{"id":"5bbff888-af73-4aeb-bbba-80df7e2784fd","arxiv_id":"2411.18640","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A quantum-inspired angle-embedding SVM trained on 194 mPower participants reports 90% accuracy and 0.98 AUC for Parkinson's screening, though the evaluation has statistical and comparison flaws.","lead":"This paper builds a Parkinson's screening model from smartphone voice, gait, tapping, and demographic data, using a quantum-inspired kernel that runs on classical hardware. The model reports 90% accuracy and 0.98 AUC, but on a 30-person test set with some benchmark comparisons made on different data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 90%/0.98 is not trustworthy because normalization and random-forest feature selection appear to be fit on the full dataframe before the 164/30 split, so the 30 test rows leak into feature choice and feature weights.","rationale":"I agree with the reader's weakest assumption: the preprocessing pipeline leaks test information into feature selection and feature weighting. This is load-bearing because the qSVM does not use raw features directly; it uses random-forest importance values to scale, order, and weight every feature in the kernel. If those importance values are estimated on the full 194-row dataframe, then the test participants have already voted on the model's inductive bias, so the reported 90% and 0.98 are not honest estimates of generalization. The single 30-sample split compounds the problem: even without leakage, the variance is large, and the paper's own Limitations section acknowledges the lack of cross-validation. I also note that the benchmark comparison is weaker because the Z and ZZ qSVMs were evaluated on a different subset. The proposed angle-embedding kernel is a legitimate, classically simulatable construction; the issue is entirely in the evaluation protocol. The reader's REJECT verdict remains appropriate because the central claim as stated is not established. A clean refit with train-only preprocessing is the decisive check.","tokens_in":9610,"tokens_out":5642,"duration_ms":53110,"concrete_test":"Re-run the pipeline from scratch with the preprocessing isolated: fit StandardScaler and the random-forest feature selector on the 164 training rows only; apply the fitted scaler and the resulting feature mask to the 30 held-out rows; then train the qSVM kernel on the 164 training rows and report test accuracy and AUC with the same benchmark models evaluated on the same 30-row test set. If the 90% accuracy and 0.98 AUC survive this protocol, the leakage concern is resolved; if they decline materially, the headline result is inflated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the reported 90% accuracy and 0.98 AUC measure generalization to unseen participants. The Feature Selection section says: 'We normalized the dataframe using Scikit Learn's StandardScaler' and then 'trained a baseline Random Forest model ... selecting features with importance values above the 80th percentile.' No train/test split is mentioned before these steps. If the scaler and the random forest are fit on all 194 participants, the 30 test rows influence which features are kept and how the qSVM kernel weights those features. Because random forest feature importance is label-dependent, this is target leakage: the feature mask and the weighted-kernel coefficients are chosen using test labels. The qSVM section then multiplies each feature by its importance, so the leakage directly enters the classifier's kernel. The paper's own Limitations section concedes only a single train/test split and no cross-validation, and with n=30 a two-sample error changes accuracy by ~6.7 points. The 'surpassing benchmark models' statement is also compromised because the Z/ZZ feature-map qSVMs were evaluated on a different 45-sample subset. None of this impugns the angle-embedding kernel construction itself, which is classically simulatable; the problem is that the evaluation protocol does not establish unbiased generalization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a multimodal Parkinson's disease classifier built on the mPower dataset, combining voice, gait, tapping, and demographic features. After extracting 64 features, the authors train a random forest to select features above the 80th percentile of importance, then weight each feature by its importance and feed them into a classically simulatable quantum support vector machine (qSVM) based on PennyLane angle embedding with Y-rotation gates. The model is reported to achieve 90% accuracy and 0.98 AUC on a 30-sample test set, outperforming classical ML models and other qSVMs. The paper claims this is a novel, simulatable architecture that can run on standard hardware rather than quantum computers.","tokens_in":104,"tokens_out":5896,"duration_ms":133452,"significance":"The idea of using a classically simulatable angle-embedding kernel for multimodal biomedical data is timely, and the paper's ambition to evaluate on a public, multimodal dataset is commendable. The authors also provide a clear conceptual explanation of how the kernel may be simulated classically and reference a public data source. However, the evaluation methodology contains a serious data-leakage flaw and the benchmark comparisons are not on equal footing; as presented, the evidence does not support the headline accuracy and AUC claims. If the experiments were redone with a valid protocol, the underlying kernel idea might still merit consideration, but the current manuscript does not establish its central claim.","major_comments":[{"comment":"The feature selection and normalization are performed on the full 194-participant dataframe before any train/test split is applied to those steps. The 'Data Description' section