From diagnosis to prognosis: machine learning-based survival prediction in breast cancer patients

Prognostication is important in personalising oncological care for breast cancer patients. Although conventional approaches, such as Cox regression, may be used to predict breast cancer outcomes, they may have limitations in capturing the complex interactions among multiple variables. As such, machine learning (ML) approaches, such as extreme gradient boosting (XGBoost), have the potential to provide satisfactory predictions for survival outcomes. Nevertheless, the use of ML in the Malaysian population remains limited. Therefore, this study was conducted to predict the survival outcomes of breast cancer patients via XGBoost. A retrospective cohort study of women diagnosed with breast cancer at Hospital Kuala Lumpur between 2011 and 2015 was conducted. Data on demographic characteristics, pathological profiles, and treatment were analysed. Survival time was calculated from the date of breast cancer diagnosis to the date of all-cause mortality or the end of the study period. The data were split into training and test sets via an 80–20 stratified approach. XGBoost was used to develop both an XGBoost regression model that predicts the logarithm of survival time and an XGBoost Cox proportional hazards (PH) model using the survival: the cox objective. Model performance was evaluated via Harrell’s concordance index (C-index) with 95% confidence intervals (CIs) estimated via bootstrap resampling (1,000 iterations), and variable importance was based on gain metrics. After excluding 303 (15.3%) patients with missing data on cancer stage and tumour grade, the training dataset included 1,336 patients, and the test set included 334 patients. The XGBoost Cox model achieved strong predictive performance, with a test C-index value of 0.7216 (95% CI: 0.6742–0.7719), compared with 0.6734 (95% CI: 0.6190–0.7247) for the XGBoost regression model and 0.2616 (95% CI: 0.2165–0.3106) for the Cox PH baseline. Age at diagnosis (34.2%) and cancer stage (28.2%) were the most influential features. The XGBoost Cox survival model showed strong predictive performance in terms of survival in women diagnosed with breast cancer, with a test C-index of 0.7216 (95% CI: 0.6742–0.7719). Caution is warranted regarding the generalisability of the study’s findings, as internal validation was performed using a single institution test set. Age and stage were identified as the most influential predictive features. The findings from this study support the utility of ML applications in improving the accuracy of prognostic factor estimates and aiding risk stratification in clinical practice, especially in resource-limited settings.