199 - Early Adaptive Interventions in Lung Cancer: Leveraging Fusion of Longitudinal CBCT Trajectories and Clinical Variables for Robust Survival Prediction

Purpose/Objective(s): Traditional prognostic models for lung cancer rely on static pre-treatment factors, failing to capture the dynamic response of tumors during radiotherapy. Although CBCT provides serial imaging of this evolution, snapshot and delta radiomics fail to capture the full response dynamics. We propose a novel cumulative longitudinal framework that integrates clinical data with CBCT-derived trajectories to enable early identification of high-risk patients.

Materials/Methods: This retrospective study analyzed 225 radiotherapy courses from 189 lung cancer patients (median age 71) treated at a single institution from 2019 to 2024. The cohort underwent diverse treatment regimens (median dose 60 Gy, range 30–69 Gy; median 30 fractions, range 3–37), consisting of 83% NSCLC, 7.3% Stage I, and 56% Stage III cases. Across 225 planning CTs and 5,067 CBCT scans, we extracted 107 radiomic features from each lesion-level contour (GTV/PTV) for every imaging set. 14 clinical variables were collected for each patient, including age, smoking history, TNM stage, histopathology type, etc. Hierarchical aggregation was used to integrate patient-level clinical data with lesion-level radiomic features. We utilized a hierarchical Gradient Boosting Survival model to aggregate multi-lesion data to the patient level, followed by cumulative temporal integration (e.g., CBCTcN incorporates all scans from fraction one through week N). We evaluated which CBCT week was most predictive of overall survival. The model's ability to predict overall survival was assessed using the concordance index (C-index), and stability was evaluated using the coefficient of variation (CV). All models underwent 5-fold patient-level cross-validation.

Results: The cumulative model reached its peak predictive accuracy by Week 2 (CBCTc2; C-index=0.72, CV=6.12%), indicating that prognostic signals emerge early in the treatment course. The complete six-week model (CBCTc6; C-index=0.72, 95% CI: 0.69–0.75) maintained accuracy while achieving exceptional stability. The CV decreased monotonically from 8.47% in Week 1 to 1.97% in Week 6, demonstrating progressive noise reduction with temporal aggregation. Compared with baseline approaches, cumulative CBCT radiomics showed performance superior to clinical-only (C-index=0.61, p<0.001), planning CT radiomics (C-index=0.66, p=0.015), and delta-radiomics (C-index=0.70, p=0.006). Integration of clinical variables did not significantly improve discrimination.

Conclusion: A cumulative CBCT-based radiomics approach provides strong prognostic accuracy by the second treatment week, enabling the earlier identification of high-risk patients. Incorporating all six weeks of imaging further stabilizes predictions. Leveraging standard-of-care imaging and clinical data, this framework offers a practical tool for adaptive radiotherapy.