Predicting the progressive-free survival of ALK-positive lung cancer: machine learning algorithm based on habitat imaging.

Background: Lung cancer associated with the harbor oncogenic driver mutations or fusions, such as anaplastic lymphoma kinase (ALK), could benefit from representative clinical first-line treatment. And patients with ALK-positive (ALK+) show longer median survival than ALK-negative patients. This study aimed to develop a machine-learning model based on habitat analysis derived from computed tomography (CT) lung images to predict disease progression and explore its prognostic value for progression-free survival (PFS) in patients with (ALK+) lung cancer. Methods: A total of 79 patients with ALK-positive lung cancer and advanced TNM stage (III-IV) at The First Affiliated Hospital of Nanjing Medical University between April 2016 and August 2022 were divided into training (n=63) and testing (n=16) cohorts. Radiomics features were extracted from CT images using a habitat-based method with clustering analysis. Subsequently, the performance of eight machine learning algorithms was compared to build a habitat-radiomics (rad_habitat) model, yielding a rad_habitat_score. Using regression, we developed both a clinical model and a combined model that integrated clinical variables with the rad_habitat_score. Subsequently, these models' performance were evaluated. Univariate and multivariate Cox regression analyses were performed on clinical variables and the rad_habitat_score. Results: Three distinct habitats were identified. Among the eight machine learning algorithms, the multilayer perceptron (MLP) model demonstrated superior performance in handling non-linearity, achieving a higher area under the curve (AUC) of 0.836 [95% confidence interval (CI): 0.628-1.000] in the testing cohort. The combined model (AUC =0.836, 95% CI: 0.628-1.000) exhibited a higher AUC than the clinical model but a similar AUC to the rad_habitat model in the testing set. Delong test revealed a significant difference in the AUC values between the clinical model and the combined model and between the clinical model and rad_habitat model in the training cohort (P=0.002 and P=0.007, respectively), while the AUC values of the other models were not statistically different. Cox regression analysis identified the rad_habitat_score as the only independent predictor of PFS (P<0.001) for patients with high-risk ALK+ lung cancer. Conclusions: The habitat-radiomics approach, utilizing three regional habitats, shows promise for accurately and interpretably identifying disease progression in ALK+ lung cancer. The rad_habitat_score, when combined with clinical features, demonstrated improved predictive accuracy for prognostic stratification compared with the clinical model alone.