2701 - Radiomics-Clinical Integration Guides Prophylactic Cranial Irradiation Decisions in Limited-Stage Small Cell Lung Cancer: A Brain Metastasis Risk Stratification Model
Presenter(s)
Y. Zhou1, S. Yang1, B. Liu2, and N. Liu1,3; 1Tianjin Medical University Cancer Institute & Hospital, National Clinical Research Center for Cancer,Tianjin’s Clinical Research Center for Cancer,Key Laboratory of Cancer Prevention and Therapy, Tianjin, Tianjin, China, 2College of Arts and Sciences, Lehigh University, Bethlehem, PA, 3Hetian District People’s Hospital, Hetian District, China
Purpose/Objective(s): To develop a multimodal prediction model integrating radiomic and clinical features for stratifying brain metastasis (BM) risk in limited-stage small-cell lung cancer (LS-SCLC) and guiding personalized prophylactic cranial irradiation (PCI) strategies.
Materials/Methods: This single-center retrospective study analyzed 141 LS-SCLC patients (2013-2021) using simulation CT images and clinical records. Patients were stratified into training (n=98) and internal validation (n=43) cohorts, with external validation from an independent institution (n=24). Radiomic features were extracted from the CT images and optimized using the Max-Relevance and Min-Redundancy (MRMR) algorithm to compute a radiomic score (RadScore). Clinical predictors were identified through univariate logistic regression. Four machine learning models—logistic regression (LR), support vector machine (SVM), random forest (RF), and XGBoost (XGB)—were used to develop clinical, radiomic, and combined predictive models. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC). The top-performing models were used for risk stratification of BM and to assess the efficacy of PCI in prolonging brain metastasis-free survival (BMFS).
Results: A total of 1,037 radiomic features were extracted from the simulated positioning CT images, with 10 optimal features selected to form the RadScore. By incorporating dynamic changes in platelet count, hemoglobin levels, and leukocyte indices before and after radiotherapy, along with the baseline lymphocyte-to-monocyte ratio (LMR), the LR combined model demonstrated superior predictive capability. The model achieved AUC values of 0.831 (95% CI: 0.753–0.909) in the training cohort, 0.831 (95% CI: 0.698–0.963) in the internal validation cohort, and 0.863 (95% CI: 0.699–1.000) in the external validation cohort. The model significantly outperformed other machine learning approaches, including SVM, RF, and XGBoost (all P < 0.05 via DeLong’s test). Risk stratification revealed differential therapeutic efficacy of PCI: in the high-risk group, PCI reduced the risk of BM (hazard ratio [HR] = 0.270, 95% CI: 0.136–0.535; log-rank P < 0.001). In contrast, no significant benefit was observed in the low-risk group (HR = 0.225, 95% CI: 0.028–1.810; log-rank P = 0.127).
Conclusion: This study provides the first evidence-based framework for PCI personalization in LS-SCLC using pretreatment biomarkers, addressing critical unmet needs in balancing therapeutic efficacy against cognitive morbidity. Prospective trials are still needed to further assess the performance of the model.