A decision support system based on an ensemble of random forests for improving the management of women with abnormal findings at cervical cancer screening.

In most cases, cervical cancer (CxCa) develops due to underestimated abnormalities in the Pap test. Today, there are ancillary molecular biology techniques available that provide important information related to CxCa and the Human Papillomavirus (HPV) natural history, including HPV DNA tests, HPV mRNA tests and immunocytochemistry techniques such as overexpression of p16. These techniques are either highly sensitive or highly specific, however not both at the same time, thus no perfect method is available today. In this paper we present a decision support system (DSS) based on an ensemble of Random Forests (RFs) for the intelligent combination of the results of classic and ancillary techniques that are available for CxCa detection, in order to exploit the benefits of each technique and produce more accurate results. The proposed system achieved both, high sensitivity (86.1%) and high specificity (93.3%), as well as high overall accuracy (91.8%), in detecting cervical intraepithelial neoplasia grade 2 or worse (CIN2+). The system's performance was better than any other single test involved in this study. Moreover, the proposed architecture of employing an ensemble of RFs proved to be better than the single classifier approach. The presented system can handle cases with missing tests and more importantly cases with inadequate cytological outcome, thus it can also produce accurate results in the case of stand-alone HPV-based screening, where Pap test is not applied. The proposed system may identify women at true risk of developing CxCa and guide personalised management and therapeutic interventions.