1229 Leveraging Machine Learning for Pre-operative Prediction of Supralesional Ablation in Laser Interstitial Thermal Therapy (LITT) for Brain Tumors

Abstract

INTRODUCTION: Although maximizing the extent of ablation (EOA) following laser interstitial thermal therapy (LITT) can improve outcomes in patients with brain tumors, the pre-operative factors that predict a supralesional ablation (EOA >100%) remain unclear. METHODS: A retrospective study of patients with glioblastoma, metastasis, or radiation necrosis receiving LITT at a tertiary care referral center between 2013 to 2023 was conducted. Using the EOA >100% as a primary outcome, several clinical and radiographic pre-operative factors were evaluated. Five ML models, Decision Trees, Random Forest, Gradient Boosting, XGBoost, and AdaBoost, were trained using an 80/20 training/testing split and 5-fold cross validation. The models were optimized using hyperparameter tuning and evaluated against traditional logistic regression. The area under the receiver operating characteristic curve (AUC-ROC) was used to evaluate model performance. The optimal pre-operative volume cutoff for constituting an EOA >100% was also explored. RESULTS: 201 patients were included in the final cohort. Multivariate logistic regression showed that lower pre-operative volume (p < 0.001) and deep-seated lesions (p = 0.041) have a greater likelihood to undergo a supralesional ablation. Of the five ML models analyzed, Random Forest had the highest F1-score of 0.889, and XGBoost had the highest AUC-ROC of 0.74. Additionally, Random Forest (p = 0.02), Gradient Boosting (p < 0.001) and XGBoost (p = 0.01) showed statistically improved predictive performance compared to logistic regression. Features of importance were pre-operative volume, deep-seated lesions, and ellipsoid ratio. The pre-operative lesion volume cutoff for achieving an EOA >100% was 4.38 cc. CONCLUSIONS: This study shows the improved predictive ability of ML models over a traditional regression model in determining pre-operative factors for a supralesional ablation of brain tumors following LITT. Notably, the ML models identified pre-operative tumor volume, deep-seated tumors, and ellipsoid tumors as important factors, which could optimize treatment decisions and patient outcomes.