ИСТИНА

Objective: It is extremely important for cancer research to predict outcomes and detect the key features from complex datasets. The Machine Learning (ML) is a powerful tool to achieve these goals. The purpose of this study is to apply the ML to predicting the overall survival (OS) after Gamma Knife radiosurgery (RS) in case of patients with brain metastases (BS). Methods: We retrospectively evaluated the total data of 589 patients with 34 patient characteristics who had received the RS in the period from 2009 to 2016. The data were gathered in the self-designed database. 146 patients were excluded from analysis due to incompleteness of the data. 443 patients with 7 features (age, Karnofsky performance status (KPS), total volume of all lesions, maximal volume of the lesion, number of lesions, presence of extracranial metastases, diagnosis) were included into analysis and subdivided into two sets - training and testing ones. We applied three ML algorithms (the Decision Tree (DT), the k-nearest neighbors algorithm (kNN) and the Random Forest (RF)) to the task of classification to predict OS. Results: For evaluating the classification quality, we use the accuracy metrics, which means the number of correct predictions. The predictors had the following classification accuracy in the cross - validation estimation as well as in the held-out dataset: the DT 78% and 61,6 % , the KNN 88,6 % and 69,9%, the RF 78,4 % and 67,6 % respectively . We have found that the most important features for predicting the OS are the total volume of the lesions, the maximal volume of the lesion, the age and the number of metastases. The features are mentioned in order of priority. Conclusions: This research is just a first step to apply the ML algorithms to prediction outcomes in case of patients with BS in the Moscow Gamma Knife center. The kNN shows the better results in the cross - validation than the other algorithms; and besides, the most important feature is the total volume of the lesions. The accuracy of the study can be improved by the number of patients and completeness of the patient data. It could be more correct to apply the other metrics to evaluate the classification. The database system integration with the machine learning algorithms can lead to a new level in development of both the predictive models and the making decision process within the clinical practice.