states that the dataset is divided into 164 training and 30 test samples, but the 'Feature Selection' section then says 'We normalized the dataframe using Scikit Learn's StandardScaler' and 'trained a baseline Random Forest model... selecting features with importance values above the 80th percentile' with no indication that these operations are restricted to the 164 training samples. The selected feature mask and the feature-importance weights are subsequently used to construct the qSVM kernel, as described in the Feature Selection and qSVM Model Architecture sections. Because random forest importance is label-dependent, the 30 test labels influence which features are kept and how the kernel is weighted. The reported 90% accuracy and 0.98 AUC therefore do not measure generalization to unseen participants.","section":"Feature Selection / Data Description"},{"comment":"The Z and ZZ feature-map qSVM benchmarks are evaluated on a different subset from the proposed model and the classical baselines. The manuscript explicitly says: 'for the alternative qSVM kernels of the Z and ZZ feature map, the full dataset was too resource-intensive to run. So, we chose to extract metrics based on a subset of the dataset including the first 30 train and 15 test samples.' This means the accuracy and AUC values in Table 1 for the Z and ZZ feature maps were not computed on the same 164/30 split used for the proposed model and for the classical ML models. Consequently, the claim of 'surpassing benchmark models' for those qSVMs is not supported by the reported comparisons.","section":"Evaluation and Comparative Analysis"},{"comment":"The test set contains only 30 samples, and the evaluation is a single train/test split with no cross-validation, repeated splits, or confidence intervals. The manuscript's own Limitations section concedes: 'With only a train and test dataset, there is a possibility of the model overfitting to the validation set.' With n=30, a one-sample error changes the accuracy by approximately 3.3 percentage points, and the standard error of the accuracy is around 5.5 percentage points. The observed differences between the proposed model (90%) and the classical baselines (e.g., linear SVM at 77%) are therefore not shown to be statistically meaningful, and any interpretation is further undermined by the leakage described in the first major comment.","section":"Data Description / Limitations"},{"comment":"The description of the proposed quantum kernel contains ambiguous or corrupted notation that prevents reproducibility. For example, the text states 'Since RX(!2) is equivalent to √𝑥' and 'the kernel can be programmed as RZ(−!2), √𝑥, RZ(x1 - x2), √𝑥 and RZ(−!2)', but the symbols '!2' and '√𝑥' are not defined. It also says that measurements are 'aggregated through a weighted sum of each measurement multiplied by the random forest feature importance of the corresponding feature, transformed using a softmax function', which does not specify whether the softmax is applied to the importances before summation or to the final kernel value. Because the novelty claim rests on this architecture, the method section needs to be self-contained and mathematically precise.","section":"qSVM Model Architecture"}],"minor_comments":[{"comment":"The abstract states a '50 percent increase in cases within just two years', while the introduction reports that PD prevalence has doubled over 25 years; please reconcile these figures or provide a specific citation for the two-year statistic.","section":"Abstract / Introduction"},{"comment":"The abstract mentions '150,000 samples' but the analysis uses 194 participants with one trial per activity per participant; the manuscript should clarify the distinction between raw mPower measurements and the final participant-level dataset.","section":"Abstract / Data Description"},{"comment":"The reference list contains duplicate numbering: two entries are labeled [41] (Kariya et al. and the PennyLane documentation), which will confuse readers tracing citations.","section":"References"},{"comment":"The paper states that race was excluded to mitigate bias, yet the diversity claim in the abstract emphasizes racial diversity; please clarify that the dataset was diverse in recruitment, but race was not used as a model feature.","section":"Feature Selection"},{"comment":"For reproducibility, the exact hyperparameters and software versions of the benchmark models should be listed; the current text only says 'the same training and testing sets' were used for most models.","section":"Benchmark Models"}],"recommendation":"reject","confidential_remarks":"The authors have chosen a relevant topic and a publicly available dataset, and the idea of a classically simulatable angle-embedding kernel is interesting. However, the central empirical claim is invalidated by the evident data leakage in feature selection/weighting and by the unequal benchmark protocol. The limitations section candidly acknowledges the lack of cross-validation, but it does not address the more serious problem that preprocessing and feature selection were performed on the full dataframe, including the test rows. In my view, this is not a matter of local revision; the experimental evaluation would need to be redone from the ground up, and the results would likely change. I cannot support publication of the current version, though a substantially revised and re-analyzed follow-up could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a standard angle-embedding qSVM applied to a multimodal Parkinson's dataset from mPower. The application is new, the method is not. The headline 90%/0.98 is not believable as reported because the preprocessing pipeline leaks test information into feature selection.