Quantum annealing outperforms gate-based machines in lung cancer prediction from tiny training sets

Researchers have shown that a quantum machine learning model trained on just 108 patient records can identify lung cancer from symptom and risk data with an accuracy of 94.31 percent in simulation and 91.90 percent on a real quantum computer. The study, published in Discover Quantum Science on 9 September 2026, systematically compared three quantum classifiers across five quantum computing platforms, and found that an annealing-based quantum support vector machine running on D-Wave hardware delivered the most reliable performance under a deliberately constrained training regime designed to mimic the data scarcity that plagues real-world clinical settings.

The team, led by Muhammad Minoar Hossain and Arslan Munir with Md. Hasibul Hassan Himal, started from a publicly available lung cancer dataset of 309 patient records with 16 attributes covering lifestyle habits and symptoms such as smoking, alcohol consumption, coughing, wheezing, chest pain, fatigue, anxiety and shortness of breath. Because the original records were imbalanced, they applied the synthetic minority over-sampling technique, known as SMOTE, to construct a balanced experimental dataset of 540 records, with 270 samples in each of the lung cancer and non-cancer classes. The class labels were converted to a standard binary representation required by quantum classifiers, and the data were checked for missing values and duplicates, of which none were found.

Before any model training, the researchers turned to quantum computing itself to solve a problem that plagues small medical datasets: redundant and correlated features. They formulated feature selection as a quadratic unconstrained binary optimization problem, or QUBO, in which each feature is represented by a binary variable indicating whether it is selected. The objective function rewards features with high absolute Pearson correlation to the disease label, penalizes pairs of highly correlated features to reduce redundancy, and applies a linear penalty that discourages unnecessary feature inclusion. Solving this optimization with quantum annealing or quantum-inspired methods selected 10 features: allergy, alcohol consuming, wheezing, swallowing difficulty, coughing, chest pain, peer pressure, yellow fingers, chronic disease and smoking. A correlation heatmap of the data had already shown that the label correlated most strongly with allergy at roughly 0.64, alcohol consumption at about 0.60, wheezing and coughing near 0.58, swallowing difficulty around 0.52 and chest pain near 0.50, while several symptoms also co-occurred with each other, confirming that some features were overlapping.

The selected features were then encoded into quantum states using dense angle encoding, a quantum feature map in which each of the 10 qubits carries one feature through three rotation gates around the X, Y and Z axes, followed by entangling connections between neighboring qubits that spread information across the system. Because the Hilbert space of a quantum system grows exponentially with qubit number, even 10 qubits yield a highly expressive representation, though the researchers note that scaling of the input data is essential to prevent some rotation angles from becoming too large or negligible.

Three classifiers were built on this encoded data: a quantum support vector machine (QSVM), a quantum k-nearest neighbor model (QKNN) and a quantum boosting classifier (QBoost). Each was implemented in two fundamentally different quantum paradigms. The gate-based versions, executed on simulators in the Qiskit, Cirq and PennyLane frameworks, encode samples into entangled multi-qubit circuits and compute similarity through quantum state fidelity, replacing classical kernels and distances with quantum-defined measures. The annealing-based versions reformulate each learning task as a QUBO optimization: support vector selection for QSVM, neighbor selection for QKNN and weak-learner weighting for QBoost are all cast as binary optimization problems solved by evolving a system toward low-energy configurations. These ran on OpenJij, a quantum-inspired simulated annealing environment on classical hardware, and on the D-Wave Ocean SDK, which accesses a physical quantum annealer.

The evaluation protocol was deliberately unusual. Instead of the conventional large-training, small-test split, the 540-sample dataset was divided into five partitions of 108 samples each. In every fold, a single partition served as training data and the remaining 432 samples estimated prediction error, so the models had to learn from a small fraction of the data while being judged on most of the unseen samples. All experiments used cross-validation, and the researchers state that the training and testing partitions were strictly disjoint, so no information leaked between model fitting and error estimation. Metrics included accuracy, sensitivity, specificity, precision, F1-score, ROC-AUC and Cohen’s kappa, and all models ran on identical CPU-only cloud infrastructure to keep the comparison fair.

