To investigate the predictive value of intratumoral heterogeneity (ITH) quantified by computed tomography (CT)-based habitat radiomic features for the response to immune checkpoint inhibitors (ICIs) in patients with non-small cell lung cancer (NSCLC).

A total of 148 NSCLC patients scheduled to receive pembrolizumab plus platinum-based doublet chemotherapy at our hospital from June 2022 to June 2025 were enrolled as the training cohort, and another 98 patients treated during the same period served as the validation cohort. Pre-treatment CT images were acquired and habitat radiomic features were extracted. LASSO regression was used to select core features for constructing the ITHscore model. After treatment completion, treatment response was assessed according to RECIST criteria. In the training cohort, patients with complete response or partial response were classified as the response group (n=79), while those with stable disease or progressive disease were classified as the nonresponse group (n=69). Clinical data were compared between the two groups. Univariate and multivariate logistic regression analyses were performed to identify independent factors associated with ICI response. Predictive performance was evaluated using receiver operating characteristic (ROC) curves and decision curve analysis (DCA).

In the training cohort, significant differences were observed between the response and nonresponse groups in age, sex, clinical stage, and programmed deathligand 1 (PDL1) expression level (all P<0.05). LASSO regression selected 8 core features with nonzero coefficients from 1, 524 radiomic features to construct the ITHscore. Multivariate logistic regression showed that age (OR=1.145, 95%CI: 1.031~1.272, P=0.011), sex (OR=2.447, 95%CI: 1.079~5.550, P=0.032), clinical stage (OR=3.555, 95%CI: 1.581~7.992, P=0.002), PDL1 expression (OR=0.887, 95%CI: 0.847~0.929, P=0.007), and ITHscore (OR=1.895, 95%CI: 1.261~2.848, P=0.002) were independent predictors of ICI response. ROC analysis showed that the AUC of the ITHscore model was 0.818 (95%CI: 0.737~0.882), and that of the clinical feature model was 0.793 (95%CI: 0.710~0.862) (P>0.05). The combined model achieved an AUC of 0.896 (95%CI: 0.827~0.944), with a sensitivity of 75.61% and specificity of 91.23%, outperforming either single model. In the validation cohort, the combined model yielded an AUC of 0.879(95%CI: 0.807~0.931), sensitivity of 70.73%, and specificity of 92.41%. DCA demonstrated that at a threshold probability of 0.25, the net benefit of the combined model (0.27) was higher than that of the ITHscore model (0.23) and the clinical feature model (0.11).

The ITHscore combined with clinical features, constructed from CTbased intratumoral heterogeneity radiomic features, demonstrates good discriminative ability and clinical net benefit for predicting ICI response in NSCLC patients, and may facilitate early identification of patients likely to benefit from immunotherapy.