3346 - Decoding Patient Heterogeneity Influencing Radiation-Induced Brain Necrosis

I. Chamseddine¹, K. Shah², H. Lee³, F. Ehret⁴, J. P. Schuemann⁵, A. Bertolet⁶, H. A. Shih⁷, and H. Paganetti²; ¹Department of Radiation Oncology, Massachusetts General Hospital, Harvard Medical School, Boston, MA, ²Massachusetts General Hospital, Harvard Medical School, Boston, MA, ³St. Jude Childrens Research Hospital, Memphis, TN, ⁴Charité - Universitätsmedizin Berlin, Berlin, Germany, ⁵Department of Radiation Oncology, Massachusetts General Hospital, Boston, MA, ⁶Harvard Medical School, Boston, MA, ⁷Department of Radiation Oncology, Massachusetts General Hospital and Harvard Medical School, Boston, MA

Purpose/Objective(s):

Outcome assessment in radiotherapy (RT) often suffers from the obscuring effect of patient heterogeneity. This effect is pronounced in brain necrosis, a significant concern in RT when treating central nervous system (CNS) and brain tumors, including proton therapy with its uncertain relative biological effect (RBE). Studying the effect of dose and linear energy transfer (LET) on necrosis is limited by this heterogeneity, impeding us to isolate the therapy impact. Consequently, we introduced a new clinically practical pipeline to elucidate patient heterogeneity, identifying variables that can clarify the relationship between dosimetric features and outcomes. Materials/

Methods: After processing patient, tumor, and treatment variables, we defined expert rules and statistical constraints to guide an analytical process. A Bayesian network was developed to determine the structural dependencies among variables. The network structure was learned from a cohort of 130 patients treated with proton therapy for tumors in the CNS, base of skull, or head and neck. The cohort included 50 necrotic cases identified based on imaging biomarkers and clinical symptoms. Each connection within the network was assessed for clinical sensibility using a 3-level grading system to evaluate the networks overall feasibility. A Markov blanket was analyzed to locate variables with a direct impact on necrosis risk. Results were evaluated using a suite of statistical metrics that collectively measure the impact of patient heterogeneity on outcome variability: log-likelihood ratio (LLR), integrated discrimination index (IDI), net reclassification index (NRI), and receiver operating characteristic (ROC).


Results:

The Markov blanket showed that key determinants influencing necrosis risk are tumor location (intracranial vs extracranial) and radiation proximity to critical brain structures—white matter, ventricles, and the frontal lobe. The network feasibility was validated, where the majority of connections (21/24) were supported by clinical evidence or were clinically sensible, and only 3/24 were considered incidental. Quantitative assessment confirmed the significance of these variables in stratifying patients, with LLR=12.17 (p=0.016), IDI=0.15, and NRI=0.74. The area under the ROC curve was 0.66 by relying solely on non-dosimetric variables without considering dose or LET, which underscores the discriminative power of the identified variables.
Conclusion:

This study identified key patient variables affecting brain necrosis after RT, which will facilitate the study of dosimetric effects by providing treatment confounders. The findings have the potential to influence clinical practice by identifying at-risk patients and offer insights for dose re-distribution. Although applied to brain necrosis, this pipeline serves as a versatile tool for analyzing the impact of patient heterogeneity on treatment outcomes and can be used in different disease sites.