Quantum kernels and classical learners, working on the same biomedical data.
QuantumX encodes multi-omics panels, longitudinal labs, and structured clinical records into quantum feature spaces where higher-order interactions stay intact. The objective is narrow and measurable: recover diagnostic signal at the pre-symptomatic stage that linear projections and tree ensembles smooth away — and prove it against the classical baseline every time.
- 12–20
- logical qubits per circuit
- IBM Heron
- QPU execution backend
- 5×2
- cross-validated splits
- p < 0.05
- reporting threshold
Where classical models stop improving
Gradient boosting on tabular clinical data is a strong, well-understood baseline, and on most tasks it should be the deployed model. The failure modes below are the ones we have repeatedly observed where it is not enough — and they share a cause: the diagnostic information is in the geometry of feature interactions, not in the features.
Epistasis is not additive
Risk in polygenic disease frequently lives in the interaction term. A model that scores loci independently — or with pairwise interactions hand-selected in advance — cannot represent a three-way epistatic effect it was never told to look for.
PCA discards the manifold
Reducing 20,000 transcripts to 50 components preserves variance, not structure. Variance is dominated by batch, tissue composition, and demographics. The curvature that separates an early lesion from benign tissue is low-variance and is the first thing a linear projection throws out.
Early signal has low amplitude
At stage I, the discriminative shift in a circulating biomarker is often within assay noise for any single analyte. Detection depends on the joint configuration across dozens of weak markers, which is precisely the regime where regularised classical models collapse toward the majority class.
p ≫ n, and the cohorts are small
Curated, well-phenotyped cohorts run to hundreds of patients against tens of thousands of features. Deep networks overfit; sparse linear models underfit. Neither failure is fixed by more epochs.
Four stages, one auditable run record
Every execution writes a versioned record: preprocessing hash, circuit ansatz, transpiled depth, backend calibration snapshot, seeds, and the classical baselines run alongside it. A result you cannot reproduce is not a result.
Non-linear feature encoding
A supervised autoencoder compresses the raw panel to a 12–20 dimensional latent vector sized to the circuit width. The bottleneck is trained jointly with the downstream objective, so compression is driven by class separability rather than reconstruction error alone. Latent dimensions are retained with their loadings intact for later attribution.
Geometric pre-screening
Before a single shot is executed, we compute the quantum kernel Gram matrix on the training fold and test whether it induces separation the classical RBF kernel does not: kernel-target alignment, effective dimension, and the spectral gap between the two Gram matrices. If the quantum kernel offers no geometric advantage, the run is flagged and the classical model is recommended. Most datasets do not pass this gate, and we say so.
Hybrid inference
Variational classifiers and quantum kernel SVMs run on IBM Heron QPUs through Qiskit Runtime, with the identical fold executed simultaneously against XGBoost, RBF-SVM, and regularised logistic regression. Statevector simulation, noisy simulation, and hardware results are stored as three distinct records — never averaged, never substituted for one another.
QXplain attribution engine
SHAP assumes a classical additive decomposition that a entangled circuit does not satisfy. QXplain instead ablates individual gates and entangling blocks, measuring the shift in decision margin to attribute the prediction to specific feature interactions. Attribution is propagated back through the autoencoder loadings to named genes, analytes, and clinical variables.
From raw cohorts to auditable discovery
QuantumX connects directly to clinical pipelines. Explore the interactive research views below to see how our hybrid framework ingests cohorts, verifies geometric advantage in Hilbert space, and attributes predictions to biological drivers.
Mechanism: Non-linear quantum feature mapping resolves low-abundance ERBB2 transcript clusters missed by regularized linear baselines.
We do not claim an advantage we cannot measure
Quantum and classical models are trained and evaluated on identical stratified splits with identical preprocessing and identical seeds. Nothing is tuned on the test fold. Comparisons use McNemar's test on paired predictions for discordance and a paired t-test across cross-validation folds for aggregate metrics, with Benjamini–Hochberg correction when several architectures are compared at once.
A difference in AUC that does not clear the significance threshold is reported as no difference. When the classical baseline wins, that is the finding we publish, and the platform recommends the classical model for that cohort.
“A negative result on a well-designed comparison is more useful to a clinical team than an unreplicable positive one.”
| Model | AUC | Sens. | Spec. | p |
|---|---|---|---|---|
| Logistic regression (L2) | 0.812 | 0.74 | 0.79 | reference |
| RBF-SVM | 0.841 | 0.77 | 0.81 | 0.041 |
| XGBoost | 0.869 | 0.81 | 0.83 | 0.012 |
| Quantum kernel SVM (sim) | 0.884 | 0.85 | 0.82 | 0.038 |
| Quantum kernel SVM (QPU) | 0.877 | 0.84 | 0.81 | 0.061 |
Figures shown are an illustrative record layout, not published performance. Hardware rows carry wider confidence intervals from shot noise and device drift; simulator results are never reported as hardware results.