Pith. sign in

REVIEW 4 major objections 6 minor 30 references

Investigating Quantum Feature Maps in Quantum Support Vector Machines for Lung Cancer Classification

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read On a 309-record lung-cancer dataset, a quantum support vector machine using the PauliFeatureMap achieves 0.96 average accuracy and perfect scores in three of six balanced subsets, outperforming the ZFeatureMap and ZZFeatureMap.

desk verdict Routine QSVM benchmark with no train/test split; the results table contradicts the abstract's central claim. read the letter →

arxiv 2506.03272 v1 pith:KC5TBZAT submitted 2025-06-03 quant-ph cs.LG

classification quant-phcs.LG
keywords QuantumMachineLearningSupportVectorFeatureMapsLungCancerDetectionHealthcareAIBinaryClassificationDiagnosisClassImbalance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper compares three quantum feature maps—ZFeatureMap, ZZFeatureMap, and PauliFeatureMap—inside a quantum support vector machine (QSVM) for classifying lung cancer from 309 patient records. Because the dataset is heavily imbalanced, the authors build six balanced subsets of 78 records each and evaluate accuracy, precision, recall, specificity, and F1-score. The central claim is that the PauliFeatureMap outperforms the other two maps overall, reaching perfect scores in three of the six subsets and an average accuracy of 0.96. This matters because feature-map choice is the main design lever in QSVM: if the comparison holds, it gives practitioners a concrete encoding to prefer for similar binary medical datasets on simulators.

What carries the argument

The load-bearing object is the quantum feature map, a parameterized circuit that encodes a patient record $x$ into the quantum state $|\psi(x)\rangle = U_{\phi(x)}|0\rangle^{\otimes n}$. The QSVM kernel is the fidelity $K_q(x_i,x_j) = |\langle\psi(x_i)|\psi(x_j)\rangle|^2$, replacing the classical inner product. The three maps differ in the rotations and couplings they apply: ZFeatureMap uses single-qubit $Z$ rotations, ZZFeatureMap adds pairwise $Z\otimes Z$ terms, and PauliFeatureMap applies $X$, $Y$, and $Z$ rotations with $Z\otimes X$ coupling. The feature map fixes the geometry of the quantum feature space, which determines what the classical SVM solver can separate.

What would settle it

Hold out a portion of each balanced subset before training, or run k-fold cross-validation, and recompute the five metrics: if the PauliFeatureMap's perfect scores and 0.96 average are not reproduced on held-out records, or if its ranking relative to the ZZFeatureMap and ZFeatureMap changes, the central claim is falsified.

Watch

Extended reading notes

Core claim

The paper argues that, for this lung-cancer dataset and simulator-based QSVM setup, the choice of quantum feature map materially changes classification performance, and the PauliFeatureMap is the best of the three tested. In its Table 2, the PauliFeatureMap achieves perfect accuracy, precision, recall, specificity, and F1-score on subsets 4, 5, and 6, and averages 0.96 across all six subsets, while the ZFeatureMap averages 0.88 accuracy and the ZZFeatureMap 0.91. The paper interprets this as evidence that the richer Pauli encoding captures the feature relationships needed to separate cancer from non-cancer records, and frames feature-map selection as the key decision in QSVM design.

Load-bearing premise

The central ranking assumes that metrics computed on the same 78 records per subset that trained the model also measure how the model will perform on patients it has not seen.

