{"id":"8af9efe3-fc0b-4f76-b9ba-fd7e82b799f9","arxiv_id":"2506.03199","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"QCML, a quantum-inspired ML method, predicts LST status from CTC morphology with balanced accuracy 70%, only marginally better than classical models in a 227-cell study.","lead":"This paper applies a quantum-inspired machine learning method called QCML to predict chromosomal instability from images of circulating tumor cells in metastatic breast cancer patients. The authors report that QCML matches or slightly beats ten classical models on a small dataset, but the improvements are within statistical uncertainty.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'out-of-sample' comparison uses cell-level folds that split patients across train/test; because each patient contributes ~3.3 correlated CTCs, this can inflate QCML's apparent advantage and does not establish generalization to new patients.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: cell-level cross-validation splits correlated cells from the same patient across train and test, so the 'out-of-sample' result can be inflated by patient leakage. This is the most direct threat to the central claim that QCML outperforms classical ML on verification CTCs, because the claim's meaning depends on generalization to unseen patients or at least to independent cells. The paper explicitly discloses the case-agnostic protocol, so this is not a hidden methodological error, but disclosure does not remove the validity problem for the stated clinical aim. My concrete test would settle whether the concern lands: a grouped patient-level split directly estimates the quantity the abstract claims to measure. I agree with the reader's rejection: the reported numerical differences are small, no significance tests are supplied, and the evaluation protocol does not support the headline generalization claim. However, the QCML framework and POVM extension are coherent, and the data represent a useful preliminary contribution, so if the grouped-split test preserves the advantage and significance tests are added, the paper could become a valid preliminary finding. The verdict should remain REJECT as currently stated, which corresponds to UNCHANGED relative to the reader's verdict.","tokens_in":17068,"tokens_out":3959,"duration_ms":47275,"concrete_test":"Re-run the complete model-comparison pipeline with patient-level grouped cross-validation: use GroupKFold (or repeated grouped splits) with patient ID as the group so that all cells from a given patient are in exactly one fold, then compare QCML and the top classical baselines (SVM Gaussian, XGBoost, Random Forest) on out-of-fold sensitivity, specificity, balanced accuracy, and ROC AUC, with bootstrap confidence intervals. If QCML's advantage over the best baseline shrinks to within one standard deviation or reverses, the claimed superiority over conventional methods is not supported. Additionally, run a paired significance test (e.g., McNemar or paired bootstrap) on the original cell-level folds; if no test reaches p < 0.05, the claim of outperformance remains unsubstantiated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that QCML 'outperforms conventional machine learning methods when tested on out of sample verification CTCs' depends on the evaluation protocol in Section 2.2. The protocol is explicitly 'case-agnostic': each cell is treated as independent and 5-fold cross-validation splits cells, not patients. The dataset contains 227 CTCs from 73 patients (mean 3.25 cells/patient), so when a patient's cells are split across folds, test cells have same-patient cells in the training set. Cells from the same patient share biological state, image-acquisition conditions, staining batches, and other technical factors, so they are not independent. The model can therefore learn patient-specific morphology without seeing a patient identifier, and the reported 'verification' performance reflects partial patient memorization rather than generalization to unseen patients. This concern applies to all models, but it does not automatically preserve the ranking: QCML's quantum-state representation may be especially sensitive to within-patient correlations, and classical baselines could be comparatively less inflated or vice versa. The abstract and Discussion frame the result as a liquid-biopsy tool for 'real-time detection' in patients, so the intended deployment is new-patient prediction. Under the current cell-level split, the headline superiority claim is not established. A secondary but related issue is that even under the reported split, the QCML advantage over the best classical baselines (balanced accuracy 70% vs 68%; AUC 0.763 vs 0.747) is within the reported standard deviations, and no paired significance test is provided. The load-bearing problem, however, is the independence assumption: the evaluation design does not support the paper's claim of out-of-sample generalization to patients.