitivity of 82%, specificity of 70%, positive predictive value of 80%, and negative predictive value of 83% (p < 0.05). External validation supported the robustness and generalizability of both models. Machine learning models based on CT-derived radiomic features may enable accurate and noninvasive early prediction of treatment response in patients with advanced NSCLC receiving chemo-immunotherapy or immunotherapy alone, supporting the potential role of radiomics as an imaging biomarker for treatment stratification and personalized therapeutic planning.