Abstract / Summary
Abstract Quantum-kernel models are increasingly proposed for clinical prediction, yet two questions block their adoption: can their outputs be explained at the level of an individual patient, and do they actually outperform classical models? We address both for breast-cancer recurrence. Our primary contribution is methodological: we show that a quantum-kernel survival model admits exact, axiom-satisfying Shapley explanations. Because a clinically grounded encoding needs only three qubits, the Shapley values can be computed by direct coalition enumeration rather than sampling; we prove the resulting attributions satisfy the efficiency axiom and verify this numerically to machine precision, and we show the sampled KernelSHAP estimator agrees with the exact values at a correlation of 0.97. We then test predictive value honestly across three independent cohorts spanning 194 to 1904 patients (WPBC, GBSG2, METABRIC) plus a controlled synthetic study. After correcting a kernel-concentration failure through embedding-bandwidth tuning, the quantum kernel is statistically indistinguishable from a classical radial-basis kernel on matched features in every real cohort, and it significantly outperforms a linear model only when the data-generating structure matches its inductive bias. Under a depolarizing-and-readout noise model the survival pipeline is robust, with concordance changing by less than 0.005. We conclude that the practical case for quantum-kernel survival models rests on exact explainability rather than accuracy, and we provide a reproducible template for evaluating such claims.