A three-step machine learning strategy for reproducible biomarker screening in ovarian cancer

Biomarkers have a significant impact in health and medicine, including cancer research, screening, and diagnosis. Traditional statistical tests, however, often fail to provide consistent and reproducible quantitative measures of biomarker effects which hinders further validation studies. In this paper, we propose a three-step machine learning strategy for generating quantitative evaluation of candidate biomarkers and provide partial explanation to the results using Shapley value analysis. For transparency and reproducibility, all experiments are performed on two widely accessible datasets on biomarkers of ovarian cancer: a timely 2025 release and an established benchmark. Our findings demonstrate consistent results that align qualitatively with existing domain knowledge with the advantage of having robust quantitative evaluations. Thus, the pipeline presents a potentially valuable approach to addressing the challenges in discovering, evaluating, and validating biomarkers, and therefore, the paper concludes by proposing opportunities for further work.