Evaluating Different Machine Learning Models for Predicting the Likelihood of Breast Cancer

Objective: With the continuous development of medical equipment and advancements in data collection, the size and features of the medical database are also increasing rapidly. However, the majority of current cancer detection still relies on doctors’ observation and test with cell tissue samples which does not seem to utilize the medical database to its full potential. Therefore, the research focus for this study was to examine if we can use artificial intelligence to identify patterns in medical databases that can be predictive of cancer detection. Accordingly, various machine learning (ML) methods were examined for their predictive performance with the perspective that how can they help the doctors in cancer diagnosis.

Methods: From the Breast Cancer Surveillance Consortium (BCSC) dataset, 154,899 screening mammograms records are applied in this research. This dataset includes 12 independent variables and 1 dependent variable, which was in the form of labeled data. We developed four prediction models using well-established machine learning algorithms: Logistic Regression, Naïve Bayes, Support Vector Machine (SVM), and Bayesian Network, by training them on 80% of the data and examined their prediction performance on remaining 20% data. We performed a comparative analysis of the performance of these machine learning models.

Results: Among the four models, Naïve Bayes showed the best prediction accuracy for the malignant samples, which is predictive of cancer likelihood. The Bayesian Network model performed the second best. Both Logistic Regression and SVM yielded poor prediction performance for predicting malignant cases and thus the breast cancer likelihood.