Predicting Tumor Dynamics Post-Staged GKRS: Machine Learning Models in Brain Metastases Prognosis

Area of Science:

• Oncology
• Radiosurgery
• Machine Learning

Background:

• Brain metastases (BM) treatment requires accurate prediction of tumor dynamics.
• Gamma Knife radiosurgery (GKRS) is a common treatment modality for BM.
• Evaluating machine learning models for predicting post-treatment tumor behavior is crucial.

Purpose of the Study:

• To assess and compare the predictive performance of six machine learning models and a 1D Convolutional Neural Network (CNN) for forecasting tumor dynamics within three months after GKRS.
• To evaluate the impact of hyperparameter tuning on model performance.
• To identify key features influencing treatment outcome predictions.

Main Methods:

• Utilized data from 77 brain metastasis patients undergoing GKRS.
• Assessed six machine learning models and a 1D CNN.
• Evaluated model performance using accuracy, AUC, and confusion matrix metrics before and after hyperparameter tuning.
• Analyzed feature importance across different models.

Main Results:

• The 1D CNN achieved 98% accuracy and 0.97 AUC.
• Before tuning, XGBoost showed the highest accuracy (0.95) and AUC (0.95).
• After tuning, Support Vector Machine (SVM) demonstrated the best performance with 0.98 accuracy and 0.98 AUC.
• XGBoost performance declined post-tuning, suggesting potential overfitting.
• Key predictive features included 'control at one year', 'age of the patient', and 'beam-on time for volume V1 treated'.

Conclusions:

• Model selection and hyperparameter tuning are critical for accurate prediction of tumor dynamics after GKRS.
• SVM and CNN models show strong potential for forecasting treatment outcomes in BM patients.
• Features related to long-term control, patient demographics, and treatment parameters are significant predictors.
• Further multicenter research is needed due to the small cohort size and single-institution data.