Development of Artificial Intelligence systems as a prediction tool in ovarian cancer

Abstract

Ovarian cancer is the 5th most common cancer in females and the UK has one of the highest incident rates in Europe. In the UK only 36% of patients will live for at least 5 years after diagnosis. The number of prognostic markers, treatments and the sequences of treatments in ovarian cancer are rising. Therefore, it is getting more difficult for the human brain to perform clinical decision making. There is a need for an expert computer system (e.g. Artificial Intelligence (AI)), which is capable of investigating the possible outcomes for each marker, treatment and sequence of treatment. Such expert systems may provide a tool which could help clinicians to analyse and predict outcome using different treatment pathways. Whilst prediction of overall survival of a patient is difficult there may be some benefits, as this not only is useful information for the patient but may also determine treatment modality. In this project a dataset was constructed of 352 patients who had been treated at a single centre. Clinical data were extracted from the health records. Expert systems were then investigated to determine the optimum model to predict overall survival of a patient. The five year survival period (a standard survival outcome measure in cancer research) was investigated; in addition, the system was developed with the flexibility to predict patient survival rates for many other categories. Comparisons with currently used prognostic models in ovarian cancer demonstrated a significant improvement in performance for the AI model (Area under the Curve (AUC) of 0.72 for AI and AUC of 0.62 for the statistical model). Using various methods, the most important variables in this prediction were identified as: FIGO stage, outcome of the surgery and CA125. This research investigated the effects of increasing the number of cases in prediction models. Results indicated that by increasing the number of cases, the prediction performance improved. Categorization of continuous data did not improve the prediction performance. The project next investigated the possibility of predicting surgical outcomes in ovarian cancer using AI, based on the variables that are available for clinicians prior to the surgery. Such a tool could have direct clinical relevance. Diverse models that can predict the outcome of the surgery were investigated and developed. The developed AI models were also compared against the standard statistical prediction model, which demonstrated that the AI model outperformed the statistical prediction model: the prediction of all outcomes (complete or optimal or suboptimal) (AUC of AI: 0.71 and AUC of statistical model: 0.51), the prediction of complete or optimal cytoreduction versus suboptimal cytoreduction (AUC of AI: 0.73 and AUC of statistical model: 0.50) and finally the prediction of complete cytoreduction versus optimal or suboptimal cytoreduction (AUC of AI: 0.79 and AUC of statistical model: 0.47). The most important variables for this prediction were identified as: FIGO stage, tumour grade and histology. The application of transcriptomic analysis to cancer research raises the question of which genes are significantly involved in a particular cancer and which genes can accurately predict survival outcomes in a given cancer. Therefore, AI techniques were employed to identify the most important genes for the prediction of Homologous Recombination (HR), an important DNA repair pathway in ovarian cancer, identifying LIG1 and POLD3 as novel prognostic biomarkers. Finally, AI models were used to predict the HR status for any given patient (AUC: 0.87). This project has demonstrated that AI may have an important role in ovarian cancer. AI systems may provide tools to help clinicians and research in ovarian cancer and may allow more informed decisions resulting in better management of this cancer.