Across the gate-based frameworks, QKNN was consistently the strongest performer, with accuracy around 93 percent and balanced precision, sensitivity and specificity, while QSVM remained stable with high sensitivity and ROC-AUC. QBoost lagged noticeably, particularly in sensitivity and kappa, which the authors attribute to the combined effects of limited training data, sensitivity to weak-learner design and parameter settings, and the QUBO formulation used for learner selection. In the annealing setting, D-Wave generally outperformed OpenJij, and the D-Wave QSVM emerged as the best configuration overall, achieving 94.31 percent accuracy in simulation. An ablation analysis showed that QUBO-based feature selection and SMOTE each contributed to performance, but their effect varied by classifier and framework, with QBoost showing the largest variation across configurations. Convergence plots of the QUBO energy landscape showed the annealing process repeatedly visiting low-energy regions rather than scattering across poor solutions, which the authors link to the model’s stable predictive behavior.

The team then executed the winning D-Wave QSVM on physical quantum hardware through D-Wave’s Leap cloud service. Using four hybrid solvers, the constrained quadratic model (CQM) solver achieved the best results with 91.89 percent accuracy, 91.32 percent precision, 91.95 percent F1-score, 83.79 percent kappa and 98.07 percent ROC-AUC, while the binary and discrete quadratic model solvers produced identical results at 91.67 percent accuracy and the nonlinear solver trailed at 87.04 percent. On direct quantum processing units, two backend systems were compared: Advantage_system4, a fifth-generation Advantage processor with Pegasus topology, and Advantage_system6, a sixth-generation Advantage2 processor with Zephyr topology offering up to 20 qubit neighbors per qubit versus 15. Advantage_system6 won on every metric, reaching 91.90 percent accuracy against 90.05 percent for Advantage_system4, with QPU annealing times ranging from roughly 0.5 to 6 seconds and problem preparation under two seconds, though Advantage_system6 carried the highest queueing and network overhead.

An analytical benchmark against IBM’s gate-based quantum platform underscored the efficiency gap between the two paradigms. Because the IBM QSVM relies on quantum kernel evaluation, a full cross-validated run would require 58,320 kernel evaluations, or about 59.7 million circuit executions at 1,024 shots each, which the authors estimate at 5.81 hours per train-test split using IBM’s baseline timing of 0.00035 seconds per execution. In practice, the free IBM API token allows only ten minutes of hardware time, and the Qiskit QSVM failed outright because the required shot count exceeded the system limit of 10 million. For a reduced benchmark on 50 samples, D-Wave executed 1,085.59 times faster on actual quantum hardware with 25 percent higher accuracy, and 144.96 times faster in overall time. The authors caution that this comparison reflects differences in algorithmic formulation and computing paradigm, not hardware performance alone, since the D-Wave approach uses a QUBO formulation while IBM uses a quantum-kernel formulation.

Statistical analysis supported the QSVM advantage: paired t-tests and Wilcoxon signed-rank tests found significant differences between QSVM and both QKNN and QBoost at p below 0.05, and the 95 percent confidence intervals of QSVM maintained a higher performance range in both simulation and hardware settings. Against classical baselines including SVM, KNN, logistic regression, random forest, XGBoost, decision trees and multilayer perceptrons, the simulated D-Wave QSVM achieved the highest accuracy of all evaluated models, while the hardware version, at 91.90 percent, remained competitive despite a 2.41 percentage point drop attributed to hardware noise, connectivity constraints and embedding effects. The researchers are explicit about the limits of the work: the dataset is symptom- and risk-factor-based only, with no imaging, pathology reports, laboratory tests, cancer stage information or multi-center validation, so the framework should be viewed as a preliminary screening-support and methodological study rather than a clinical diagnostic system. Future work, they say, will extend the approach to larger multi-institutional datasets, integrate imaging and genomic information, explore explainability techniques and expand the binary task to multi-class lung cancer classification.

Subject of Research: Quantum machine learning for lung cancer prediction from limited symptom-based patient data using gate-based and annealing-based quantum computing platforms

Article Title: Inferring lung cancer under limited learning data through gate-based and annealing-based quantum techniques

Article References: Hossain, M. M., Himal, M. H. H., & Munir, A. (2026). Inferring lung cancer under limited learning data through gate-based and annealing-based quantum techniques. Discover Quantum Science, 2(1), Article 34. https://doi.org/10.1007/s44464-026-00038-4

Image Credits: AI Generated

DOI: 10.1007/s44464-026-00038-4

Keywords: quantum machine learning, lung cancer, quantum annealing, D-Wave, QSVM, QUBO, feature selection, SMOTE, cross-validation, quantum feature map, Qiskit, healthcare decision support