Hybrid quantum medical models get two-level explanations: exact attribution over 8 semantic dims, and CNOT pruning that raises accuracy under noise
Synopsis
Hybrid quantum-classical medical models can predict but cannot say which semantic factors or which quantum gates drove a decision. FQXAI compresses inputs into an eight-dimensional proxy-supervised semantic bottleneck, turns it into IF-THEN rationales through Gaussian fuzzy membership states, and computes exact Shapley values over all 256 semantic coalitions without Monte Carlo sampling. On MIT-BIH Arrhythmia, PneumoniaMNIST and PathMNIST the models reached accuracies of 0.890, 0.768 and 0.726, and under 5% depolarizing noise SVQX-guided pruning cut CNOT counts from 14 to 11, 12 and 10 while raising accuracy by 2.20, 1.74 and 1.51 percentage points.
Interpretation
A hybrid quantum-classical medical model can now report both semantic-level and gate-level attributions instead of only a probability. Earlier quantum explainability methods either analyzed only circuit gates or returned raw numerical weights without structured clinical logic, while classical SHAP-style methods typically rely on sampling approximations in high-dimensional feature spaces. On MIT-BIH Arrhythmia, PneumoniaMNIST and PathMNIST, hybrid FQXAI reached accuracies of 0.890, 0.768 and 0.726, in the same range as capacity-matched classical MLPs sharing the same upstream feature extractor at 0.888, 0.761 and 0.721.
Semantic attribution is carried out on an eight-dimensional bottleneck, so it can be enumerated rather than sampled. An eight-dimensional semantic space contains only 28, or 256, feature coalitions, so exact Shapley values can be computed for all of them without Monte Carlo sampling variance. Exact FQ-SHAP gave AUDC values of 0.780, 0.786 and 0.423 on the three datasets, where a lower AUDC means confidence drops faster when top-ranked features are deleted, indicating higher attribution fidelity.
Circuit-level analysis treats each CNOT gate as a player in a cooperative game and flags negatively contributing gates as pruning candidates. SVQX uses class-balanced true-label probabilities as the coalition value under a controlled depolarizing noise model, and CNOTs with negative contribution are marked for pruning under a fixed threshold of zero. Under 5% simulated depolarizing noise, SVQX-guided pruning reduced CNOT counts from 14 to 11, 12 and 10 with accuracy gains of 2.20, 1.74 and 1.51 percentage points, while a matched random-pruning control removing the same number of CNOTs reached accuracies of 0.898, 0.770 and 0.723.
Perspective
The demonstrated scope is simulation on three public datasets: single-beat binary classification on MIT-BIH Arrhythmia, 28×28 grayscale chest radiographs from PneumoniaMNIST, and nine-class colorectal tissue images from PathMNIST, with the quantum part run on a PennyLane density-matrix simulator under a 5% depolarizing channel. The intended readers are researchers and engineers working on quantum model interpretability and circuit noise analysis. The authors propose extending the noise model to backend-calibrated noise and physical QPU execution, and adding expert-annotated clinical concepts with inter-rater agreement studies.
The semantic concepts are programmatically constructed proxy targets rather than expert-annotated clinical concepts, and the fuzzy labels describe regions of the learned concept space. Attribution depends on the reference baseline: AUDC for the zero semantic baseline versus the training-mean baseline was 0.780 versus 0.792, 0.786 versus 0.801, and 0.423 versus 0.438 on the three datasets. The authors describe eight qubits as a trade-off between semantic capacity and exact explainability under current simulators and near-term hardware.