2217 - Machine Learning-Based Clinico-Radiomic Model for Predicting Brain Metastasis Velocity after Stereotactic Radiosurgery: A Multi-Institutional Validation Study
Presenter(s)
C. S. Shieh1, C. Y. J. Hsu2, W. Wang3, and H. L. Lee4; 1Taipei Medical University Hospital, Taipei 10449, Taiwan, 2Department of Oncology, National Taiwan University Hospital, Taipei, Taiwan, 3National Taiwan University, Taipei, Taiwan, 4TMU Proton Center, Taipei City, Taiwan
Purpose/Objective(s): Brain metastasis velocity (BMV) quantifies the rate of distant brain failure (DBF) after stereotactic radiosurgery (SRS) and predicts overall survival (OS). However, because BMV is defined only after the first DBF, its value at initial brain metastases (BMs) diagnosis is limited. We developed and externally validated a machine learning–based clinicoradiomic (CR) model using pre-SRS imaging features to predict high BMV (BMV-H) at the time of initial BMs diagnosis.
Materials/Methods: A total of 319 patients with newly diagnosed BMs treated with upfront SRS were retrospectively included from two institutions: National Taiwan University Hospital (NTUH; n=256, development cohort) and Taipei Medical University Hospital (TMUH; n=63, external validation cohort). BMV-H was defined as =4 new BMs per year or leptomeningeal disease at first DBF; others were classified as BMV-L. Radiomic features were extracted from pre-SRS contrast-enhanced T1-weighted MRI and contrast-enhanced CT across two tumor subregions (GTV core and edge) using PyRadiomics. Features with intraclass correlation coefficient <0.75 and highly correlated features (r>0.9) were excluded. Four subregion-specific LightGBM models were developed using nested five-fold cross-validation with hyperparameter optimization, and combined into a radiomic signature (RS). The final clinicoradiomic (CR) model integrated RS with clinical predictors (number of BMs, perilesional edema, extracranial progression) using LightGBM. Performance was assessed by AUC (DeLong test) and calibration (Hosmer–Lemeshow test). Survival between CR-predicted BMV groups was compared using Kaplan–Meier analysis and the log-rank test.
Results: Median follow-up was 20.2 months (NTUH) and 16.7 months (TMUH). DBF occurred in 42.5% and 47.6%, with BMV-H in 23.8% (61/256) and 14.3% (9/63), respectively. The CR model achieved an AUC of 0.78 (95% CI, 0.72–0.84) in the development cohort and 0.76 (95% CI, 0.61–0.91) in external validation. At the optimal threshold (Youden index; specificity =85%), sensitivity/specificity were 73.8%/85.6% (development) and 66.7%/87.0% (validation). CR-predicted BMV-L patients had significantly longer median OS after initial SRS than BMV-H patients in both cohorts (48.3 vs. 16.5 months, P<.001; 29.1 vs. 11.8 months, P=.008). Among patients with DBF, median OS after first DBF was also longer in CR-predicted BMV-L vs. BMV-H (25.4 vs. 12.6 months, P=.019; 23.2 vs. 10.3 months, P=.035).
Conclusion: Our multi-institutional study demonstrates that a LightGBM-based CR model integrating pre-SRS radiomic features with clinical predictors provides reliable prediction of BMV and significantly stratifies survival outcomes after both initial SRS and first DBF. This model may serve as a decision-support tool to personalize radiation treatment strategies and surveillance frequencies for patients with newly diagnosed BMs.