LOCD: a novel LOWESS based computational methodology for early detection of ovarian cancer with biomarker screening

Ovarian cancer is the fifth most common cause of cancer-related deaths among women and the survival rate for ovarian cancer varies significantly based on the stage at which the cancer is diagnosed. Therefore, early detection of ovarian cancer is crucial and can substantially improve the outcome in these women. Screening using protein biomarkers for the average-risk population has been trialed in the UK, but their effectiveness remains a major challenge. Nowadays, computational methods play an increasingly important role in early detection given their low cost, efficiency and ability to complement expert judgment. However, the screening data for ovarian cancer have several characteristics that make computational detection challenging. To this end, we developed a new methodology LOCD (LOWESS-based Ovarian Cancer Detection) that enables efficient and robust prediction of the outcome based on the screening history. We demonstrate the benefit of incorporating the longitudinal trajectory as well as the superiority of our method over the state-of-the-art using the gold-standard biomarker cancer antigen 125 (CA125) from a well-studied UKCTOCS dataset. Particularly, we adopted a repeated half-half-splitting strategy to conduct the internal validation to enhance the pertinence of evaluation. The performance was comprehensively assessed with the ROC AUC and the sensitivity scores and reported in the form of the median and its confidence interval. Furthermore, the extension to the multi-marker panel (CA125, HE4 and Glycodelin) is also laid out and compared with the CA125-algorithm alone.