Machine Learning-Driven Prediction of Gamma Knife Treatment Parameters and Outcomes

Abstract

Introduction:

Gamma Knife Radiosurgery offers a non-invasive treatment option to treating brain metastases often in surgically inoperable regions. However, determining optimal dosing regimens remains a clinical challenge due to patient outcome variability.

Methods:

In this study, we propose tree-based machine-learning models for personalized dose recommendations, tumor prediction progression, and follow-up treatment assessment using a dataset of over one thousand metastases treated at the UVA Health Gamma Knife Center between 1980 and 2025. The models leverage heterogeneous patient data from three distinct cancer types (breast, melanoma, and renal cell carcinoma), including histology, genetic markers, tumor location, and volume.

Results:

Our best-performing dose classification model achieved an F1-score of 0.82, while regression demonstrated low RMSE values at around 1.5 Gy. The models for tumor progression and patient-followup likelihood were impacted by class imbalance and are not yet powerful enough for decision making.

Discussion:

These exploratory models underscore the viability of data-driven approaches to support individualized Gamma Knife radiosurgical planning while generating insights into key predictors to inform clinical decision-making and improve patient outcomes.