💻 computer science

Exact Shapley Explanations for Quantum-Kernel Survival Models: A Three-Cohort Breast-Cancer Recurrence Study

This study demonstrates that quantum-kernel survival models for breast-cancer recurrence can be exactly explained via Shapley values due to their low-qubit encoding, while revealing that their predictive performance is statistically comparable to classical radial-basis kernels and robust to noise, suggesting their primary practical value lies in interpretability rather than superior accuracy.

Original authors: Bhanwar Gupta, Sanjeev Rana

Published 2026-10-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bhanwar Gupta, Sanjeev Rana

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

When a surgeon removes a breast tumor, the most pressing question for the patient is not just whether the cancer is gone, but when it might return. For the doctor, the challenge is different: they need to know why a computer model predicts a specific risk for a specific person. In medicine, a prediction without an explanation is often useless; a clinician cannot trust a "high risk" warning if they cannot see which factors—tumor size, cell shape, or lymph node count—drove that conclusion. This is the realm of survival analysis, a statistical method designed to handle the uncertainty of time and the fact that some patients leave a study before their cancer returns. Recently, scientists have begun asking if quantum computers, which process information in ways fundamentally different from standard machines, could improve these predictions. The hope has been that these machines could spot complex patterns in patient data that traditional computers miss, offering better forecasts for when a disease might recur.

A new study by researchers at Maharishi Markandeshwar University in India tackles this promise with a rigorous, three-part investigation into breast-cancer recurrence. The team did not simply run a quantum model and hope for the best; they built a system to test two specific questions. First, can a quantum model explain its own reasoning to a doctor in a way that is mathematically sound and trustworthy? Second, does it actually predict recurrence better than the best classical computers we have today? The researchers focused on three distinct groups of patients, ranging from a small set of 194 individuals to a large group of nearly 2,000, to ensure their findings were not just a fluke of a single dataset.

The most significant finding of the work concerns the first question: explainability. The researchers demonstrated that a quantum model designed for survival analysis can produce exact, step-by-step explanations for its predictions. In the world of artificial intelligence, "Shapley values" are a standard method for breaking down a prediction to show how much each factor contributed to the final result. For many complex models, calculating these values requires guessing or sampling, which introduces a small amount of uncertainty. However, because the quantum model in this study used a simplified, three-dimensional representation of patient data, the researchers could calculate the exact contribution of every single factor without any guessing. They proved mathematically that these calculations were perfect and verified them against a standard sampling method, finding a near-perfect match. This means that for this specific type of quantum model, a doctor could receive a clear, undeniable list of why a patient was flagged as high-risk, a level of transparency that is often missing in advanced machine learning.

On the second question—whether the quantum model actually predicts better—the answer was more sobering. The researchers tested their quantum system against powerful classical models, including standard statistical methods and advanced tree-based algorithms, across all three patient groups. They found that once the quantum model was carefully tuned to avoid a common technical pitfall where the data becomes too uniform to learn from, it performed just as well as the best classical models, but not better. In every real-world dataset they examined, the quantum model's accuracy was statistically indistinguishable from a classical model using the same features. It did not outperform the established methods, nor did it perform worse. The study explicitly ruled out the idea that quantum computers offer a magical boost in accuracy for this specific medical task on current data.

The researchers did find a specific condition where the quantum approach showed an edge, but it was a narrow one. In a controlled experiment using synthetic data designed to mimic a very specific type of complex interaction between variables, the quantum model significantly outperformed a basic linear model. However, even in this ideal scenario, it did not beat a well-tuned classical model. This suggests that the quantum advantage is not a general superpower but a specialized tool that only works when the data structure perfectly matches the way the quantum computer is built to think. Furthermore, the study showed that the system was robust against the noise and errors typical of current quantum hardware, with the predictions changing by less than a tiny fraction even when the simulation included realistic hardware imperfections.

Ultimately, the study reframes the value of quantum computing in medicine. The practical case for using these models does not rest on the hope that they will suddenly predict cancer recurrence with superhuman accuracy. Instead, the value lies in the ability to provide exact, mathematically verified explanations for every prediction. For a field where trust and clarity are as important as the numbers themselves, the ability to say "this patient is at risk because of these three specific factors, and we know this with absolute certainty" is a tangible benefit. The researchers conclude that while quantum models may not yet be the superior predictor for breast cancer, they offer a reproducible, honest, and fully explainable template for how such models should be evaluated, shifting the focus from a search for impossible accuracy gains to the very real goal of trustworthy, transparent medicine.

Technical Summary: Exact Shapley Explanations for Quantum-Kernel Survival Models

1. Problem Statement

The adoption of quantum machine learning (QML) in clinical settings faces two primary barriers: the lack of patient-level interpretability and unverified claims of predictive superiority over classical models. While quantum kernels are proposed for clinical prediction, existing literature often fails to demonstrate that these models can provide additive, per-feature attributions trusted by clinicians (e.g., via SHAP values). Furthermore, predictive claims are frequently based on single, small datasets without rigorous multi-cohort validation, bandwidth tuning, or honest significance testing to distinguish real effects from noise. This study addresses these gaps specifically in the context of breast-cancer recurrence survival analysis.

2. Methodology

2.1 Data and Encoding

The study utilizes three independent breast-cancer cohorts:

WPBC: 194 patients (46 events).
GBSG2: 686 patients (299 events).
METABRIC: 1,903 patients (1,103 events).

