PREDICTING SURGICAL OUTCOME IN PATIENTS WITH ADVANCED OVARIAN CANCER

Abstract

Introduction: Epithelial ovarian cancer affects more than 7,000 women each year in the UK, and approximately 70% of patients present with advanced disease. Treatment involves one of the following options: initial primary debulking surgery (PDS) followed by chemotherapy; initial chemotherapy followed by interval debulking surgery (IDS); or palliative management. Removing all visible disease at the time of surgery (complete cytoreduction) is the most important prognostic marker in these patients, and it is important that the treatment option that is most likely to result in complete cytoreduction is chosen. Complete cytoreduction at the time of PDS holds a survival advantage over IDS. Failing to achieve complete cytoreduction equates to reduced overall survival, increased morbidity and delayed chemotherapy start. Under the current decision-making process 9 - 67% of patients suffer residual disease, which highlights a significant area for improvement. Many clinical, biochemical, genomic and radiological predictors have been linked with surgical outcome. Despite this, there is no accepted tool or guideline to aid clinicians in this decision-making process. This thesis aims to review currently published models predicting surgical outcome, externally validate a pre-existing model, explore new predictors and create a new prognostic model combining all available predictors. Methods: A systematic review of all published multimodal prognostic models predicting outcome of PDS in patients with stage II-IV epithelial ovarian cancer was performed. Data extraction was performed using the checklist for critical appraisal method (CHARMS) and risk of bias was assessed using the prediction model study assessment tool (PROBAST). The external validation of a three-protein signature was performed via immunohistochemistry (IHC) on a validation cohort from the ICON5 trial. The association between homologous recombination (HR) and surgical outcome was assessed in two separate cohorts. The cancer genome atlas (TCGA) cohort had HR status assessed using an established gene panel, and the Manchester database cohort via a functional assay. Both logistic regression and random forest models were developed combining multiple predictors including operating surgeon to predict surgical outcome at the time of PDS. Results: The systematic review included 26 publications describing 27 prognostic models. Predictors included clinical, biochemical, genomic and radiological features. All but one model was developed by logistic regression. Validated performance measured by AUC ranged between 0.50 and 0.89, with low levels of external validation. The majority of models showed high risk of risk of bias. The three-protein signature was validated via IHC on a cohort of 238 HGSOC tumour samples. Staining intensity scores from each protein were combined to create a combined prognostic model. Validation failed, with AUC dropping from 0.866 in the original cohort to 0.593 in the validation cohort. Two patient cohorts were used to assess association between HR status and surgical outcome. The TCGA cohort (n=258) assessed via a 14 gene panel demonstrated association between HR status and surgical outcome (p=0.033). The Manchester cohort (n=38) assessed via a functional assay did not show any association (p=0.5205). Finally, the developed prognostic prediction models developed on a cohort of stage III-IV HGSOC patients (n=100) incorporated 18 predictors. When internally validated, they performed with AUC values of 0.688 and 0.734 for logistic regression and Random Forest models, respectively. Conclusions: Models incorporating single modalities rarely show accurate prediction upon external validation and there are currently no prognostic models validated successfully enough for use in clinical practice. Models combining multiple predictors including surgeon heterogeneity show the most promise, and further validation would be the next step to progress these models with a view to app

Date of Award	31 Dec 2022
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Emma Crosbie (Supervisor) & Richard Edmondson (Supervisor)