\n\nWhat the paper does well: It uses a public, reasonably diverse dataset with voice, gait, tapping, and demographics. The writing is clear, and the authors are upfront in the Limitations section that they only used a single train/test split and no cross-validation. The decision to use a classically simulatable angle-embedding kernel instead of entanglement-heavy ZZ feature maps is a sensible practical choice, and the kernel construction itself is standard and correct.\n\nThe soft spots are structural. The Feature Selection section says they normalized the dataframe and trained a random forest on it, then selected features above the 80th percentile. No split is mentioned before this step. The train/test split is described earlier in the Results, but the order of operations in the text suggests the scaler and the random forest see all 194 samples, including the 30 test participants. Since those 30 rows are used to pick features and to compute the importance weights that directly multiply the kernel features, the test labels leak into the model. That alone invalidates the reported generalization metrics. The same section shows the feature importances in Figure 1, which are computed on the full dataframe. Even if the split happened before, the paper never says so.\n\nThe evaluation also rests on a single 30-sample test set. With n=30, a two-sample shift changes accuracy by ~7 points. No confidence intervals, no repeated splits. The benchmark qSVMs (Z and ZZ feature maps) were run on a different, smaller 45-sample subset, so the claim of outperforming them is not supported. The classical ML benchmarks appear to use the same split, so those comparisons are fair, but they are secondary.\n\nIs there anything to salvage? The feature engineering is thoughtful, and the idea of weighting kernel inputs by a trained feature importance could be useful if the weighting were learned on the training set only. But as submitted, the central claim is not established.\n\nWho is this for? Someone working on quantum kernel applications might skim it, but the methodological flaw means I would not trust the numbers. It could be a good teaching example of target leakage.\n\nRecommendation: If I were editor, I would not send this to peer review as is. A revision that fixes the preprocessing order, adds cross-validation or bootstrap CIs, and reruns the qSVM benchmarks on the same split could make it a legitimate contribution, but it would need those changes first.","headline":"Standard angle-embedding qSVM, new application, but the 90% accuracy is not trustworthy because feature selection leaks test information.","tokens_in":10412,"tokens_out":3740,"would_cite":false,"duration_ms":32038,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A quantum-inspired kernel machine simulated on ordinary hardware predicts Parkinson's disease from smartphone-measured voice, gait, tapping, and demographic features with 90 percent accuracy and 0.98 AUC, beating the classical and quantum…","keywords":["Parkinson's disease","quantum machine learning","quantum support vector machine","angle embedding","multimodal biomarkers","mPower dataset","random forest feature selection","smartphone screening"],"falsifier":"Recompute the full pipeline with feature selection and scaling performed only within training folds, and evaluate on truly held-out participants; if the accuracy drops to the level of the classical baselines or below, the reported 0.90 and 0.98 scores are likely artifacts of test information leaking through the preselect fit. A second check is to rerun the Z and ZZ feature-map qSVMs on the same full 164/30 split used for the proposed model; if they match or exceed the proposed model, the claimed kernel advantage does not hold.","tokens_in":9373,"feed_emoji":"🧠","tokens_out":6336,"duration_ms":54328,"temperature":0.7,"pith_summary":"This paper tries to establish that a quantum-inspired classifier can serve as an accessible screening tool for Parkinson's disease without requiring quantum hardware. Using the mPower smartphone dataset, the authors combine 64 features from voice, gait, tapping, and demographics, select the most important features with a random forest, and feed them into a quantum support-vector machine whose kernel uses angle-embedding rotations instead of entanglement, making it simulatable on classical hardware. The reported 90 percent accuracy and 0.98 AUC exceed the classical and quantum baseline models in the paper. If true, this would mean high-accuracy PD screening could run on ordinary computers from data collected by a phone app.","feed_headline":"A quantum-inspired model predicts Parkinson's at 90%","feed_subtitle":"Combining voice, gait, tapping, and demographics, the simulated quantum SVM reports 0.90 accuracy and 0.98 AUC.","key_machinery":"The load-bearing object is a quantum kernel built from per-feature Y-axis rotations. Each qubit represents one feature, and the kernel encodes a pair of data points by rotating the qubit by $\\theta = x_1 - x_2$, so similar feature values nearly cancel and leave the qubit close to its starting state; measuring the overlap gives a similarity score. The per-qubit measurements are combined through a softmax-weighted sum using the random forest feature importances, and that weighted kernel matrix is passed to a classical SVM for classification. This design avoids entanglement, which is expensive to simulate, and the paper argues it is the reason the model outperforms both classical kernels and entanglement-heavy quantum feature maps in simulation.","core_discovery":"The central claim is that a simulatable quantum SVM whose kernel uses angle embedding—rotation gates rather than entanglement—can classify Parkinson's disease from a multimodal smartphone dataset better than standard classical