Integrating multivariable logistic regression, ensemble machine learning, and SHAP explainability for precision breast cancer mortality risk stratification and prevention strategy

The identification of predictors and precise mortality risk stratification help optimize targeted interventions in breast cancer screening and management. The combination of explainable machine learning and parametric statistical modeling lends clinical credibility, while the model's non-linear predictive power is also strong. This research is aimed at the development, validation, and creation of an explanatory framework for an integrated predictive framework using Multivariable Logistic Regression, Random Forests, and Gradient Boosting (based on XGBoost), and the integration of SHAP analysis of feature importance, as well as ROC-AUC analysis for the evaluation of breast cancer mortality predictors and treatment-related clinical risk factors, and ensuring treatment of breast cancer patients. A retrospective cohort study of 4,024 patients with breast cancer (3,408 still living [84.7%], 616 deceased [15.3%]) was conducted systematically. Baseline characteristics were compared using independent t-tests and the chi-squared test (χ2). Training of the models was performed on 80% of the data, and performance on the holdout set was evaluated (n = 805; 682 living, 123 deceased). The discrimination ability is quantified using the values of the confusion matrix (TP = 55, FP = 18, TN = 664, FN = 68), precision, sensitivity, and specificity values obtained over 1,000 bootstrap resampling iterations to obtain 95% confidence intervals (CIs). Also, Adjusted Odds Ratios (ORs) and SHAP feature weights were obtained to interpret the variables that most influence the death rate. The Multivariable Logistic Regression model provided the highest AUC among all tested machine learning algorithms with an ROC-AUC value of 0.853 (95% CI 0.810-0.892) similar to Gradient Boosting but better than Random Forests, which showed a lower performance (ROC-AUC=0.828, 95%CI 0.783-0.871). All models have shown very high values for Specificity (97.4%-97.7%) as well as Accuracy (89.3%). From this multivariate logistic analysis, our study found out that T4 advanced stages of tumors were the most important prognostic factor for patients with poor outcome, having an odds ratio of 4.095(95%CI 2.073–8.092), P<0.0001. High histologic grade was also significantly predictive of poor outcome (OR=2.463; 95%CI 1.645-3.688). However, progesterone receptor positivity predicted better prognosis as shown by reduced odds ratios (OR=0.596, 95%CI 0.443–0.801, p = 0.0006), which represents a 40.4% decreased risk of death among those positive to this marker. Similarly, estrogen positivity was significantly associated with 36.9% decreased risk of death (OR=0.631,SHAP showed that the most important variables used by our model for predicting survival include the duration of survival (66.1%) weight. A combination of multivariable logistic regression, and ensemble machine learning with SHAP, served as a framework to derive highly accurate, auditable, and interpretable breast cancer prognoses. Parametric effect sizes from the multivariable logistic regression directly informed and integrated various strategies and methods in oncology and cancer care.