{"id":"55f0711e-ccd2-4b02-becc-c1c3419b91e6","arxiv_id":"2509.02863","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"QI-SMOTE claims quantum-inspired oversampling improves imbalanced medical classification, but the described VQE step is self-defeating and the evidence is internally inconsistent.","lead":"The paper proposes QI-SMOTE, an oversampling method that adds quantum-inspired encoding and a variational quantum eigensolver loop before applying SMOTE, and reports F1 gains on MIMIC-III and MIMIC-IV mortality prediction. The method as printed is internally inconsistent: the core quantum step, by the paper's own equations, either discards or fails to add the information it claims to preserve, so the reported gains cannot be attributed to the described mechanism.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 2.1.1 defines H=|Ψ⟩⟨Ψ|; VQE then minimizes |⟨ψ|Ψ⟩|², so the optimized state is driven toward orthogonality with the minority sample it was supposed to enrich—the central mechanism of QI-SMOTE is internally inverted.","rationale":"The reader's weakest-assumption analysis correctly isolates the VQE step: the Hamiltonian in Algorithm 2.1.1 is a projector onto the encoded state, so minimizing its expectation actively drives the evolved state away from that state. This is an internal, mechanism-level defect, not a disagreement with external consensus or a matter of tuning. A successful VQE does not produce an enriched version of |Ψ⟩; it produces a state that has been optimized to be orthogonal to it. The appendix's alternative Ising Hamiltonian does not rescue the method because it contradicts the main text and leaves the 'learned' coupling weights unspecified. Without a well-defined enrichment mechanism, the central claim that QI-SMOTE outperforms SMOTE-family methods cannot be explained by the proposed algorithm, even if some reported numbers are reproducible. I also credit the paper for providing a runtime benchmark and for attempting qualitative analysis, but those do not address the core defect. A corrected paper would need to redefine H (e.g., as I−|Ψ⟩⟨Ψ| or a data-dependent Ising Hamiltonian with a specified learning rule), define the feature-append operation, and show an ablation isolating the quantum step's contribution. Since the reader already rejected the paper on essentially these grounds, the verdict should remain REJECT (no change).","tokens_in":19415,"tokens_out":4000,"duration_ms":48320,"concrete_test":"Implement Algorithm steps 1–2 literally for one 9-feature MIMIC-IV minority sample: build |Ψ⟩ with RY/CNOT/CZ/Toffoli, set H=|Ψ⟩⟨Ψ|, run the Appendix A.6 COBYLA VQE (100 iterations), and record |⟨Ψ|Ψ_evolved⟩|² and the final variational energy. If the overlap is near zero (or the energy near 0), the step has discarded the encoded sample. Then train RF on the actual QI-SMOTE features and on a control where the VQE step is replaced by an arbitrary vector orthogonal to |Ψ⟩; if F1/G-Mean are unchanged, any measured gain is not attributable to the claimed quantum enrichment. If the implementation instead follows Appendix A.6, the authors must specify exactly how w_ij and b_i are learned; without that, the algorithm is not reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Algorithm 2.1.1–2.1.2. The paper constructs H = |Ψ⟩⟨Ψ| and asks VQE to minimize ⟨ψ(θ)|H|ψ(θ)⟩ = |⟨ψ(θ)|Ψ⟩|². This is a rank-one projector with eigenvalues 1 and 0; its ground state is any state in the orthogonal complement of |Ψ⟩. Thus a converged VQE produces |Ψ_evolved⟩ with zero overlap with the encoded minority sample, not a state that 'preserves and enriches' it. The paper's own Appendix A.6 defines a different Ising Hamiltonian H = Σ w_ij Z_i Z_j + Σ b_i Z_i with 'learned' weights, but no learning procedure or weight values are given, so the algorithm specification is self-contradictory: either the main-text Hamiltonian discards the sample, or the appendix Hamiltonian is under-specified. Since the only difference between QI-SMOTE and classical SMOTE is this VQE-transformed feature space, the reported performance advantage has no described causal basis. The separate step 3.1, appending a 2^n-dimensional real vector to an n-feature table, is also undefined (for 9 features, 512 new columns per row), but the VQE inversion alone is sufficient to invalidate the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes QI-SMOTE, a preprocessing method that augments SMOTE with