Background: Ovarian cancer is the 6th most common cancer to affect UK women and has the worst prognosis of any gynaecological cancer. Most women are not diagnosed until the disease is advanced, which leads to poor outcomes. Earlier ovarian cancer diagnosis has the potential to improve these outcomes. Cancer antigen 125 (CA125) is recommended by the National Institute for Health and Care Excellence (NICE) as the first line test for ovarian cancer in symptomatic women presenting to primary care in England. However, the performance of CA125 in this setting is unknown. The overarching aim of this thesis was to determine the diagnostic performance of CA125 for the detection of ovarian cancer when used in primary care, and to develop and evaluate novel approaches to improve its performance and clinical utility. Key methods: I used routinely collected primary care and cancer registry data from 50,780 women who underwent CA125 testing in England between 1st May 2011 - 31st December 2014. First, I performed a diagnostic accuracy study, calculating the performance of CA125 within the cohort, at the national cut-off (≥35 U/ml), for the detection of ovarian cancer. Diagnostic accuracy metrics were also calculated for other types of cancer and all cancer types combined (secondary study outcomes). I used logistic regression to estimate the probability of ovarian cancer at specific CA125 levels (1-1000 U/ml) for women of different ages. CA125 levels equating to a 3% ovarian cancer probability (the "risk threshold" at which NICE advocates urgent specialist cancer investigation) were identified. Next, I examined the associations between CA125 test result and time from testing to diagnosis, tumour type and cancer stage, in those women with ovarian cancer. Finally, I developed and internally validated ovarian cancer diagnostic prediction models (of varying complexity) in a sub-group of women with a relevant symptom recorded prior to CA125 testing (n=29,962). To inform the development of these models, I conducted a systematic review of existing ovarian cancer detection tools. Key results: CA125 had a sensitivity of 77%, a specificity of 94% and a Positive Predictive Value (PPV) of 10% for ovarian cancer at the national cut-off (≥35 U/ml). The PPV for all cancers combined was 21% overall, and 33% in women ≥50 years of age. 20% of women ≥50 years with a raised CA125 level, but no ovarian cancer, had another type of cancer. A CA125 value of 53 U/ml equated to a 3% probability of ovarian cancer overall, but this varied markedly by age (40- year-old: 104 U/ml, 70-year-old: 32 U/ml). Women with a 'normal' CA125 (<35 U/ml) prior to ovarian cancer diagnosis took twice as long to be diagnosed as those with an 'abnormal' CA125, but more frequently had indolent tumour types and were more likely to be diagnosed at an early stage. An ovarian cancer prediction model, incorporating patient age and CA125 level, outperformed CA125 alone. This model showed excellent discrimination on internal validation (AUC: 0.94). Including symptoms, baseline risk factors and other routine blood tests did not improve model performance. Conclusions: My findings demonstrate that CA125 is a useful test for ovarian cancer detection in primary care. They also indicate that clinicians should consider other types of cancer in women with high CA125 levels, especially if ovarian cancer has been excluded, in order to prevent diagnostic delay. The models presented in this thesis will allow patients and clinicians to determine the estimated probability of ovarian cancer at any given CA125 level and age. This information could inform individual patient decisions on the need for further investigation. If incorporated into the diagnostic pathway, the models would enable patients to be referred on the basis of ovarian cancer risk rather than a generic CA125 cut-off.