models and better than the Z and ZZ feature-map qSVMs. On 194 participants split into 164 training and 30 test samples, the model reports accuracy 0.90, ROC/AUC 0.98, and precision, recall, and F1 score each 0.90, beating the best classical baseline in the comparison (linear SVM, 0.77 accuracy) and the best alternative qSVM (Z feature map, 0.87 accuracy) under this study's evaluation setup.","pith_inferences":["A testable extension would be to apply the same angle-embedding weighted-kernel recipe to other clinical multimodal datasets, replacing the random forest importances with any gradient-based feature importance measure.","Because the Z and ZZ feature-map qSVMs were evaluated on a smaller subset (30 train and 15 test samples) than the proposed model, their lower scores may reflect dataset size rather than kernel quality; refitting them on the full 164/30 split would settle which kernel truly wins.","The \"quantum-inspired\" label is doing the work of a classical kernel trick: the RY-difference rotation is mathematically close to a cosine-like kernel, so the practical contribution may be the multimodal feature engineering plus weighted kernel construction, not a quantum advantage.","If the leakage concern about fit-before-split is confirmed, a properly validated classical SVM with the same feature set and weights might reach comparable accuracy, making the multimodal feature set itself the more durable contribution."],"forward_implications":["If the reported results hold, clinicians could screen for Parkinson's disease from a 20-second tapping test, a 10-second \"ahh\" recording, and phone accelerometer data, with no specialized equipment.","A qSVM kernel that avoids entanglement can be simulated on standard hardware, removing the main practical barrier to applying quantum machine learning in clinical settings.","Combining four biomarker modalities yields higher accuracy than single-feature classifiers, which the paper contrasts with prior single-feature results in the 60 to 85 percent range.","The model is positioned as a screening tool rather than a standalone diagnostic, so its immediate value would be in triage and early referral rather than definitive diagnosis.","The proposed kernel's sensitivity to feature ordering and magnitude is turned into a design choice: weighting and sorting features by random forest importance appears to improve performance."],"supporting_citations":[{"why":"Supplies the entire multimodal mPower dataset of smartphone voice, gait, tapping, and demographic measurements that the model is built on.","marker":"[20]"},{"why":"Provides the random-forest feature-importance threshold used to select features and also the ZZ feature-map quantum kernel used as a benchmark.","marker":"[21]"},{"why":"Defines the angle-embedding construction that maps each feature to a qubit rotation, the core of the proposed kernel.","marker":"[42]"},{"why":"Cites prior evidence that entanglement-heavy quantum feature maps perform poorly in classical simulation, motivating the no-entanglement design.","marker":"[39]"},{"why":"Supports the claim that qSVMs are sensitive to feature ordering and magnitude, which the paper exploits by sorting and scaling features.","marker":"[22]"},{"why":"A CNN voice-only benchmark whose 0.60 accuracy contrasts with the multimodal model's performance.","marker":"[18]"},{"why":"An FNN benchmark used in the comparison table, representing deep-learning alternatives.","marker":"[25]"},{"why":"Supplies the KNN and DNN benchmarks used in the comparison table.","marker":"[26]"}],"fun_headline_variants":["Simulated quantum SVM screens Parkinson's at 90% accuracy","Quantum-inspired model flags Parkinson's with 0.98 AUC","Multimodal data powers 90% accurate Parkinson's predictor","Simulatable quantum SVM hits 90% on Parkinson's detection","Parkinson's screening via quantum-inspired SVM: 90% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The feature selection and normalization are fit on the full dataset before the train/test split, so the 30 test samples may have influenced which features and scalings the model uses; if so, the reported 90 percent accuracy is not a clean estimate of performance on new people.","fun_headline_variants_meta":{"raw":{"variants":["Simulated quantum SVM screens Parkinson's at 90% accuracy","Quantum-inspired model flags Parkinson's with 0.98 AUC","Multimodal data powers 90% accurate Parkinson's predictor","Simulatable quantum SVM hits 90% on Parkinson's detection","Parkinson's screening via quantum-inspired SVM: 90% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000564,"raw_usage":{"total_tokens":2652,"prompt_tokens":897,"completion_tokens":1755,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":1667}},"tokens_in":513,"tokens_out":1755,"duration_ms":11667,"temperature":1.0,"reasoning_tokens":1667,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:52:24.761215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the full pipeline with feature selection and scaling performed only within training folds, and evaluate on truly held-out participants; if the accuracy drops to the level of the classical baselines or below, the reported 0.90 and 0.98 scores are likely artifacts of test information leaking through the preselect fit. A second check is to rerun the Z and ZZ feature-map qSVMs on the same full 164/30 split used for the proposed model; if they match or exceed the proposed model, the claimed kernel advantage does not hold.","supporting_citations":[{"cited_title":"Generating quantum feature maps for SVM classifier","cited_arxiv_id":"2207.11449","evidence_quote":"Supports the claim that qSVMs are sensitive to feature ordering and magnitude, which the paper exploits by sorting and scaling features."}],"review_version":1}