quantum-inspired operations. Minority samples are encoded into qubit rotations and entangled with CNOT/CZ/Toffoli gates; the resulting state |Psi> is evolved by a VQE loop whose Hamiltonian is defined as H = |Psi><Psi|; the real part of the evolved state is appended to the dataset, and SMOTE is applied on the augmented feature space. The method is evaluated on MIMIC-III and MIMIC-IV mortality prediction at imbalance ratios 2.05-20 with six classifiers and compared to SMOTE-family baselines, reporting F1, G-Mean, and AUC improvements. The central claim is that QI-SMOTE produces more informative synthetic data and outperforms conventional resampling methods on imbalanced medical data.","tokens_in":19718,"tokens_out":5118,"duration_ms":57588,"significance":"If the claims were supported, QI-SMOTE would be a reusable, structure-aware oversampling method for imbalanced clinical tabular data. The paper deserves credit for a fairly detailed algorithmic description, evaluation on public medical datasets, a runtime/memory benchmark, and a Wilcoxon test. However, the defining quantum-evolution step is internally inconsistent as specified, and the append operation is undefined; therefore the paper does not establish a mechanism for the reported gains. The significance of the empirical tables depends on a mechanism that the manuscript itself contradicts.","major_comments":[{"comment":"The main text defines H = |Psi><Psi| (step 2.1.1) and then asks VQE to minimize <psi(theta)|H|psi(theta)> = |<psi(theta)|Psi>|^2. This rank-one projector has ground states in the orthogonal complement of |Psi>. Thus a converged VQE returns |Psi_evolved> with zero overlap with the minority sample it was supposed to preserve and enrich; even partial convergence systematically reduces that overlap. Appendix A.6 instead defines an Ising Hamiltonian H = sum w_ij Z_i Z_j + sum b_i Z_i with 'learned' weights, but gives no learning procedure or weight values. The two specifications are contradictory, and neither supports the claim that the evolution step enriches the sample. Because this VQE step is the only component that distinguishes QI-SMOTE from classical SMOTE, the reported performance advantage has no described causal basis.","section":"3.2, Algorithm steps 2.1.1-2.1.2; Appendix A.6"},{"comment":"The operation 'Append |Psi_real> to the original dataset D' is undefined. For an n-feature sample, |Psi_real> is a vector of length 2^n (512 components for the 9 features used here). The paper does not specify whether these components become new columns, how they are aligned with the original rows, how negative or complex-valued entries are handled, or how SMOTE interpolates in the resulting mixed feature space. This step is load-bearing for the 'enriched feature space' claim, and its undefined nature makes the algorithm irreproducible.","section":"3.2, step 3.1"},{"comment":"The statistical significance analysis covers only one comparison (QI-SMOTE vs. SMOTE-ENN on MIMIC-IV (10)), while the conclusion claims broad superiority over all conventional resampling methods across six dataset variants and many classifiers. No correction for multiple comparisons is reported. This single paired test does not substantiate the general claim, even setting aside the mechanism problem raised in Major Comment 1.","section":"5.3"}],"minor_comments":[{"comment":"The text describes QI-SMOTE's improvements as 'exponential'; the reported values are ordinary percentage improvements over a low baseline and should be described as such.","section":"5.2"},{"comment":"The abstract promises evaluation with Neural Networks, but the results tables do not report a Neural Network classifier.","section":"Abstract and Section 4"},{"comment":"Equation (1) is mis-typeset/torn in the manuscript; the formula for percentage improvement should be typeset cleanly.","section":"Equation (1)"},{"comment":"Several figures are labeled 'Figure 2. (continued)' and 'Figure 3. (continued)' with no explanatory captions for the repeated panels; captions and numbering should be consolidated.","section":"Figures 2-3"},{"comment":"Feature selection and preprocessing for MIMIC-III/IV are described only at a high level; no code or detailed pipeline is provided, which hinders reproducibility of the empirical tables.","section":"Section 4.2"}],"recommendation":"reject","confidential_remarks":"To the editor: the header stating that the paper has already been accepted for