Machine Learning-based Predictive Model for Mortality in Female Breast Cancer Patients Considering Lifestyle Factors

Objective: Current breast cancer mortality prediction models primarily rely on clinical and genetic data, which are costly and less accessible. Emerging evidence highlights the significance of lifestyle factors in postdiagnosis mortality among breast cancer patients. This study aimed to enhance predictive models by integrating lifestyle variables, creating an economical, practical, and interpretable prognostic tool. Methods: In this retrospective study, we utilized a ten-year follow-up dataset of female breast cancer patients from a major Chinese hospital and included 1,390 female breast cancer patients with a 7% (96) mortality rate. We employed six machine learning algorithms (ridge regression, k-nearest neighbors, neural network, random forest, support vector machine, and extremegradient boosting) to construct a prognostic model for breast cancer. The 10-fold cross-validation results indicated that the random forest model exhibited the best performance (average area under the curve (AUC)= 0.918; 1-yr AUC = 0.914, 2-yr AUC = 0.867, 3-yr AUC = 0.883). These models also performed strongly in an independent external dataset (external validation average AUC = 0.782; 1-yr AUC = 0.809; 2-yr AUC = 0.785; 3-yr AUC = 0.893). We developed a user-friendly web tool using Shiny-Web for easy model access. Cox regression identified postsurgery sexual activity, totally implantable venous access port use, and prosthetic breast wear as independent protective factors, while considering pain to be the greatest difficulty postoperatively was associated with a worse prognosis. Conclusion: Our postoperative mortality prediction model is cost-effective and accurate, aiding healthcare professionals in treatment decisions and promoting healthier lifestyles for breast cancer patients.