Interpretable machine learning for prognostic prediction in young women with triple-negative invasive ductal breast cancer: a Boruta-SHAP integrated approach.

Background: Young women with triple-negative invasive ductal breast cancer (TN-IDC) exhibit high invasiveness and poor prognosis, rendering accurate prognostic prediction crucial for individualized diagnosis and treatment. This study aimed to identify core prognostic factors and establish an optimal binary predictive model for cancer-specific survival (CSS) through an interpretable machine learning (ML) framework comprising "feature selection-model construction-result interpretation". Methods: Clinical data of 2,311 young female TN-IDC patients aged 20-40 years were extracted from the Surveillance, Epidemiology, and End Results (SEER) database, including 560 deceased cases and 1,751 surviving cases. The dataset covered multi-dimensional indicators such as demographics, tumor characteristics, treatment regimens, and metastatic status. The Boruta algorithm was used to screen key prognostic features for CSS, and 8 ML models were established based on the selected features. Model performance was comprehensively evaluated using 9 metrics (including accuracy, sensitivity, and specificity) as well as receiver operating characteristic (ROC) curves and calibration curves. Finally, the Shapley Additive Explanations (SHAP) algorithm was employed to interpret model logic and quantify feature contributions to CSS. Results: Boruta feature selection identified node (N) stage, tumor (T) stage, bone, lung, and liver metastases, and surgical approach as core prognostic factors for CSS. Among the 8 established ML models, the multilayer perceptron (MLP) model demonstrated the optimal overall performance, with an accuracy of 0.711 [95% confidence interval (CI): 0.683-0.739], Matthews Correlation Coefficient of 0.402 (95% CI: 0.372-0.423), balanced accuracy of 0.732 (95% CI: 0.704-0.759), area under the curve (AUC) of 0.782 (95% CI: 0.737-0.828). Additionally, its calibration curve showed the highest alignment with the ideal line. Several models achieved competitive performance, including logistic regression (AUC =0.736), xgboost (AUC =0.741), random forest (rf, AUC =0.726) and elastic net (enet, AUC =0.738), yet none surpassed MLP in overall predictive ability. AUC differences between MLP and logistic regression (P=0.17) or xgboost (P=0.22) were not statistically significant, but MLP showed consistently superior performance across multiple metrics, confirming its robustness for this task. SHAP analysis revealed that the N3 stage (N stage subclass) and bone metastasis had the most significant impacts on predictive outcomes, with higher N stages and positive bone metastasis status significantly increasing the risk of adverse prognosis. Conclusions: Through the Boruta-SHAP interpretable ML framework, this study clarified the core prognostic characteristics of TN-IDC in young women. The constructed MLP model shows favorable prognostic performance, providing supportive evidence and a supplementary practical tool for clinical risk stratification, individualized treatment decision-making, and follow-up management.