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies a quantum-inspired machine learning framework, QCML, to predict chromosomal instability (LST status) from 112 morphological and intensity features extracted from circulating tumor cells (CTCs) in metastatic breast cancer patients. The authors compare QCML regression and a new QCML POVM classification extension against 10 classical machine learning baselines using 5-fold cross-validation with 5 repeats, reporting that QCML achieves the highest out-of-sample balanced accuracy (70%) and ROC AUC (0.763) among all models. They also present a gradient-based feature importance analysis and a QCML-distance visualization. The paper concludes that QCML is a promising tool for liquid-biopsy-based real-time detection of chromosomal instability.","tokens_in":17395,"tokens_out":4276,"duration_ms":46223,"significance":"If the reported superiority were statistically reliable and generalizable to new patients, the paper would make a valuable contribution to high-dimensional, low-sample-size biomedical classification and to the emerging field of quantum-inspired machine learning. The manuscript is clearly written, the biological feature-importance findings are plausible and connect to known mechanisms of chromosomal instability, and the QCML POVM extension is described in detail. However, the central empirical claim rests on an evaluation protocol that does not establish generalization to unseen patients, and the reported differences are not supported by statistical testing. These issues are load-bearing for the paper's main conclusion.","major_comments":[{"comment":"The cross-validation procedure treats each CTC as an independent observation and splits cells, not patients. The dataset contains 227 CTCs from 73 patients (mean 3.25 cells per patient), so the same patient's cells appear in both the training and test folds. Cells from the same patient share biological state, staining batch, imaging conditions, and other technical factors, and are therefore not independent. The abstract and Discussion frame the method as enabling real-time detection for patients, so the intended use case is prediction on new patients. The current 'case-agnostic' protocol measures performance on cells with same-patient cells in the training set, which can inflate out-of-sample metrics and does not establish generalization to new patients. The authors must re-run the comparison using patient-level or group-aware cross-validation and report whether the QCML advantage persists. Without this, the headline superiority claim is not supported.","section":"Section 2.2, Tables 1 and 2"},{"comment":"The reported performance differences lack statistical support. In Table 1, QCML's out-of-sample balanced accuracy is 70% ± 8% versus 68% ± 7% for the SVM Gaussian kernel, a difference well within the reported standard deviations. In Table 2, the QCML POVM ROC AUC of 0.763 versus XGBoost's 0.747 is presented without confidence intervals, per-fold values, or a significance test (e.g., DeLong test, bootstrap, or paired test across folds). The claim that QCML 'outperforms' the classical baselines requires formal uncertainty quantification and hypothesis tests, especially given the small absolute differences and overlapping error bars.","section":"Tables 1 and 2"},{"comment":"The claim that QCML shows a 'smaller disconnect between in-sample and out-of-sample performance' is made qualitatively. For balanced accuracy, the in-sample to out-of-sample drop for QCML is 78% to 70%, while for SVM Gaussian it is 74% to 68%; these gaps are similar in magnitude. If the authors wish to argue that QCML generalizes better, they should quantify the generalization gap across folds and test whether the difference is statistically significant, rather than relying on visual or qualitative comparison.","section":"Section 3.2, Table 1"}],"minor_comments":[{"comment":"The phrase 'LST¿12' appears to be a typographical error for 'LST > 12'; please correct it.","section":"Section 2.3"},{"comment":"The description of the 'gradient-free' loss reweighting is unclear. The text says that adjusting each loss component by the corresponding gradient-free loss 'forces each component's loss to be equal to 1,' but in practice the gradient is scaled by the detached loss value, so the effective weighting is not necessarily equal. Please clarify the intended effect and the implementation.","section":"Section 2.3, Equation (2)"},{"comment":"The caption states that at lower specificity QCML POVM outperforms both XGBoost and Random Forest, but this is a qualitative claim based on visual inspection. Please provide the actual ROC curves with confidence bands or numerical comparisons at selected operating points.","section":"Figure 4"},{"comment":"The feature importance table is difficult to read in the current formatting. Please present it as a proper table with clear column headers, and ensure the feature names and importance scores are legible.","section":"Section 3.4, Figure 6"},{"comment":"The phrase 'out of sample' should be hyphenated as 'out-of-sample' for consistency with standard terminology.","section":"Abstract and throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant clinical problem and presents an interesting application of quantum-inspired methods, but the central comparative claim is currently not supported by the evaluation design or statistical analysis. The authors should be required to perform patient-level cross-validation and provide proper uncertainty quantification before the manuscript can be considered for publication. I also note that the method and its theoretical support come largely from the authors' own prior work, so an independent validation or application by another group would strengthen the claims substantially."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a real application of a quantum-inspired ML method to a clinically relevant problem, with a genuine formal addition (the POVM extension) and some biologically plausible feature findings. But the headline claim that QCML outperforms classical ML on out-of-sample CTCs is not established by the reported evaluation.\n\nThe POVM construction for full predictive densities is a reasonable extension of the authors' prior QCML work, and the paper is the first to use QCML for morphology-based LST prediction. The feature importance results—DAPI cytoplasmic signal, fractal features, cell size—line up with the cited biology, which gives the model some external credibility. The data come from a real clinical platform, and the authors are transparent that this is a small, preliminary study.