Development and validation of a computed tomography-based radiomics-clinical model to preoperatively predict high-grade patterns within lung invasive adenocarcinoma.

Background: Lung cancer remains a leading cause of cancer mortality globally, with non-small cell lung cancer (nSCLC), particularly invasive adenocarcinoma (IAC), being predominant. Within IAC, high-grade patterns (HGPs) are strongly linked to aggressive behavior and poor prognosis. Unfortunately, preoperative identification of HGPs is challenging: biopsy (histopathology gold standard) suffers from sampling limitations, while conventional computed tomography (CT) lacks sensitivity. Radiomics offers promise by extracting quantitative features from CT images. This study aimed to develop and validate a CT-based radiomics-clinical integrated model for preoperative prediction of HGPs in IAC patients, thereby guiding optimal surgical decision-making. Methods: A total of 278 patients (63.3±10.6 years) were divided into an internal cohort (n=240) for model development and validation, and a preoperative prediction cohort (n=38). The internal cohort was randomly split into training (70%) and test (30%) sets. CT radiomics features were extracted, followed by feature selection via the least absolute shrinkage and selection operator (LASSO) regression. Machine learning algorithms were used to construct a radiomics model and a clinical model, with a combined radiomics-clinical model being generated via logistic regression (LR). Model performance was evaluated by area under the curve (AUC), accuracy, sensitivity, and specificity, with subgroup analysis and preoperative prediction cohort test. Results: The combined radiomics-clinical model showed the best predictive performance. In the training set, an AUC of 0.934, an accuracy of 0.869, a sensitivity of 0.765, and a specificity of 0.966 were achieved. In the test set, an AUC of 0.854, an accuracy of 0.806, a sensitivity of 0.846, and a specificity of 0.758 were achieved. The calibration curve showed good alignment between the predicted and actual outcomes, and decision curve analysis (DCA) demonstrated greater clinical benefits compared to the other models. Subgroup analysis revealed better accuracy for lesions with a CT diameter ≤2 cm and a mixed ground-glass appearance, with AUCs of 0.907 and 0.909, respectively. The preoperative prediction cohort test achieved an accuracy of 92.1%. Conclusions: We have developed and validated a CT-based radiomics-clinical model for preoperatively predicting HGPs in IAC, particularly for lesions with a diameter smaller than 2 cm or presenting as mixed ground-glass nodules, which provides valuable predictive insights for clinical diagnosis and treatment.