Editorial extensions

If this is right

  • If the comparison is correct, practitioners working with similarly structured binary medical data on quantum simulators should prefer the PauliFeatureMap over the ZFeatureMap and ZZFeatureMap to maximize accuracy, recall, and F1-score.
  • The RBF-kernel classical SVM baseline reaches 0.95 accuracy, close to the PauliFeatureMap QSVM's 0.96, so on this dataset the practical gain over a classical model is small; the main contribution is identifying which encoding uses the quantum feature space most effectively.
  • The paper's own discussion implies that on noisy real hardware the PauliFeatureMap's advantage may shrink or invert because its $X$ and $Y$ gates are sensitive to bit-flip errors, while the ZFeatureMap should remain the most noise-resilient.
  • The ZZFeatureMap's variability across subsets, with perfect scores on one subset but lower recall on others, suggests that pairwise-entangling encodings are more sensitive to which positive samples appear in a given balanced subset.
  • Feature-map choice should therefore be coupled to both the data's correlation structure and the error rates of the target quantum device, not treated as a single universally best encoding.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Our inference: because no train/test split is described, the reported perfect scores are best read as an expressivity result—the PauliFeatureMap can separate these particular 78-point subsets—and not yet as evidence that it will generalize to unseen patients.
  • Our inference: a stronger evaluation would use repeated random train/test splits or cross-validation and report the variance of the scores; with only 39 negative samples, six fixed subsets cannot by themselves establish generalization.
  • Our inference: comparing the PauliFeatureMap against a classical kernel that also includes feature products, such as a polynomial kernel, would help isolate whether the quantum feature space rather than mere nonlinearity drives the improvement.
  • Our inference: the paper's noise discussion yields a concrete hardware prediction—on a real noisy device, the ZFeatureMap should degrade least and the PauliFeatureMap most, potentially reversing the simulator ranking.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper compares three quantum feature maps (ZFeatureMap, ZZFeatureMap, PauliFeatureMap) in Quantum Support Vector Machines for lung cancer classification, using a Kaggle dataset of 309 patients (39 negative, 270 positive). Six balanced subsets of 78 samples each are constructed, and QSVM performance is evaluated with accuracy, precision, recall, specificity, and F1-score. The abstract claims that the PauliFeatureMap consistently outperformed the other feature maps, achieving perfect classification in three subsets. The paper also compares classical SVM kernels (linear, polynomial, RBF) as a baseline. The central empirical claim rests on Table 2 and the surrounding discussion.

Significance. If the results were valid, the paper would provide a useful systematic comparison of quantum feature maps in a medical classification setting, with a reasonable approach to class imbalance through balanced subsets and a clear presentation of the mathematical formulations of the feature maps and evaluation metrics. However, the central claim of consistent PauliFeatureMap superiority is not supported by the reported data, and the evaluation is in-sample with no held-out data, no error bars, and unreported hyperparameters. As presented, the paper does not establish that any feature map generalizes to unseen patients, so its significance for clinical or quantum-machine-learning practice is currently limited.

major comments (4)
  1. [Abstract; Section 4, Table 2] The claim that the PauliFeatureMap 'consistently outperformed' the other feature maps is directly contradicted by Table 2. On Subset 1, ZZFeatureMap achieves accuracy 1.000 while PauliFeatureMap achieves 0.875; on Subset 5, ZFeatureMap also achieves 1.000, tying with PauliFeatureMap. PauliFeatureMap is thus not uniformly the best across the six subsets, and the word 'consistently' is unsupported even for in-sample performance.
  2. [Section 4, Table 2] No train/test or validation split is described, and all metrics appear to be computed on the same 78 samples used to fit each model. With only 39 samples per class and 15 features, a quantum fidelity kernel is sufficiently expressive to memorize labels, so the perfect scores in Subsets 4–6 likely reflect interpolation rather than generalization. Without held-out evaluation or cross-validation, the reported accuracy and F1-scores cannot support the paper's claim that PauliFeatureMap is the best feature map for unseen patients.
  3. [Section 4; Table 2] The experimental setup is under-specified and uncontrolled. The manuscript does not report the SVM regularization parameter C, the RBF kernel parameter gamma, the polynomial kernel degree, the number of qubits or how the 15 features are mapped onto them, the number of circuit repetitions for each feature map, or the entanglement pattern (beyond the generic circuit diagrams). Also, each balanced subset is evaluated only once, with no repeated runs, error bars, or confidence intervals, so the differences between feature maps (e.g., 0.937 vs 0.875) cannot be assessed for statistical reliability.
  4. [Section 5; Conclusion] The Discussion misreports Table 2. It states that ZZFeatureMap 'achieves perfect scores for Subset 1, Subset 5, and Subset 6,' but Table 2 lists ZZFeatureMap accuracy 0.875 for both Subset 5 and Subset 6; only Subset 1 is perfect. It also asserts that PauliFeatureMap 'often avoids entanglement,' which contradicts Eq. (14), where the unitary explicitly contains the Z⊗X entanglement term. These inconsistencies undermine the qualitative narrative about expressivity and noise sensitivity.
minor comments (6)
  1. [References] Reference [24] appears to be a self-citation with a malformed author list ('M., Y., A. El Hafidi, M. Toufah, A., and Kadim') and no identifiable journal or year; please correct or remove it.
  2. [Section 4, Table 2] In Table 2, ZZFeatureMap Subset 3 precision is reported as 0.8000 with an unnecessary trailing zero; unify the number of decimal places across the table.
  3. [Section 3.1] The text says the six balanced subsets are 'without overlap' and 'preserving data integrity,' but the negative samples (D0) are shared across all six subsets; please clarify that only the positive samples are disjoint.
  4. [Section 4] The sentence 'The PauliFeatureMap was configured with Pauli rotations on the Z, X and Y axes' is too vague; specify the exact circuit parameters, including any repetition count and entanglement structure.
  5. [Section 4, Figure 5] The PCA visualization in Figure 5 is referenced but never discussed in the text; either add a brief analysis of what it shows or remove it.
  6. [General] There are formatting inconsistencies such as 'ZF eatureMap' for 'ZFeatureMap' and 'ZZF eatureMap' for 'ZZFeatureMap'; please fix the spacing throughout the manuscript.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; the self-citation [24] is minor and non-essential, so the empirical comparison remains self-contained.