\n\nThe main problem is the evaluation protocol. Section 2.2 states a \"case-agnostic approach\": cells are treated as independent and folds split cells, not patients. With 3.25 cells per patient on average, test cells often have same-patient cells in the training set. That means the reported out-of-sample scores reflect partial patient memorization, not generalization to unseen patients. Since the abstract frames this as enabling real-time detection for patients, the missing patient-level analysis is a load-bearing gap.\n\nEven under the reported split, the QCML advantage is within the shown standard deviations: 70% vs 68% balanced accuracy and AUC 0.763 vs 0.747, with no paired significance test. This is fixable without a huge rewrite—a grouped or leave-one-patient-out CV plus a paired test would settle it. The lack of an external validation cohort is a minor additional caveat, acceptable for a pilot.\n\nThis paper deserves serious refereeing, not a desk reject: the method is formally grounded, the application is legitimate, and the evaluation flaw is correctable. I would send it out but expect major revision, asking specifically for patient-level CV and significance testing. I would not cite it in my own work until that is done.","headline":"A legitimate first application of QCML to CTC morphology with a sound POVM extension, but the cell-level cross-validation does not support the new-patient generalization claim.","tokens_in":18003,"tokens_out":1691,"would_cite":false,"duration_ms":18964,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A quantum-inspired machine learning model predicts chromosomal instability in circulating tumor cells from microscopy features more accurately than ten classical methods.","keywords":["Quantum Cognition Machine Learning","Circulating tumor cells","Chromosomal instability","Liquid biopsy","pLST prediction","Positive Operator-Valued Measure","Metastatic breast cancer","High-dimensional low-sample-size classification"],"falsifier":"Run the same 5-fold cross-validation comparison with all cells from each patient held out together (patient-level splits), or on an independent cohort of CTCs; if QCML's balanced-accuracy and ROC-AUC advantages shrink or vanish, the claimed generalization advantage is an artifact of shared-patient overlap.","tokens_in":16921,"feed_emoji":"🧬","tokens_out":9573,"duration_ms":86890,"temperature":0.7,"pith_summary":"The paper claims that Quantum Cognition Machine Learning (QCML) predicts which circulating tumor cells carry chromosomal instability — measured by large-scale state transitions (LST) — from microscopy-derived morphology better than ten classical machine-learning baselines. In a dataset of 227 CTCs from 73 metastatic breast cancer patients, QCML reports the highest out-of-sample specificity (57%) with sensitivity 84% and balanced accuracy 70%, and its POVM variant reports the highest ROC AUC (0.763). If correct, the finding matters because morphology-only prediction could flag high-metastatic-potential CTCs in hours rather than after whole-genome sequencing. The paper also argues that QCML's learned feature importance aligns with known biology of unstable cells, such as larger cells, fractal nuclear complexity, and abnormal cytoplasmic DAPI signal.","feed_headline":"Quantum-inspired ML beats 10 baselines at spotting unstable CTCs","feed_subtitle":"It reports the top balanced accuracy and ROC AUC on 227 tumor cells, suggesting faster biopsy readouts.","key_machinery":"The central object is the error Hamiltonian $H(x_t) = \\frac{1}{2}\\sum_k (A_k - x_{t,k} I)^2$, where each $A_k$ is a learned Hermitian feature observable on an $N$-dimensional Hilbert space and $x_t$ is a cell's feature vector. The map from data to quantum state is the ground state $|\\psi_t\\rangle$ of this Hamiltonian, and a forecast observable $B$ gives predictions through $\\hat{y}_t = \\langle \\psi_t | B | \\psi_t \\rangle$. The QCML POVM extension replaces point forecasts with a full probability density $p(y) = \\langle \\psi | \\hat{Y}^\\dagger(y) \\hat{Y}(y) | \\psi \\rangle$, with the operators $\\hat{Y}(y)$ parameterized by Legendre polynomials so the density integrates to one. This representation carries the argument: the parameter count grows linearly with the number of features, and the ground-state map performs an intrinsic dimensionality reduction, which the paper claims is why QCML generalizes in a high-dimensional, low-sample-size regime.","core_discovery":"On the paper's own terms, the central discovery is that encoding each cell's 112 morphology, intensity, and texture features as a quantum state — the ground state of an error Hamiltonian built from learned observables — yields better out-of-sample classification of LST status than support vector machines, neural networks, random forests, XGBoost, and other classical models. In verification folds QCML reached balanced accuracy 70% ± 8%, sensitivity 84% ± 9%, and specificity 57% ± 10%, and QCML POVM reached ROC AUC 0.763, the best among all models tested. The authors interpret the smaller gap between in-sample and out-of-sample performance as evidence that the quantum-state representation controls overfitting without curated feature selection. They further claim that the model's feature importance ranking recovers biologically plausible correlates of chromosomal instability, including cell size, fractal nuclear shape, cross-channel colocalization, and DAPI