A comparative analysis of parametric survival models and machine learning methods in breast cancer prognosis

Abstract

Accurate prediction of breast cancer survival is critical for optimizing treatment strategies and improving clinical outcomes. This study evaluated a combination of parametric statistical models and machine learning algorithms to identify the most influential prognostic factors affecting the survival of patients. Two commonly used parametric models, log-gaussian regression and logistic regression, were applied to assess the relationship between survival and a set of clinical variables, including age at diagnosis, tumor grade, primary tumor site, marital status, American Joint Committee on Cancer (AJCC) stage, race, and receipt of radiation therapy or chemotherapy. Machine learning methods, such as neural networks, support vector machines (SVMs), random forests, gradient boosting machines (GBMs), and logistic regression classifiers, were employed to compare the predictive performance. Among these, the neural network model exhibited the highest predictive accuracy. The random forest model achieved the best balance between model fit and complexity, as indicated by its lowest akaike information criterion and bayesian information criterion values. Across all models, five variables consistently emerged as significant predictors of survival: age, tumor grade, ajcc stage, marital status, and radiation therapy use. These findings highlight the importance of combining traditional survival analysis techniques with machine learning approaches to enhance predictive accuracy and support evidence-based personalized treatment planning in breast cancer care.

Keywords: Accuracy; Hazard ratio; Probability density function; Survival probability.

Fig. 1

Flow Diagram of Data Analysis and Methodology.

Fig. 2

(a) Survival Probability, (b) Cumulative Survival, and (c) Hazard Function Curves for Parametric Survival Distributions.

Fig. 3

ROC Curve Illustrating Performance of (a) GBM, (b) LR, (c) Neural Network, (d) RF, and (e) SVM.