An interpretable machine-learning signature predicts early triple-negative breast cancer recurrence but adds no prognostic value beyond clinicopathological factors in external validation

Triple-negative breast cancer (TNBC) is an aggressive subtype that lacks targeted therapies and carries a high risk of early distant metastasis, yet interpretable tools for predicting early recurrence remain scarce. We develop an end-to-end, leakage-safe machine-learning framework that predicts 3-year distant recurrence from the gene-expression profiles of 101 TNBC patients in GSE58812, and evaluate it in an independent, clinically annotated cohort—the triple-negative subset of METABRIC ( \(n=320\) ). Supervised feature selection is embedded within every fold of repeated stratified cross-validation, so that seven classifiers are compared on honest out-of-fold predictions; an \(\ell _2\) -regularised logistic regression on a fifty-gene panel attained a mean per-fold cross-validated ROC-AUC of 0.689 (standard deviation 0.137) and a repeat-averaged out-of-fold ROC-AUC of 0.738 (95% CI 0.626–0.837), with good calibration (Brier score 0.171). The two are not interchangeable: 0.689 estimates the discrimination of one fitted model and is the primary, more conservative figure, whereas 0.738 describes the repeat-averaged ensemble and is the value to which the calibration, decision-curve and internal survival analyses correspond. Both retain the optimism of having chosen the reported learner as the best of seven on the same resampling: nested selection of the learner gives 0.648. The panel size was fixed a priori; nested tuning did not improve honest discrimination (0.648), and performance was flat for panels of 30–100 genes. Coefficients and SHAP attributions agreed almost perfectly, which for a standardised linear model holds by construction and is reported as an internal consistency check; a model-agnostic explainer applied to three non-linear learners gave much weaker cross-model agreement (Spearman \(\rho \) 0.54–0.94, median 0.62). Functional annotation showed natural-killer-cell-mediated cytotoxicity and antigen presentation, with proliferation up (E2F targets, NES \(+2.23\) ) and interferon- \(\gamma \) response down (NES \(-2.77\) ) in early-recurring tumours. Because GSE58812 records no clinicopathological variable beyond age, METABRIC was used to test the signature against, and on top of, the routine factors that govern treatment decisions. A clinicopathological Cox model discriminated relapse-free survival well (concordance index 0.673), driven chiefly by nodal burden, whereas only 20 of the 50 genes are quantified on the Illumina array, so what could be scored externally is the cross-platform-transferable component. It carried only a weak univariable signal (hazard ratio 1.23 per standard deviation; 3-year AUC 0.55) and did not improve discrimination over clinicopathological factors (combined concordance index 0.652; likelihood-ratio \(p=0.13\) ). Re-deriving the panel within the gene universe common to both platforms yielded a fully scoreable signature that separated relapse-free survival externally (hazard ratio 1.30 per standard deviation, \(p=0.004\) ) but still added nothing ( \(p=0.09\) ), and ComBat changed the external AUC by less than 0.01 while a within-METABRIC refit reached only 0.56–0.57. The endpoint mismatch does not account for this: the signature performs no better against a metastasis-dominated cause-specific endpoint, and even if three first relapses in ten were local the corrected concordance index would rise only to 0.58. A transcriptomic signature can therefore be developed and interpreted reliably, but clinicopathological variables remain the dominant determinants of recurrence risk. The signature is best regarded as a hypothesis-generating molecular adjunct rather than a clinically ready tool.