localization.","pith_inferences":["Editorial extension: Because the reported cross-validation splits cells rather than patients, the paper's claim of generalization to new patients is not yet fully tested; a patient-grouped split or independent cohort would be the decisive check.","Editorial extension: The QCML distance defined by quantum fidelity is a supervised similarity measure; a natural next step the paper does not take is to use it for patient-level clustering or survival analysis in metastatic breast cancer.","Editorial extension: The paper mentions quantum-assisted feature mapping from raw pixels as future work; that same framework could be tested against the hand-crafted 112-feature pipeline used here.","Editorial extension: The POVM densities should be checked for calibration, since threshold-based clinical use requires trustworthy probability forecasts; the paper does not report calibration curves."],"forward_implications":["Morphology-only pLST prediction could cut liquid-biopsy turnaround time from sequencing-scale workflows to image-analysis-scale workflows, if the reported performance holds in validation.","QCML's gradient-based feature ranking yields a testable biological hypothesis: unstable CTCs are larger, have more fractal nuclei, and show disrupted DAPI localization and cross-channel texture alignment.","QCML POVM's full predictive densities allow clinicians to choose operating points on specificity/sensitivity and report quantiles rather than single point predictions.","A smaller in-sample/out-of-sample performance gap than classical models suggests QCML may generalize in other high-dimensional, low-sample-size biomedical datasets without curated feature selection."],"supporting_citations":[{"why":"Establishes morphology-predicted LST as a chromosomal-instability biomarker and provides the clinical rationale.","marker":"[3]"},{"why":"Supplies the CTC assay, genomic LST ground truth, and the LST > 12 cutoff used for labels.","marker":"[4]"},{"why":"Provides the automated CTC segmentation and the 112-feature extraction pipeline that defines the input space.","marker":"[30]"},{"why":"Introduces QCML and its dimensional-reduction/generalization properties that the paper builds on.","marker":"[25]"},{"why":"Gives the supervised QCML formulation used here for forecasting a target observable.","marker":"[28]"},{"why":"Supplies the quantum-cognition background motivating state-vector representations of data.","marker":"[34]"},{"why":"Provides the quantum measurement and POVM formalism underlying QCML POVM.","marker":"[35]"},{"why":"Defines the low-pass sequencing approach used to estimate LST from CTC genomes.","marker":"[6]"}],"fun_headline_variants":["Quantum-inspired AI beats 10 baselines in unstable CTC prediction","Quantum-cognition ML bests classical models on CTC instability","Quantum-inspired model outperforms 10 classical ML baselines","Quantum-cognition AI predicts unstable CTCs better than classical models","QCML forecasts chromosomal instability better than standard ML"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that splitting individual cells, not patients, into training and test sets measures how well the model generalizes to new patients; cells from the same patient share biological and technical factors that can inflate apparent out-of-sample performance.","fun_headline_variants_meta":{"raw":{"variants":["Quantum-inspired AI beats 10 baselines in unstable CTC prediction","Quantum-cognition ML bests classical models on CTC instability","Quantum-inspired model outperforms 10 classical ML baselines","Quantum-cognition AI predicts unstable CTCs better than classical models","QCML forecasts chromosomal instability better than standard ML"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000624,"raw_usage":{"total_tokens":2884,"prompt_tokens":934,"completion_tokens":1950,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":1869}},"tokens_in":550,"tokens_out":1950,"duration_ms":14107,"temperature":1.0,"reasoning_tokens":1869,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:35:12.720606+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 5-fold cross-validation comparison with all cells from each patient held out together (patient-level splits), or on an independent cohort of CTCs; if QCML's balanced-accuracy and ROC-AUC advantages shrink or vanish, the claimed generalization advantage is an artifact of shared-patient overlap.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes morphology-predicted LST as a chromosomal-instability biomarker and provides the clinical rationale."},{"cited_title":"A novel liquid biopsy assay for detection of erbb2 (her2) amplification in circulating tumor cells (ctcs),","cited_arxiv_id":null,"evidence_quote":"Supplies the CTC assay, genomic LST ground truth, and the LST > 12 cutoff used for labels."},{"cited_title":"Fully automated ctc detection, segmentation and classification for multi-channel if imaging,","cited_arxiv_id":null,"evidence_quote":"Provides the automated CTC segmentation and the 112-feature extraction pipeline that defines the input space."},{"cited_title":"Robust estimation of the intrinsic dimension of data sets with quantum cognition machine learning","cited_arxiv_id":"2409.12805","evidence_quote":"Introduces QCML and its dimensional-reduction/generalization properties that the paper builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the quantum measurement and POVM formalism underlying QCML POVM."},{"cited_title":"Chromosomal instability estimation based on sequencing,","cited_arxiv_id":null,"evidence_quote":"Defines the low-pass sequencing approach used to estimate LST from CTC genomes."}],"review_version":1}