publication is unusual for a submitted manuscript and does not influence my technical assessment. The decisive issue is the internally contradictory specification of the VQE step: with the Hamiltonian defined in the main text, the evolution step drives the state toward orthogonality with the input, so the central mechanism cannot be as claimed. Even if the authors intended the Ising Hamiltonian of Appendix A.6, no learning procedure or weights are supplied, and the augmentation dimension remains undefined. I would not consider this paper acceptable without a corrected, code-backed specification and new experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Sarah,\n\nYou're right about this one, and the flaw is worse than a missing ablation. Algorithm step 2.1.1 sets H = |Ψ⟩⟨Ψ|. VQE minimizes ⟨ψ|H|ψ⟩ = |⟨ψ|Ψ⟩|². The ground state of that rank-one projector is any state orthogonal to |Ψ⟩. So a converged VQE doesn't enrich the minority sample; it erases it. The appendix tries to fix this with an Ising Hamiltonian with 'learned' weights, but no learning procedure is given. So the algorithm, as specified, is self-contradictory.\n\nWhat's new is the specific combination: layered CNOT/CZ/Toffoli entanglement plus a VQE loop before SMOTE. That's a fresh mix, and the motivation is real—SMOTE can create implausible medical samples (BP 160 with HR 40), and structure-preserving oversampling would be useful. The survey of SMOTE variants is competent. But the mechanism doesn't deliver that preservation.\n\nOther problems: appending a 2^n-dimensional state vector to an n-feature table is never defined—for 9 features that's 512 columns per row, and SMOTE would interpolate in that space without explanation. The Wilcoxon analysis uses fold-level F1 scores around 0.78, while Table 3 reports 0.705 and 0.688 for the same configuration; those can't both be right. Only one comparison on one dataset was tested, and it produced the minimum possible p-value. No code, no data, no seeds. The abstract claims neural network results, but no NN results appear. References [11]-[13] are never cited. The limitations section mentions runtime and hyperparameter tuning but misses the mechanism-level failures.\n\nThe qualitative claim that QI-SMOTE keeps synthetics on the minority manifold is supported by one PCA picture, not by a measured distance-to-manifold or plausibility metric.\n\nSo the paper's central claim—that QI-SMOTE outperforms SMOTE-family methods because it preserves structure—has no described causal basis. The reported performance gains, if real, would need a different algorithm to explain them.\n\nThis is a cautionary read for anyone working on quantum-inspired preprocessing or oversampling. A serious referee should see it, because catching the Hamiltonian issue requires reading the equations carefully, so it deserves review rather than desk reject. But my verdict is reject as is; it needs a redesigned evolution step, consistent statistics, and released code before any of the results can be trusted.","headline":"The VQE step in QI-SMOTE drives the evolved state orthogonal to the sample it's supposed to enrich, and the numbers don't match the tables.","tokens_in":20335,"tokens_out":4738,"would_cite":false,"duration_ms":47355,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"QI-SMOTE, a quantum-inspired oversampling method, raises F1, G-Mean, and AUC for imbalanced MIMIC mortality prediction beyond SMOTE-family baselines—by up to 165 percent at the most severe imbalance.","keywords":["quantum-inspired computing","class imbalance","synthetic oversampling","SMOTE","medical machine learning","MIMIC dataset","variational quantum eigensolver","data preprocessing"],"falsifier":"Take any minority sample used in the paper, form |Ψ⟩ by the algorithm's encoding, run the VQE routine with H = |Ψ⟩⟨Ψ|, and record |⟨Ψ|Ψ_evolved⟩|. If that overlap is near zero (as minimization of ⟨ψ|H|ψ⟩ implies), the central mechanism—evolved states that preserve and enrich the sample—is not operative, and the reported gains must come from elsewhere in the pipeline.","tokens_in":19156,"feed_emoji":"⚛️","tokens_out":6613,"duration_ms":68387,"temperature":0.7,"pith_summary":"The paper introduces QI-SMOTE, a preprocessing method that applies quantum-inspired state encoding—superposition, layered entanglement via CNOT, CZ, and Toffoli gates, then VQE-based evolution—to minority-class samples before classical SMOTE interpolation, and