Abstract
BACKGROUND: Although immune checkpoint inhibitors (ICIs) have shown durable clinical benefits in a subset of patients with non-small cell lung cancer (NSCLC), robust biomarkers for predicting treatment response and guiding individualized immunotherapy remain lacking. Current prognostic models based on programmed death-ligand 1 (PD-L1) expression and clinical factors are insufficient for precise risk stratification. The aim of this study was to develop and validate a multi-institutional habitat imaging-based model to predict clinical outcomes of first-line immunotherapy in advanced NSCLC. METHODS: This retrospective multi-cohort study included a discovery cohort of 128 stage IIIB-IV NSCLC patients treated with anti-PD-(L)1 combination therapy from the ORIENT-11 trial, and two external validation cohorts consisting of 60 and 32 real‑world patients, respectively. Progression-free survival (PFS) was used as the primary outcome. For each patient, arterial‑phase contrast-enhanced computed tomography (CECT) images were processed using a habitat analysis approach to segment intratumoral subregions, extract radiomic features, and construct machine learning models. The predictive value of habitat imaging alone and in combination with PD-L1 tumor proportion score (TPS) and clinical factors was evaluated using the area under the receiver operating characteristic curve (AUC). RESULTS: Across the three cohorts, the mean age ranged from 59.8±9.1 to 62.4±9.7 years, with a predominance of male patients (77-92%), stage IV disease, and adenocarcinoma histology; the distribution of PD-L1 TPS was comparable among cohorts. Patients with high and low risk of disease progression showed significantly different proportions of specific intratumoral habitat clusters. Using intratumoral habitat imaging alone, the model achieved an AUC of 0.758 in predicting response to anti-PD-(L)1 combination therapy. When integrating habitat imaging with PD-L1 TPS and clinical metrics, the AUC reached 0.869. Furthermore, Kaplan-Meier survival analysis for PFS showed a statistically significant difference for grouping based on TPS ≥50% (P=0.03) and for grouping based on intratumoral habitat imaging (P=0.007). CONCLUSIONS: Habitat imaging is a potential valuable approach for predicting ICI efficacy in NSCLC patients. While this approach stratifies patients into distinct prognostic groups, its clinical utility requires further validation in larger prospective, multi center studies, inclusion of lymph node and metastatic lesions, and assessment across different histological subtypes.