Machine Learning and Personalized Breast Cancer Risk Prediction

In the past decades the incidence of breast cancer has shown an increasing trend worldwide, while survival has improved through screening, especially if tumors are diagnosed at early stages, and through advances in therapeutic approaches. Early detection is currently the best option to reduce cancer morbidity and mortality. Although many risk factors have been established for breast cancer, e.g., age, family history, genetic predisposition, hormone and reproductive factors, and history of benign breast disease, few are applicable for primary prevention. In most western countries breast cancer screening programs target women over 50 years old and age is considered the sole risk factor for entering a population-based screening program. Many societies and groups propose that a risk-stratified screening strategy could be more effective, less morbid and more cost-effective. Breast cancer risk prediction models use established clinical and epidemiological factors to provide a risk estimate for individual woman. Clinicians can use these models to facilitate stratification of preventive interventions and personalized clinical management, including risk stratified screening at a younger age, chemoprevention, lifestyle change interventions, and follow-up care.
As an essential tool in precision medicine, several breast cancer risk prediction models are developed in past decades and some have been incorporated in clinical guidelines to support clinical decision making. The biggest limitation of these models is their low discriminatory accuracy (Area Under the Receiver Operating Characteristics curve around 0.65). This is slightly better than a coin toss and limits utility in clinical practice, especially at the individual patient level. These classical model-based prediction methods always rely on implicit assumptions that each risk factor relates to breast cancer in a linear way. These assumptions oversimplify complex relationships and non-linear interactions among multiple risk factors. Although these models have been updated and extended for decades, there is a very limited improvement in accuracy. Machine learning (ML) offers an alternative approach which can address current limitations and has the potentials to improve model performance. However, very few studies applied ML for personalized breast cancer risk prediction. The comparison of predictive accuracy and reliability for breast cancer lifetime risk prediction between ML and models commonly used in clinical practice has never been performed. Moreover, no ML-based model has been carried forward to explore its clinical utility, e.g. impact on screening practices.
This thesis addresses the above-mentioned limitations and gaps in knowledge. The overall aim was to develop a breast cancer risk prediction model based on ML techniques, to compare its performance with classic models commonly used in clinical practice, and to explore the clinical impact of ML-based models under the current screening settings and guideline-based recommendations.
The most important findings were the superior performance in the predictive accuracy of ML-based models to commonly used models when using the same risk factors from the US and Switzerland retrospective datasets. Bringing this advance of more accurate ML prediction into screening settings can result in about one in three women being classified into a different risk group. Women younger than 50 years old would be most influenced because clinical decision making for their initiation of screening would be changed.