Treffer: EXPERIMENTAL STUDY OF THE IMPORTANCE OF DATA FOR MACHINE LEARNING-BASED BREAST CANCER OUTCOME PREDICTION

EXPERIMENTAL STUDY OF THE IMPORTANCE OF DATA FOR MACHINE LEARNING-BASED BREAST CANCER OUTCOME PREDICTIONWid Yamani, Ph.D. George Mason University, 2024 Dissertation Director: Dr. Janusz Wojtusiak Researchers have used various large-scale datasets to develop and validate predictive models in breast cancer outcome prediction. However, a notable gap exists due to the lack of a systematic comparison among these datasets regarding predictive performance, feature availability, and suitability for different analytical objectives. While each dataset has unique strengths and limitations, no comprehensive studies evaluate how these differences impact model performance, particularly across diverse timeframes, survival, and recurrence outcomes. This gap limits researchers in making informed choices about the most appropriate dataset for specific research questions.Effective modeling and prediction of breast cancer outcomes (such as cancer survival and recurrence) rely on the dataset's quality, the pre-processing techniques used to clean and transform data, and the choice of predictive models. Therefore, selecting a suitable dataset and identifying relevant variables are as crucial as the choice of the model itself. This thesis addresses this gap by systematically comparing five prominent datasets for predicting breast cancer outcomes. This dissertation compares five datasets—SEER Research 8, SEER Research 17, SEER Research Plus, SEER-Medicare, and Medicare Claims data—focusing on breast cancer survival and recurrence. It evaluates the predictive performance of each dataset using supervised machine learning methods, including logistic regression, random forest, and gradient boosting. The models were tested on metrics such as AUC, accuracy, recall, and precision, with gradient boosting delivering the most accurate results. The findings indicate that SEER-Medicare, which integrates cancer registry data with three years of retrospective claims, outperformed the other datasets, achieving AUCs of 0.891 for 5-year survival and ...