Improving the prediction of surgical outcome at secondary cytoreduction in patients with ovarian cancer: Results from retrospective part of HELP-ER study NOGGO TR2/ENGOT OV47-TR.

Abstract

5558 Background: Complete resection at secondary cytoreductive surgery is associated with prolonged progression free and overall survival for patients with relapsed ovarian cancer. Secondary cytoreductive surgery has no impact on survival rates, if macroscopically tumor clearance cannot be achieved. Therefore, in order to avoid unnecessary perioperative morbidity and mortality, selection of patients who will undergo secondary tumor debulking is crucial. This study aims to improve upon the contemporary Arbeitsgemeinschaft Gynäkologische Onkologie (AGO) score by including additional clinical variables like circulating HE4 and CA125 levels to predict surgical outcome at secondary cytoreduction. Methods: A total of 90 patients underwent secondary cytoreductive surgery and were retrospectively assigned a positive AGO score. Of those patients, 62 (68.9%) achieved optimal surgical outcome at secondary debulking with 28 (31.1%) patients retaining residual tumor mass ( > 0mm). Utilizing clinical variables including circulating HE4 and CA125 levels, we implemented a machine learning workflow to predict suboptimal surgical outcome in patients despite a positive AGO score. Results: We elucidated significantly lower levels of circulating HE4 (p = 0.0038) in patients with optimal surgical outcome compared to patients that retain macroscopic residual tumor at secondary cytoreductive surgery. Moreover, machine learning algorithms trained on clinical variables (e.g. serum HE4 level, serum CA125 level, age, Risk of Ovarian Malignancy Algorithmus (ROMA) score and occurrence of peritoneal carcinomatosis) achieved a mean area under the curve (AUC) of 78.4% based on 100 consecutive executions with randomized training and test sets. Conclusions: The application of machine learning allows to further improve the prediction of patients with high likelihood of achieving optimal surgical outcome at secondary cytoreduction. In turn, it might identify patients that would benefit from amplified treatment efforts. However, machine learning relies on large amounts of data to account for biological and clinical variation and produce predictions of sufficient/adequate quality. Given this limitation, we would validate this data within the prospective multicentric cohort of patients collected within NOGGO/ENGOT HELP_ER Trial.