Machine learning-driven identification and immunohistochemical validation of an integrated immune-inflammatory phenotype for disease-free survival stratification in breast cancer.

Background: Recurrence risk in breast cancer remains heterogeneous, and conventional clinicopathological variables may not fully capture the contribution of immune-related factors. We compared multiple survival modeling strategies and evaluated whether an integrated immune-inflammatory phenotype could improve disease-free survival (DFS) stratification. Methods: This retrospective single-center study included 503 patients with surgically treated breast cancer between January 2020 and December 2025. Stromal tumor-infiltrating lymphocytes (TILs) were assessed from pathological sections, and the systemic immune-inflammation index (SII) was calculated from pre-treatment blood counts. An integrated immune phenotype was defined as favorable (high TILs/low SII), poor (low TILs/high SII), or intermediate (all remaining combinations). A base clinical Cox model, an immune-extended Cox model, LASSO-Cox, CoxBoost, and random survival forest (RSF) were compared using C-index, time-dependent area under the curve (AUC), integrated Brier score (IBS), and decision curve analysis. Conventional survival analyses, restricted cubic spline analysis, and immunohistochemical validation with CD8 and CD163 staining were also performed. Results: During follow-up, 107 patients (21.3%) experienced a DFS event. RSF achieved the best overall performance, with time-dependent AUCs of 0.867, 0.880, 0.879, 0.893, and 0.911 at 12, 24, 36, 48, and 60 months, respectively, and the lowest IBS (0.100). In the RSF model, pathological N stage was the most important predictor, followed by SII, integrated immune phenotype, Ki-67, and lymphovascular invasion. Kaplan-Meier analysis showed no significant DFS difference according to TIL category alone, whereas high SII and the poor integrated immune phenotype were associated with significantly worse DFS. In the final multivariable Cox model, the poor phenotype remained independently associated with worse DFS compared with the favorable phenotype (hazard ratio 2.53, 95% confidence interval 1.39-4.60; Conclusion: RSF provided the best prognostic performance in this cohort. SII and the integrated immune phenotype emerged as clinically relevant predictors, and the integrated phenotype showed tissue-level biological support. Combining machine learning-based survival modeling with pragmatic immune-inflammatory markers may improve recurrence risk stratification in breast cancer.