full rationale

The paper's central claim is an empirical ranking of three Qiskit feature maps on an external Kaggle dataset, with the QSVM kernel defined by Eq. 9 and the feature map unitaries given by Eqs. 12-14. None of these equations is defined in terms of the reported accuracy, precision, recall, specificity, or F1-score results, and no parameter is fitted to a subset and then renamed as a prediction; the Table 2 metrics are direct outputs of the described pipeline. The only circularity-adjacent item is the self-citation [24] ('When classical svm fails: How quantum svm recovers recall in binary classification', by the same authors), cited in Sec. 3.3.1 for the standard statement that ZFeatureMap performs single-qubit Z-rotations; that statement is independently stated in Eq. 12, so the citation is not load-bearing. The absence of a described train/test split is a real validity and generalization weakness, and the 'consistently outperformed' wording is not supported by Table 2 (e.g., ZZFeatureMap reaches 1.000 on Subset 1), but these are correctness concerns, not circular reductions. Score 2 reflects the minor self-citation without treating it as central.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical or mathematical entities are introduced; the feature maps are standard Qiskit circuits. The main assumptions are standard QSVM background and dataset-specific domain choices. The critical assumption is in-sample evaluation, which is ad hoc to the paper and undermines the central claim.

free parameters (4)
  • Classical SVM hyperparameters (C, gamma, degree) = not reported
    Optimized on each balanced subset without a validation split, so they are fitted to the same data used for evaluation (Section 4).
  • QSVM regularization parameter C = not reported
    Used in the dual SVM objective (Eq. 10-11) but no value or search range is given.
  • Quantum feature map circuit parameters = not reported
    Number of qubits, number of circuit repetitions, and Pauli rotation coefficients are not stated; these determine the kernel matrix in Eq. (9).
  • Six random balanced subsets = six subsets of 78 samples each
    Random draws of 39 positives without replacement; no seed is given and all subsets share the same 39 negatives, so the six evaluations are not independent (Eq. 2).
assumptions (5)
  • standard math The quantum fidelity kernel K_q(x_i,x_j) = |<psi(x_i)|psi(x_j)>|^2 is a valid positive semidefinite kernel for SVM.
    Used in Eq. (9) following Havlicek et al.; the paper relies on this standard result without proof.
  • domain assumption The Kaggle lung cancer questionnaire features are sufficient for a meaningful diagnosis task.
    The paper treats 15 encoded self-report features as predictors for lung cancer without external clinical validation (Sections 3.1 and 4).
  • domain assumption Qiskit qasm simulator results are representative of quantum hardware behavior.
    All QSVM metrics come from noiseless simulation, yet the discussion makes claims about noise sensitivity (Sections 4 and 5).
  • ad hoc to paper Metrics computed on the same data used to fit a model measure classification performance.
    No train/test split or cross-validation is described; Table 2 reports in-sample metrics. This is the load-bearing assumption for the central claim.
  • domain assumption The six balanced subsets provide independent evaluations.
    All subsets share the same 39 negative samples, so they are correlated; the paper averages them without accounting for this (Eq. 2, Table 2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Investigating Quantum Feature Maps in Quantum Support Vector Machines for Lung Cancer Classification." pith.science (2026). https://pith.science/paper/KC5TBZAT

@misc{pith2026250603272,
  author       = {Pith},
  title        = {Pith review of: Investigating Quantum Feature Maps in Quantum Support Vector Machines for Lung Cancer Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KC5TBZAT}},
  note         = {Machine review of arXiv:2506.03272}
}
read the original abstract

In recent years, quantum machine learning has emerged as a promising intersection between quantum physics and artificial intelligence, particularly in domains requiring advanced pattern recognition such as healthcare. This study investigates the effectiveness of Quantum Support Vector Machines (QSVM), which leverage quantum mechanical phenomena like superposition and entanglement to construct high-dimensional Hilbert spaces for data classification. Focusing on lung cancer diagnosis, a concrete and critical healthcare application, we analyze how different quantum feature maps influence classification performance. Using a real-world dataset of 309 patient records with significant class imbalance (39 non-cancer vs. 270 cancer cases), we constructed six balanced subsets for robust evaluation. QSVM models were implemented using Qiskit and executed on the qasm simulator, employing three distinct quantum feature maps: ZFeatureMap, ZZFeatureMap, and PauliFeatureMap. Performance was assessed using accuracy, precision, recall, specificity, and F1-score. Results show that the PauliFeatureMap consistently outperformed the others, achieving perfect classification in three subsets and strong performance overall. These findings demonstrate how quantum computational principles can be harnessed to enhance diagnostic capabilities, reinforcing the importance of physics-based modeling in emerging AI applications within healthcare.

