Ovarian cancer recurrence prediction: comparing confirmatory to real-world predictors with machine learning.

Background: Ovarian cancer is one of the deadliest cancers in women, frequently diagnosed at an advanced stage, with a 5-year survival rate of 17%-28% in advanced-stage (International Federation of Gynecology and Obstetrics IIB-IV) disease. Machine learning (ML) may provide a better tool for survival prognosis than traditional methods and could provide insight into predictive factors. This study focuses on advanced-stage ovarian cancer and contrasts expert-derived predictive factors with data-driven ones from the Netherlands Cancer Registry (NCR) to predict progression-free survival. Materials and methods: A Delphi questionnaire was conducted to identify 14 predictive factors which were included in the final analysis. ML models (regularized Cox regression, random survival forests, and XGBoost) were used to compare the Delphi expert-based set of variables with a real-world data (RWD) variable set derived from the NCR. A non-regularized Cox model was used as the benchmark. Results: While regularized Cox models with the RWD variable set outperformed the traditional Cox regression with the Delphi variables (c-index: 0.70 versus 0.64, respectively), XGBoost showed the best performance overall (c-index: 0.75). The most predictive factors for recurrence, not identified by Delphi, were surgery type and debulking results, post-operative chemotherapy administration, number of platinum cycles, and socioeconomic status. Conclusions: Our results highlight that ML algorithms have higher predictive power compared with the traditional Cox regression. Moreover, RWD from a cancer registry identified more predictive variables than a panel of experts. Overall, these results have important implications for artificial intelligence (AI)-assisted clinical prognosis and provide insight into the differences between AI-driven and expert-based decision making in survival prediction.