Öznacar Tuğçe, Aral İpek Pınar, Zengin Hatice Yağmur, Tezcan Yılmaz
Department of Biostatistics, Faculty of Medicine, Ankara Medipol University, 06570 Ankara, Turkey.
Radiation Oncology Clinic, Faculty of Medicine, Ankara Yıldırım Beyazıt University, 06800 Ankara, Turkey.
Brain Sci. 2025 Mar 1;15(3):266. doi: 10.3390/brainsci15030266.
Accurate survival prediction for brain metastasis patients undergoing stereotactic radiotherapy (SRT) is crucial for personalized treatment planning and improving patient outcomes. This study aimed to develop a machine learning model to estimate survival times, providing clinicians with a reliable tool for making informed decisions based on individual patient characteristics. The goal was to compare the performance of multiple algorithms and identify the most effective model for clinical use.
We applied a hybrid machine learning approach to predict survival in brain metastasis patients treated with SRT, utilizing real-world data. Four algorithms-XGBoost, CatBoost, Random Forest, and Gradient Boosting-were compared within a meta-model framework to identify the most accurate for survival prediction. Model performance was evaluated using metrics such as MSE, MAE, MAPE, and C index.
XGBoost outperformed all other algorithms, achieving an MSE of 0.14%, MAE of 0.10%, and MAPE of 0.093%, with a high C-index of 100%. CatBoost showed reasonable performance, while Gradient Boosting had higher error rates (MSE of 6.99%, MAE of 21.04%, MAPE of 19.29%). Random Forest performed the weakest, with the highest MSE (14.39%), MAE (30.23%), and MAPE (33.58%).
Inputting relevant clinical variables into the model enables clinicians to obtain highly accurate survival predictions for patients with brain metastasis. This enhances clinical decision making by providing a more precise understanding of expected outcomes. The XGBoost-based hybrid model showed exceptional accuracy in predicting survival for brain metastasis patients after SRT, offering valuable support for clinical decision making. Integrating machine learning into clinical practice can improve treatment planning and personalize care for these patients.
对于接受立体定向放射治疗（SRT）的脑转移患者，准确的生存预测对于个性化治疗计划和改善患者预后至关重要。本研究旨在开发一种机器学习模型来估计生存时间，为临床医生提供一个基于个体患者特征做出明智决策的可靠工具。目标是比较多种算法的性能，并确定临床应用中最有效的模型。
我们应用一种混合机器学习方法，利用真实世界数据预测接受SRT治疗的脑转移患者的生存情况。在一个元模型框架内比较了四种算法——XGBoost、CatBoost、随机森林和梯度提升——以确定生存预测最准确的算法。使用均方误差（MSE）、平均绝对误差（MAE）、平均绝对百分比误差（MAPE）和C指数等指标评估模型性能。
XGBoost的表现优于所有其他算法，MSE为0.14%，MAE为0.10%，MAPE为0.093%，C指数高达100%。CatBoost表现出合理的性能，而梯度提升的错误率较高（MSE为6.99%，MAE为21.04%，MAPE为19.29%）。随机森林表现最差，MSE最高（14.39%），MAE（30.23%）和MAPE（33.58%）。
将相关临床变量输入模型可使临床医生获得脑转移患者高度准确的生存预测。这通过更精确地了解预期结果来加强临床决策。基于XGBoost的混合模型在预测SRT后脑转移患者的生存方面表现出卓越的准确性，为临床决策提供了有价值的支持。将机器学习整合到临床实践中可以改善治疗计划并为这些患者提供个性化护理。