Figures

Figures reproduced from arXiv: 2506.03272 by the authors.

Figure 1
Figure 1. Illustration of a soft-margin SVM with support vectors, margin boundaries, and decision [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Quantum Circuit for ZFeatureMap [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Quantum Circuit for ZZFeatureMap 5 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Quantum Circuit for PauliFeatureMap The ZFeatureMap is a Pauli-based feature map that applies single-qubit rotations around the Z-axis [24]. While it does not involve entanglement, which limits its capacity to capture feature correlations, it is simple and efficient to…
Figure 5
Figure 5. Figure 5: PCA visualization of the six balanced subsets. Each plot represents a two-dimensional projec [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Grouped bar chart showing the average performance metrics of SVM models across different [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Grouped bar chart showing the average performance metrics of QSVM models across different [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 24 canonical work pages

  1. [24]

    El Hafidi, M

    M., Y., A. El Hafidi, M. Toufah, A., and Kadim. When classical svm fails: How quantum svm recovers recall in binary classification. Unknown, 2025

  2. [1]

    An Introduction to Support Vector Machines and Other Kernel-based Learning Methods

    Nello Cristianini and John Shawe-Taylor. An Introduction to Support Vector Machines and Other Kernel-based Learning Methods. Cambridge University Press, 2000

  3. [2]

    An introduction to quantum machine learning

    Maria Schuld, Ilya Sinayskiy, and Francesco Petruccione. An introduction to quantum machine learning. Contemporary Physics, 56(2):172–185, 2015

  4. [3]

    Supervised learning with quantum-enhanced feature spaces

    Vojtˇ ech Havl ´ ıˇ cek, Antonio D C´ orcoles, Kristan Temme, Aram W Harrow, Abhinav Kandala, Jerry M Chow, and Jay M Gambetta. Supervised learning with quantum-enhanced feature spaces. Nature, 567(7747):209–212, 2019

  5. [4]

    Quantum machine learning

    Jacob Biamonte, Peter Wittek, Nicola Pancotti, Patrick Rebentrost, Nathan Wiebe, and Seth Lloyd. Quantum machine learning. Nature, 549(7671):195–202, 2017

  6. [5]

    Application of support vector machine to diagnosis of lung cancer

    Chih-Cheng Huang, Hung-Yu Chen, and Chih-Cheng Lee. Application of support vector machine to diagnosis of lung cancer. Expert Systems with Applications , 22(1):63–71, 2002. 12

  7. [6]

    Lung cancer diagnosis using artificial immune recognition system and fuzzy resource allocation mechanism

    Kemal Polat and Salih G¨ une¸ s. Lung cancer diagnosis using artificial immune recognition system and fuzzy resource allocation mechanism. International Journal of Approximate Reasoning , 44(2):179– 194, 2007

  8. [7]

    Multi-stage lung cancer detection and predic- tion using multi-class svm classifier

    Janee Alam, Sabrina Alam, and Alamgir Hossan. Multi-stage lung cancer detection and predic- tion using multi-class svm classifier. Proceedings of the 2018 IEEE International Conference on Computational Intelligence and Communication Networks (IC4ME2) , pages 118–122, 2018

Show all 30 references
  1. [8]

    Shaik Parveen

    S. Shaik Parveen. Classification of lung cancer nodules using svm kernels. International Journal of Computer Applications, 95(25):1–5, 2014

  2. [9]

    Quantum machine learning for breast cancer detection on ibm quantum computers

    Santiago Ramos-Calderer, Oscar Viyuela, Luis Garc ´ ıa-´Alvarez, et al. Quantum machine learning for breast cancer detection on ibm quantum computers. Scientific Reports, 11(1):15655, 2021

  3. [10]

    Svore, and Nathan Wiebe

    Maria Schuld, Alex Bocharov, Krysta M. Svore, and Nathan Wiebe. Circuit-centric quantum clas- sifiers. Physical Review A , 101(3):032308, 2020