claims this produces synthetic medical data that better preserve feature dependencies. The authors test it on MIMIC-III and MIMIC-IV mortality prediction at imbalance ratios from 2.05 to 20, comparing with SMOTE, ADASYN, Borderline-SMOTE, SMOTE-ENN, SMOTE-TOMEK, and SVM-SMOTE across RF, SVM, LR, KNN, GB, and neural classifiers. Their headline empirical claim is that QI-SMOTE consistently raises F1, G-Mean, and AUC beyond these baselines, with the largest relative gains at the most severe imbalance (165% F1 improvement over the original dataset on MIMIC-IV(20)). The paper also argues the entanglement step prevents unrealistic physiological combinations (e.g., high blood pressure with very low heart rate) that classical SMOTE can generate. A sympathetic reader would care because, if correct, it offers a reusable, algorithm-agnostic preprocessing step that makes imbalanced clinical ML both more accurate and more plausible at the sample level.","feed_headline":"Quantum-inspired oversampling lifts ICU mortality F1 by up to 165%","feed_subtitle":"Hybrid SMOTE that entangles features before interpolation beats SMOTE-family baselines on MIMIC-III/IV.","key_machinery":"The load-bearing object is the layered-entanglement state |Ψ⟩ produced per minority sample: Hadamard + RY(f_i) encode features, CNOT/CZ/Toffoli layers couple qubits, and VQE with Hamiltonian H = |Ψ⟩⟨Ψ| is used to 'evolve' the state. The paper uses this evolved state, after taking its real part, as an enriched feature vector appended to the original data, so that SMOTE interpolates in a space the method claims preserves physiological co-dependencies rather than raw coordinate space.","core_discovery":"QI-SMOTE encodes each minority sample into an n-qubit state by mapping features to RY rotation angles after a Hadamard initialization, entangles the qubits with CNOT, CZ, and Toffoli gates, then forms the Hamiltonian H = |Ψ⟩⟨Ψ| from the resulting state vector and runs VQE to obtain an 'evolved' state. The real part of that evolved state is appended to the original feature table, and the enlarged dataset is passed to classical SMOTE for interpolation. The paper's central claim is that this pipeline yields synthetic minority samples that preserve high-order feature correlations—unlike linear interpolation in raw feature space—and that classifiers trained on QI-SMOTE-balanced data outperform th","pith_inferences":["Because minimizing ⟨ψ|H|ψ⟩ for H = |Ψ⟩⟨Ψ| pushes the optimized state toward the orthogonal complement of |Ψ⟩, a straightforward check of the inner product ⟨Ψ|Ψ_evolved⟩ on actual minority samples would show whether VQE is preserving or erasing the encoded data; this is a testable consequence the paper does not report.","The algorithm never specifies how a 2^n-dimensional real vector (512 components for 9 features) is appended to an n-feature table, so the empirical gains may be an artifact of SMOTE interpolating a high-dimensional expanded space rather than of the quantum-inspired correlations; a control that appends random or classically projected noise would isolate the mechanism.","If the gains survive that noise-control ablation, QI-SMOTE's contribution reduces to 'expand features then SMOTE', which would transfer directly to other tabular imbalance problems; if not, the quantum gates and VQE are not the active ingredient.","The runtime/memory gap suggests that until real quantum hardware is available, the practical value of QI-SMOTE depends on whether the F1/AUC gains justify roughly a thousand-fold slowdown relative to SMOTE; comparing at a fixed compute budget would be a more decision-relevant benchmark."],"forward_implications":["If QI-SMOTE works as claimed, F1, G-Mean, and AUC on imbalanced medical tabular data improve across tree, kernel, and neural classifiers whenever conventional SMOTE-family resampling is replaced by QI-SMOTE preprocessing.","The improvement is largest at severe imbalance: the paper reports 89.5% and 165.3% F1 improvements over the unaltered dataset on MIMIC-IV(10) and MIMIC-IV(20), respectively, where baselines plateau.","Because the method is a preprocessing step, it can be dropped into existing ML pipelines without modifying the classifier.","Entanglement co-modulates features, so synthetic samples avoid physiologically implausible combinations such as high systolic blood pressure with very low heart rate.","In its current classical-simulation form, QI-SMOTE costs orders