Evaluation of acute hematological toxicity by machine learning in gynecologic cancers using postoperative radiotherapy
Melek Akcay1, Durmus Etiz1, Ozer Celik2, Alaattin Ozen1
1 Department of Radiation Oncology, Medical Faculty of Osmangazi University, Eskişehir, 26480, Turkey
2 Department of Mathematics - Computer Eskisehir Osmangazi University, Eskişehir, 26480, Turkey
Background: The aim of the study is to investigate the factors affecting acute hematologic toxicity (HT) in the adjuvant radiotherapy (RT) of gynecologic cancers by machine learning.
Methods: Between January 2015 and September 2018, 121 patients with endometrium and cervical cancer who underwent adjuvant RT with volumetric-modulated arc therapy (VMAT) were evaluated. The relationship between patient and treatment characteristics and acute HT was investigated using machine learning techniques, namely Logistic Regression, XGBoost, Artificial Neural Network, Random Forest, Naive Bayes, Support Vector Machine (SVM), and Gaussian Naive Bayes (GaussianNB) algorithms.
Results: No HT was observed in 11 cases (9.1%) and at least one grade of HT was observed in 110 cases. There were 55 (45.5%) cases with ≤grade 2 HT (mild HT) and 66 (54.5%) cases with grade ≥3 HT (severe HT). None of the patients developed grade 5 HT. Of 24 variables that could affect acute HT, nine were determined as important variables. According to the results, the best machine learning technique for acute HT estimation was SVM (accuracy 70%, area under curve (AUC): 0.65, sensitivity 71.4%, specificity 66.6%). Parameters affecting hematologic toxicity were evaluated also by classical statistical methods and there was a statistically significant relationship between age, RT, and bone marrow (BM) maximum dose.
Conclusion: It is important to predict the patients who will develop acute HT in order to minimize the side effects of treatment. If these cases can be identified in advance, toxicity rates can be reduced by taking necessary precautions. These cases can be predicted with machine learning algorithms.
Department of Radiation Oncology, Medical Faculty of Osmangazi University, Eskişehir, 26480
Source of Support: None, Conflict of Interest: None