Robust predictive model for evaluating breast cancer survivability

ObjectiveMany machine learning models have aided medical specialists in diagnosis and prognosis for breast cancer. Accuracy has been regarded as a primary measurement for the performance evaluation of the models, but stability which indicates the robustness of the performance to model parameter variation also becomes essential. A stable model is in practice of benefit to the medical specialists who may have little expertise in model tuning. The main purpose of this work is to address the importance of the stability of a model and to suggest one of such models.MethodsA comparative study of three prominent machine learning models was carried out for the prognosis of breast-cancer survivability: support vector machines, artificial neural networks, and semi-supervised learning models.MaterialThe surveillance, epidemiology, and end results database for breast cancer was used, which is known as the most comprehensive source of information on cancer incidence in the United States.ResultsThe best performance was obtained from the semi-supervised learning model. It showed good overall accuracy and stability under model parameter variation. The sharpening procedure enhanced the stability of the model via the noise-reduction.ConclusionWe suggest that semi-supervised learning model is a good candidate that medical professionals readily employ without consuming the time and effort for parameter searching for a specific model. The ease of use and faster time to results of the predictive model will eventually lead to the accurate and less-invasive prognosis for breast cancer patients.