of magnitude more compute and memory than SMOTE; the authors argue this tradeoff would disappear on real quantum hardware."],"supporting_citations":[{"why":"Supplies the baseline interpolation method that QI-SMOTE extends.","marker":"[3]"},{"why":"Borderline-SMOTE, one of the baseline oversampling methods QI-SMOTE must beat.","marker":"[15]"},{"why":"ADASYN, an adaptive synthetic sampling baseline used in the comparison.","marker":"[16]"},{"why":"SMOTE-ENN, a baseline and the primary technique used in the Wilcoxon significance test.","marker":"[17]"},{"why":"SMOTE-TOMEK, a baseline it must beat in the empirical comparison.","marker":"[18]"},{"why":"SVM-SMOTE, a baseline that generates samples in support regions.","marker":"[19]"},{"why":"Prior quantum SMOTE work that the paper extends by adding layered entanglement and VQE.","marker":"[36]"},{"why":"Supplies the VQE optimization routine used at algorithm step 2.1.2.","marker":"[39]"},{"why":"MIMIC-III dataset, the first benchmark for mortality detection.","marker":"[40]"},{"why":"MIMIC-IV dataset, the second benchmark for mortality detection.","marker":"[44]"}],"fun_headline_variants":["QI-SMOTE: quantum-inspired oversampling lifts ICU mortality F1","Entangle features, beat imbalance: QI-SMOTE boosts ML in medicine","Quantum SMOTE improves classifiers by 165% on MIMIC mortality","QI-SMOTE: quantum-inspired resampling cuts medical data bias","Quantum entanglement for oversampling: QI-SMOTE on MIMIC"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The method's gains rest on the assumption that the VQE step enriches each minority sample while keeping its information; the Hamiltonian it defines, H = |Ψ⟩⟨Ψ|, instead drives the optimized state toward the complement of the encoded sample, so the assumed preservation may not be what happens.","fun_headline_variants_meta":{"raw":{"variants":["QI-SMOTE: quantum-inspired oversampling lifts ICU mortality F1","Entangle features, beat imbalance: QI-SMOTE boosts ML in medicine","Quantum SMOTE improves classifiers by 165% on MIMIC mortality","QI-SMOTE: quantum-inspired resampling cuts medical data bias","Quantum entanglement for oversampling: QI-SMOTE on MIMIC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1203,"prompt_tokens":855,"completion_tokens":348,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":250}},"tokens_in":599,"tokens_out":348,"duration_ms":4636,"temperature":1.0,"reasoning_tokens":250,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:22:08.058594+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any minority sample used in the paper, form |Ψ⟩ by the algorithm's encoding, run the VQE routine with H = |Ψ⟩⟨Ψ|, and record |⟨Ψ|Ψ_evolved⟩|. If that overlap is near zero (as minimization of ⟨ψ|H|ψ⟩ implies), the central mechanism—evolved states that preserve and enrich the sample—is not operative, and the reported gains must come from elsewhere in the pipeline.","supporting_citations":[{"cited_title":"In: D.-S","cited_arxiv_id":null,"evidence_quote":"Borderline-SMOTE, one of the baseline oversampling methods QI-SMOTE must beat."},{"cited_title":"A., and Li, S.: ADASYN: Adaptive synthetic sampling approach for imbalanced learning","cited_arxiv_id":null,"evidence_quote":"ADASYN, an adaptive synthetic sampling baseline used in the comparison."},{"cited_title":"E., et al.: SMOTE –ENN-Based Data Sampling and Improved Dynamic Ensemble Selection for Imbalanced Medical Data Classification","cited_arxiv_id":null,"evidence_quote":"SMOTE-ENN, a baseline and the primary technique used in the Wilcoxon significance test."},{"cited_title":"In: 2016 IEEE International Conference of Online Analysis and Computing Science (ICOACS)","cited_arxiv_id":null,"evidence_quote":"SMOTE-TOMEK, a baseline it must beat in the empirical comparison."},{"cited_title":"V., and Krasser, S.: SVMs Modeling for Highly Imbalanced Classification","cited_arxiv_id":null,"evidence_quote":"SVM-SMOTE, a baseline that generates samples in support regions."},{"cited_title":"A Quantum Approach to Synthetic Minority Oversampling Technique (SMOTE)","cited_arxiv_id":"2402.17398","evidence_quote":"Prior quantum SMOTE work that the paper extends by adding layered entanglement and VQE."},{"cited_title":"P., and Jayapandian, N.: Challenges and opportunities: Quantum computing in ML","cited_arxiv_id":null,"evidence_quote":"Supplies the VQE optimization routine used at algorithm step 2.1.2."}],"review_version":1}