Deep Feature Synthesis for Accurate Breast Cancer Prediction

With 2.3 million women diagnosed with breast cancer and 685.000 deaths globally in 2020, breast cancer is the most prevalent form of cancer worldwide. However, 7.8 million women are alive after being diagnosed with breast cancer in the past five years, indicating that breast cancer is survivable if diagnosed early. One in eight women will have invasive breast cancer during her lifetime. Therefore, it is vital that regular screenings are conducted for early detection and, thus, survival. Previous studies have been completed on the Breast Cancer Coimbra Data Set from the UCI repository to identify biomarkers with reliable confidence intervals for further investigation with less emphasis on a perfect model due to lack of enough data. Here, Deep Feature Synthesis and Conditional Generative Adversarial Networks (CTGANs) are implemented for data generation and to increase the performance of the models in terms of specificity, accuracy, and sensitivity. Logistic Regression outperforms the other classifiers with 100% specificity, accuracy, and sensitivity when only Deep Feature Synthesis is used. Performance decrease of models, when introduced to data CTGANs-generated data, indicates that CTGANs do not generate similar enough data. However, the large range of variability in the model's performance shows that BMI, Resistin, Age, and Glucose are not trustable biomarkers for early-stage breast cancer.