ESTRO 2024 - Abstract Book

S794

Clinical - CNS

ESTRO 2024

785

Proffered Paper

Machine Learning to predict time to local failure & radionecrosis post-brain metastasis radiosurgery

Sreenija Yarlagadda 1 , Yanjia Zhang 2 , Anshul Saxena 2 , Tugce Kutuk 1 , Ranjini Tolakanahalli 1,3 , Haley Appel 1 , Alonso La Rosa 1 , Matthew D Hall 1,3 , D Jay Wieczorek 1,3 , Yongsook C Lee 1,3 , Michael W McDermott 4 , Alonso N Gutierrez 1,3 , Minesh P Mehta 1,3 , Rupesh Kotecha 1,3 1 Miami Cancer Institute, Baptist Health South Florida, Department of Radiation Oncology, Miami, USA. 2 Miami Cancer Institute, Baptist Health South Florida, Department of Biostatistics, Miami, USA. 3 Herbert Wertheim College of Medicine, Florida International University, Department of Radiation Oncology, Miami, USA. 4 Miami Neuroscience Institute, Baptist Health South Florida, Department of Neurosurgery, Miami, USA

Purpose/Objective:

Predicting time to local failure (LF) or development of radiation necrosis (RN) requires consideration of numerous patient-related, disease-specific, and treatment-related variables. The objective of this study was to develop machine-learning (ML) algorithms that would systematically evaluate all of these pertinent variables to estimate the time to LF or RN using a comprehensive database of relevant patient, disease-specific, and treatment-related (radiotherapy and systemic therapy) factors.

Material/Methods:

Consecutive patients with small brain metastasis (≤2 cm in maximum dimension) treated with single fraction stereotactic radiosurgery (SRS) to 20-24 Gy between 2017 and 2021 at a single tertiary-care institution were evaluated. Patient characteristics, treatment details, and outcomes were extracted. Random survival forest analysis (with death as a competing risk) was utilized to predict the duration of local control (LC) and time to RN accounting for various factors, such as Karnofsky Performance Status (KPS), the total number of lesions, tumor size, status and burden of extra-cranial disease, and type of post-SRS systemic therapy administered (immunotherapy, targeted therapy, chemotherapy). The data were divided into training (75%) and test (25%) sets to develop and evaluate the models. In each model, log rank was used to split the nodes, and the cause-specific cumulative incidence function (CIF) at the time points of interest was used as the terminal node statistic. Then, the averaged out-of-bag (OOB) performance error based on Harrell’s C -index was used to evaluate the model, and the variable importance was estimated by comparing the performance of the estimated model with and without the variables.

Results:

From a dataset of 1318 brain metastasis in 250 patients, 1160 lesions that had at least one follow-up imaging post treatment were analyzed to build these models. The median age was 65 years (range: 18-90), and 63% were female. The median KPS was 90 (range: 50-100) and the median number of lesions was 8 (range: 1-25). With a median follow-up of 13 months, 119 (10.3%) LFs in 42 patients and 62 (5.3%) RN events in 44 patients were observed. The median time to LF was 10 months (95% CI-5-25 months) and the median time to RN was 12 months (95% CI: 5-25