ИСТИНА |
Войти в систему Регистрация |
|
ИСТИНА ИНХС РАН |
||
The aim of this study was to predict overall survival in patients with brain metastases after stereotactic radiosurgery and identify the prognostic factors using machine learning (ML) approach. 1023 patients treated with SRS between January 2005 and December 2017 were analyzed. 26 features (clinical, biological and morphological) were selected to develop a prognostic model to predict overall survival for these patients. The target variable was time from the date of oncological diagnosis to the date of death. The data were divided into two sets: a training dataset and a test dataset. The training dataset consisted of 577 patients with different features, for which the target variable was known. The overall survival classes were: less than 8 months (292 patients), more than 10 months (285 patients). The second class ( OS > 10 months) included 92 patients that were alive at the day of the analysis. Splitting the group of patients on those classes supports class balancing and minimizes overfitting probability. The machine learning technique “Gradient boosting” was used to identify the favorable prognostic factors associated with long term overall survival. The 5 - folds cross validation technique was performed to estimate the accuracy of the predictive model. The most significant features were: age at the time of diagnosis, total volume of brain metastases, maximum metastasis volume, time from primary diagnosis to brain metastases, time from brain metastases to the first radiosurgery. The accuracy of the classification model was 0.81. The accuracy computed from the confusion matrix was 0.80. The predictor had an average area under the curve (AUC) of 0.87. To classify patients into risk classes is the important step in making therapeutic decisions. The predictive model demonstrates accuracy of the brain metastases patients classification. Machine learning techniques seem to be a very promising tool for clinical decision making. But the principal challenge and the key ingredient of successful application of ML in radiation oncology is data collection from the multimodal data sources. ML techniques will be fully integrated in clinic routine only if it combines with modern databases provided routine data collection