  4. [11]

    Quantum circuit learning

    Kosuke Mitarai, Makoto Negoro, Masahiro Kitagawa, and Keisuke Fujii. Quantum circuit learning. Physical Review A , 98(3):032309, 2018

  5. [12]

    Lung cancer dataset, 2025

    Ahmet S ¸endil. Lung cancer dataset, 2025. https://www.kaggle.com/datasets/ahmetsendil/ lung-cancer-dataset

  6. [13]

    Jain and Richard C

    Anil K. Jain and Richard C. Dubes. Algorithms for Clustering Data . Prentice-Hall, Upper Saddle River, NJ, 2000

  7. [14]

    Chawla, Kevin W

    Nitesh V. Chawla, Kevin W. Bowyer, Lawrence O. Hall, and W. Philip Kegelmeyer. Smote: Syn- thetic minority over-sampling technique. Journal of Artificial Intelligence Research , 16:321–357, 2002

  8. [15]

    B. E. Boser, I. M. Guyon, and V. N. Vapnik. A training algorithm for optimal margin classifiers. In Proceedings of the Fifth Annual Workshop on Computational Learning Theory , pages 144–152. ACM, 1992

  9. [16]

    Combarro, Samuel Gonzalez-Castillo, and Alberto Di Meglio

    Elias F. Combarro, Samuel Gonzalez-Castillo, and Alberto Di Meglio. A Practical Guide to Quantum Machine Learning and Quantum Optimization . Packt Publishing, 2023

  10. [17]

    Pontil and A

    M. Pontil and A. Verri. Support vector machines for 3d object recognition. Image and Vision Computing, 16(9):479–485, 1998

  11. [18]

    Sch¨ olkopf and A

    B. Sch¨ olkopf and A. J. Smola. Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond . MIT Press, Cambridge, MA, 2002

  12. [19]

    Cortes and V

    C. Cortes and V. Vapnik. Support-vector networks. Machine Learning, 20(3):273–297, 1995

  13. [20]

    Quantum machine learning in feature hilbert spaces

    Maria Schuld and Nathan Killoran. Quantum machine learning in feature hilbert spaces. Physical Review Letters, 122(4):040504, 2019

  14. [21]

    Quantum support vector machines for classification and regression on a trapped-ion quantum computer

    Teppei Suzuki, Takashi Hasebe, and Tsubasa Miyazaki. Quantum support vector machines for classification and regression on a trapped-ion quantum computer. Quantum Machine Intelligence , 6:31, 2024

  15. [22]

    The effect of linear discriminant analysis and quantum feature maps on qsvm performance for obesity diagnosis.International Journal of Quantum Computing, 3(2):45–58, 2022

    Mohamed Saad, Hossam Ahmed, and Nourhan Ali. The effect of linear discriminant analysis and quantum feature maps on qsvm performance for obesity diagnosis.International Journal of Quantum Computing, 3(2):45–58, 2022

  16. [23]

    Experimental kernel-based quantum machine learning in finite feature space

    Karol Bartkiewicz, Clemens Gneiting, Anton ´ ınˇCernoch, Kateˇ rina Jir´ akov´ a, Karel Lemr, and Franco Nori. Experimental kernel-based quantum machine learning in finite feature space. arXiv preprint arXiv:1906.04137, 2019

  17. [25]

    Sokolova and G

    M. Sokolova and G. Lapalme. A systematic analysis of performance measures for classification tasks. Information Processing & Management , 42(1):247–265, 2006. 13

  18. [26]

    D. M. W. Powers. Evaluation: From precision, recall and f-measure to roc, informedness, markedness & correlation. Journal of Machine Learning Technologies , 2:37–63, 2011

  19. [27]

    A survey of normalization techniques in deep learning

    Sergey Ioffe and Christian Szegedy. A survey of normalization techniques in deep learning. arXiv preprint arXiv:1609.02907, 2016

  20. [28]

    Handling imbalanced datasets in health- care machine learning: Challenges and solutions

    Alistair E W Johnson, Tom J Pollard, and Roger G Mark. Handling imbalanced datasets in health- care machine learning: Challenges and solutions. IEEE Engineering in Medicine and Biology Mag- azine, 39(5):28–35, 2020

  21. [29]

    Principal Component Analysis

    Ian T Jolliffe. Principal Component Analysis. Springer, 2nd edition, 2002

  22. [30]

    Principal component analysis

    Herv´ e Abdi and Lynne J Williams. Principal component analysis. Wiley Interdisciplinary Reviews: Computational Statistics , 2(4):433–459, 2010. 14

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.