Article Text

Download PDFPDF
233 Predicting survival in advanced ovarian cancer; strategies to overcome national heterogeneity and model the causal impact of treatment
  1. Kathryn Baxter1,
  2. Glen Martin2 and
  3. Richard Edmondson1,2
  1. 1St Mary’s, Manchester, UK
  2. 2University of Manchester, Manchester, UK


Introduction/Background The geographical heterogeneity seen in treatment patterns for patients with advanced ovarian cancer is profound, long standing, worrying, and impacts upon survival. A tool that could demonstrate the impact of a patient’s treatment on their predicted survival is needed to counsel patients about their treatment options and reduce treatment variance.

We used detailed clinical datasets to develop a model for predicting treatment dependent survival in ovarian cancer patients.

Methodology Data were collected using a data dictionary for all cases of ovarian cancer presenting to six cancer centres in England between 1/1/2018 and 31/12/2019.

A Cox Proportional Hazard model was built using internal-external cross validation to estimate data heterogeneity between centres. Variables were assessed for non-linear relationships. Backwards selection was used to optimise fit. The hazard ratios for surgical treatment from the model were compared to existing RCT results, to investigate our ability to estimate causal effects within the observational data. Prediction accuracy of all models was assessed by calibration and discrimination (Harrell’s C statistic).

Results 991 patient records were included in survival analysis. The concordance of the models developed ranged from 0.64–0.78 across centres, with meta-analysis of the final model producing a Harrell’s C statistic of 0.73 (CI 0.70–0.75). The observed to expected ratio for three-year survival was 1.07 (CI 0.47–2.47). Some model recalibration was needed for each centre to achieve good calibration performance.

The hazard ratios for surgical treatment were similar to, and fell within, the confidence intervals of RCT results, confirming the model’s ability to predict survival under differing treatment conditions. This causal model retained discriminative power (C-stat 0.73, CI 0.65–0.76).

Conclusion Our new model has excellent predictive power, and this methodology ensures this will translate to new datasets. If validated, this model could be used to predict the treatment dependant survival in ovarian cancer.

Disclosures None.

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.