Using machine learning algorithms in determining the stage of breast cancer from pathology reports

Introduction: After a cancer diagnosis, the most important thing is to determine the stage and grade of the cancer. Pathology reports are the main source for cancer staging, but they do not contain all the information needed for the staging. However, the text of these reports is sometimes the only available information. We were interested in knowing whether text mining methods can be used to predict staging only from pathology reports. Material and Methods: A total of 698 pathology reports of breast cancer cases and their TNM staging collected from multiple centers in West Azerbaijan Province, Iran were used for this study. After preparing the semi-structured reports, the texts of the reports were imported into a program written by Python V3. Three machine learning algorithms of Logistic Regression, SVM, and Naïve Bayes and a simple pipeline were used for the purpose of text mining. The performance of the algorithms was evaluated in terms of accuracy, precision, recall, and F1 score. Results: The Naïve Bayes algorithm achieved excellent results and a value rate of higher than 91% in all evaluation criteria (accuracy, precision, recall and F1 score). This means that the Naïve Bayes algorithm could classify the reports with high efficiency and its predictions were more correct than the other two algorithms. Naïve Bayes also outperformed SVM and Logistic Regression in terms of accuracy, recall and F1 score. In addition, Naïve-Bayes showed faster inference due to its simplicity and lower computational and training time. Conclusion: We suggest using the proposed design in this study for predicting breast cancer staging, where there is a need but not all necessary information except pathology reports. This method may not be a useful for clinical management of cancer patients, but it can be safely used for epidemiological estimations.