Development and evaluation of clinical prediction models for risk-stratified early detection, prevention and management of breast cancer

Thesis

Accurately estimating individual-level risks of breast cancer incidence and mortality could inform stratified approaches to screening, prevention, and management that help reduce deaths from breast cancer and are more cost-effective. Increasing interest in ‘machine learning’ techniques and the integration of phenotypic and genetic data present uncertainty around the best approaches to prognostication in these settings.

In a scoping review, evidence deficiencies were highlighted regarding risk-based breast screening, and parallels explored between this and risk-based breast cancer prevention and management. Using the QResearch primary care database and its linked datasets, clinical prediction models were developed using regression and machine learning methods to estimate individual women’s 10-year risks of incident breast cancer, 10-year risks of developing and then dying from breast cancer, and 10-year risks of breast cancer mortality after diagnosis. These were comparatively evaluated in an internal-external cross-validation framework to assess performance and transportability. Second, the incremental effects of integrating genome-wide polygenic risk scores to models for the first two outcomes were assessed using the UK Biobank.

This thesis sought to develop and evaluate models in accordance with modern practice, with an emphasis on robust, fair comparisons between different approaches. A key output was the first work to develop models that estimate the risk of breast cancer mortality in women without breast cancer at baseline, which could be important in the context of screening-associated overdiagnosis. Second, models were produced that may provide reliable risk stratification for any woman with breast cancer. Third, integrating genome- wide polygenic risk to ‘phenotypic’ information may increase the performance of breast cancer incidence and mortality models in women currently eligible for screening. The clinical impact and cost-effectiveness of strategies structured around these models needs further assessment.

Authors: Clift, A

• DOI: 10.5287/ora-nbv5a0rrq
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: risk prediction; breast cancer; machine learning
• Subjects: Breast cancer; Clinical prediction modelling