Scientists use machine learning to uncover new predictor of postmenopausal breast cancer

Breast cancer is one of the most common types of cancer affecting women worldwide. Multiple predictors of this disease have been identified, including genetic genetic factors, reproductive factors, and lifestyle.

Previous studies have highlighted differences in the etiology of premenopausal and postmenopausal breast cancer. Recently, scientists combine different approaches to make accurate predictions. breast cancer in women.

study: Combining machine learning and Cox models to identify predictors of postmenopausal breast cancer development in the UK Biobank. Image credit: aslysun / Shutterstock.com

Background

Machine learning (ML) techniques can analyze large datasets on predictors and handle complex nonlinear relationships. Previous studies have used ML to predict breast cancer risk, but not to identify predictors.

Consisting of a broad and detailed cohort, the UK Biobank (UKB) offers an opportunity to adopt a hypothesis-independent approach to identifying novel predictors of breast cancer. The recently developed polygenic risk score (PRS) uses genome-wide association studies (GWAS) to predict the impact of hundreds or thousands of genetic variants associated with a particular disease or trait.

PRS can be used to identify people at high risk of disease and target them for early statin prescription. In particular, PRS improved the accuracy of existing coronary artery disease risk predictors such as the Framingham risk score.

Breast cancer PRS has been combined with risk prediction models such as the Tyrer-Cuzick model and the Breast and Ovarian Incidence and Carrier Estimation Algorithm (BOADICEA). In breast cancer, interactions between PRS and phenotypic traits such as gene-environment interactions have been analyzed, but conflicting results have been reported.

About research

Recent scientific report This study utilized machine learning (ML) techniques for feature selection, followed by Cox models for risk prediction. The main purpose of this study was to demonstrate effective application of ML methods for feature selection in support of classical statistical methods.

We investigated potential interactions between phenotypic features and PRS using SHapley Additive exPlanation (SHAP) feature dependence plots. The study used UKB data and included more than 500,000 participants from England, Wales and Scotland. Baseline data were collected through oral interviews with trained nurses, questionnaires, biological samples, and physical examinations.

Postmenopausal women with ages between 40 and 69 years at baseline were recruited because of the heterogeneity of etiology due to menopausal status described above. The incidence of breast cancer was identified using the International Classification of Diseases code, and PRS313 and PRS120k were considered potential genetic signatures.

research result

The study included a total of 104,313 participants, of whom 4,010 developed breast cancer during the 11.9-year follow-up period. Combining ML with conventional cancer epidemiological statistical approaches identified several known and unknown risk factors for postmenopausal cancer incidence.

Known risk factors identified include age at menopause, testosterone, and age. Five new predictors were also identified, including blood chemistry, blood counts, and urinary biomarkers.

Newly identified predictors were strongly associated with the incidence of postmenopausal breast cancer. Further research is needed to understand whether these are potentially modifiable risk factors for breast cancer in the future.

The XGBoost model chose detailed body composition measurements instead of body mass index (BMI). This suggests that accurate body composition measurements are an important predictor of breast cancer. Basal metabolic rate was also found to be an important predictor of breast cancer, which contradicts previous studies that found no association between basal metabolic rate and breast cancer.

Plasma urea, a blood biomarker associated with renal function, was also associated with breast cancer. This is the first time an association between urinary plasma phosphate, sodium, or creatinine and breast cancer has been reported.

Two polygenic risk scores were ranked as the strongest risk factors by the agnostic ML model. Cox regression proved PRS to be an important predictor of postmenopausal breast cancer.

Conclusion

The current study identified five new, statistically significant correlations with postmenopausal breast cancer, including urinary biomarkers, blood counts, and blood biochemistry. Adding these five new features to the baseline Cox model maintained the discriminative performance. In addition, two prespecified PRSs were found to be the most important features by SHAP value.

These findings motivate further research on the use of more accurate anthropometric measurements to improve breast cancer prediction. External validation of results is an important next step prior to introduction into clinical practice.