To align with near-term quantum hardware constraints, the authors introduce the Tumour-Aggressiveness Composite (TAC), a three-dimensional feature vector encoding:

Log-transformed tumor diameter.
Log-transformed positive node count.
A summary of the worst-cell nuclear morphology.

This 3D encoding maps one-to-one onto a three-qubit circuit, utilizing a ZZ feature map with two repetitions and full entanglement.

2.2 Quantum Kernel and Survival Model

The quantum kernel is defined as the fidelity between encoded states: $K_q(x, x') = |\langle \phi(x) | \phi(x') \rangle|^2$ . To prevent the "concentration" phenomenon (where the kernel matrix approaches the identity, rendering the model unable to generalize), the embedding bandwidth ( $\beta$ ) is tuned via nested cross-validation. A classical Radial Basis Function (RBF) kernel is tuned with a corresponding length-scale ( $\gamma$ ) to ensure fair comparison.

The survival models employ:

Kernel Survival Support Vector Machine (SurvSVM) for the primary WPBC analysis.
Kernel Ridge Regression as a risk surrogate for scalable multi-cohort evaluation.
Baselines: Cox Proportional Hazards (on raw and composite features), Random Survival Forests, Gradient-Boosted Survival, and XGBoost.

2.3 Exact Shapley Attribution

The core methodological innovation is the derivation of exact Shapley values for the quantum-kernel survival model.

Theoretical Basis: Since the fitted risk is a fixed weighted sum of kernel similarities ( $f(x) = \sum \alpha_i K_q(x, x_i)$ ), the Shapley value can be computed by enumerating all $2^d$ feature coalitions (where $d=3$ ).
Efficiency: For the 3-qubit encoding, this enumeration is computationally tractable without sampling.
Validation: The authors prove that these attributions satisfy the efficiency axiom (the sum of attributions equals the model output minus the baseline) and verify this numerically to machine precision. They also compare these exact values against the sampled KernelSHAP estimator.

2.4 Noise and Evaluation

Robustness is assessed using an Aer density-matrix simulation incorporating single- and two-qubit depolarizing errors and readout errors, with simple mitigation via diagonal renormalization. Evaluation uses 5-fold stratified cross-validation, reporting Harrell's concordance index ( $C$ ) and paired bootstrap significance tests.

3. Key Contributions

Exact Explainability: The paper demonstrates that quantum-kernel survival models admit exact, axiom-satisfying Shapley explanations. For low-dimensional encodings (3 features/qubits), these can be computed via direct coalition enumeration, avoiding sampling error.
Honest Predictive Characterization: The study provides a transparent, multi-cohort evaluation of quantum kernels against classical baselines, correcting for kernel concentration via bandwidth tuning.
Noise Robustness: The survival pipeline is shown to be robust to depolarizing and readout noise levels representative of current hardware.
Reproducible Artifact: A fully reproducible framework where all reported numbers are generated by the analysis code.

4. Results

4.1 Explainability Performance

Axiom Satisfaction: The exact Shapley estimator satisfies the efficiency axiom to machine precision ( $\le 10^{-15}$ ).
Correlation: The exact values correlate with the sampled KernelSHAP estimator at $r = 0.97$ , validating the use of sampling methods when exact enumeration is infeasible.
Interpretability: The attributions align with clinical intuition (Tumor size and nodal burden dominate), providing per-patient explanations in oncologist-legible quantities.

4.2 Predictive Performance

Real-World Cohorts (WPBC, GBSG2, METABRIC):
- After bandwidth tuning, the quantum kernel is statistically indistinguishable from the classical RBF kernel on matched features across all three cohorts.
- The quantum model does not outperform strong classical nonlinear baselines (XGBoost, Random Survival Forests) or the Cox model on the composite features.
- Models using the 3-feature composite generally trail full-feature classical models (e.g., Cox on 9+ features).
Synthetic Study:
- The quantum kernel significantly outperforms a linear Cox model ( $\Delta C = +0.21$ ) only when the data-generating structure consists of pure pairwise-product interactions (matching the ZZ map's inductive bias).
- Even in this favorable synthetic case, the quantum kernel ties the bandwidth-tuned classical RBF kernel and is outperformed by a linear model given explicit interaction features.

4.3 Noise Robustness

Under depolarizing and readout noise models, the concordance index changes by less than 0.005, indicating the pipeline is robust at near-term noise levels.

5. Significance and Claims

The paper concludes that the practical case for quantum-kernel survival models rests on exact explainability rather than predictive accuracy.

Accuracy: There is no evidence of a quantum advantage in predictive discrimination on real breast-cancer data. The quantum kernel performs comparably to a well-tuned classical RBF kernel but offers no superiority over classical nonlinear baselines.
Explainability: The primary contribution is the ability to generate exact, sampling-free Shapley values for quantum models in the low-dimensional regime favored by near-term hardware. This provides a rigorous, axiom-verified explanation framework that matches the trustworthiness of tree-based SHAP.
Inductive Bias: Any potential advantage is strictly tied to data structures matching the kernel's inductive bias (e.g., specific feature interactions), which can also be captured by classical methods if the structure is known.

The authors explicitly state that the study does not support a claim of quantum advantage in accuracy for these clinical tasks but provides a reproducible template for evaluating such claims, emphasizing that the value proposition for quantum clinical models lies in interpretability and compact encoding, not in outperforming classical models on classical data.