A quantum inspired predictor of Parkinsons disease built on a diverse, multimodal dataset

Anya Iyer; Ashwin A. Nair; Diya Vatsavai

REVIEW 4 major objections 5 minor 3 references

A quantum inspired predictor of Parkinsons disease built on a diverse, multimodal dataset

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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…

desk verdict Standard angle-embedding qSVM, new application, but the 90% accuracy is not trustworthy because feature selection leaks test information. read the letter →

arxiv 2411.18640 v1 pith:3XTQDNOM submitted 2024-11-25 q-bio.QM cs.LG

Diya Vatsavai , Anya Iyer , Ashwin A. Nair This is my paper · ORCID

classification q-bio.QMcs.LG

keywords Parkinson'sdiseasequantummachinelearningsupportvectorangleembeddingmultimodalbiomarkersmPowerdatasetrandomforestfeatureselectionsmartphonescreening

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 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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

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

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.

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.

Desk Editor's Note

read the 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.

What 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.

The 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.

The 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.

Is 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.

Who 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.

Recommendation: 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.

Referee Report

4 major / 5 minor

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.

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 (4)

[Feature Selection / Data Description] 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.
[Evaluation and Comparative Analysis] 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.
[Data Description / Limitations] 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.
[qSVM Model Architecture] 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.

minor comments (5)

[Abstract / Introduction] 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.
[Abstract / Data Description] 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.
[References] The reference list contains duplicate numbering: two entries are labeled [41] (Kariya et al. and the PennyLane documentation), which will confuse readers tracing citations.
[Feature Selection] 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.
[Benchmark Models] 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.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported 90%/0.98 is partly forced: random-forest feature selection and weighting are fit on the full 194-participant dataframe, so the 30 test rows leak into the kernel before evaluation.

fitted input called prediction [Results, Feature Selection (pp. 5–7) and qSVM Model Architecture (p. 15)]
"We normalized the dataframe using Scikit Learn’s StandardScaler to ensure a consistent magnitude for each feature. Then, we trained a baseline Random Forest model to identify the top-performing features for the final qSVM model, selecting features with importance values above the 80th percentile [21] (Figure 1). ... For input into the proposed qSVM model, which is highly sensitive to feature ordering and magnitude [22], we multiplied each feature by its importance."

The Random Forest feature importances are fitted parameters. The paper describes the split into 164 training and 30 test samples in the preceding paragraph, but the Feature Selection paragraph operates on 'the dataframe' with no statement that the scaler or Random Forest is restricted to the training rows. The qSVM kernel then multiplies each feature by its importance, so the chosen feature mask and kernel weights depend on the PD labels of the 30 test rows. The reported 90% accuracy and 0.98 AUC are computed on those same 30 rows, making the headline 'prediction' a fitted evaluation: the test labels helped construct the very kernel that is then said to predict them.

full rationale

The central methodological circularity is target leakage in feature selection. The Random Forest importance values that select the feature set and set the qSVM kernel weights are computed on the full 194-participant dataframe, before any explicit split of the feature-selection procedure, so the 30 test rows influence the model. The qSVM then reports accuracy and AUC on those same 30 rows, so the headline numbers are partially forced by the evaluation set itself, not an out-of-sample prediction. The paper's own Limitations section concedes only a single train/test split and no cross-validation, which is consistent with this fragility; a later sentence saying the model was trained 'on our dataframe of 194 samples' further compounds the issue if read literally. However, the proposed angle-embedding kernel is a standard, externally documented construction (PennyLane AngleEmbedding) with independent content; no load-bearing self-citation or imported uniqueness theorem is present, and the kernel is not defined in terms of the labels. The Z/ZZ benchmark qSVMs were also evaluated on a different 45-sample subset, so the comparative 'surpassing' claim is not on a common test set. For these reasons the circularity is partial and localized to the evaluation protocol, not a wholesale derivation-by-definition. Score 6.

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

The central performance claim rests on several choices and premises: an unseeded 164/30 split, an 80th-percentile RF feature-selection threshold, RF importances that double as kernel weights, a scaling factor of 10, PCA-reduced MFCCs to 10 components, and the assumption that normalization and feature selection do not use test data. The mPower self-reported labels and fairness of the Z/ZZ subset comparison are additional load-bearing assumptions.

free parameters (5)

80th percentile random forest feature importance threshold = not reported
Hand-chosen threshold determines how many of 64 features enter the qSVM; affects all reported metrics and was chosen after seeing the full dataset.
random forest feature importance weights = not reported
Fitted on the full 194-sample dataframe and used both to select features and as multiplicative weights in the kernel's softmax aggregation.
feature scaling factor of 10 = 10
Hand-selected to bring scaled features close to magnitude 1; changes kernel values and therefore the reported accuracy.
number of MFCC components after PCA = 10
Choice reduces voice feature dimensionality; no stated criterion for 10 components.
train/test split assignment = 164/30 balanced, random, no seed
Random split with no seed or repeated runs; metric variance on 30 test samples is large.

assumptions (5)

domain assumption mPower participant-reported Parkinson's status is an accurate ground-truth label.
The mPower cohort labels come from self-reported diagnosis via ResearchKit; the paper treats these labels as ground truth without clinical adjudication.
domain assumption Random forest feature importances computed on the full dataframe are unbiased and can be used to select and weight features for the qSVM.
This is violated if test samples were included, and the text in the Feature Selection section suggests normalization and RF training happened on all 194 participants.
domain assumption StandardScaler normalization and feature selection use only training-set statistics.
The paper states it normalized the dataframe before splitting; a load-bearing premise for unbiased evaluation.
domain assumption A classical simulation of the angle-embedding fidelity kernel captures the same classification behavior as a true quantum kernel.
The paper relies on this to claim 'quantum inspired' value from a simulatable architecture; no experiment on quantum hardware is provided.
domain assumption The benchmark comparison is fair, including qSVM Z/ZZ results computed on a different subset.
Methods section states Z and ZZ feature maps were evaluated on the first 30 train and 15 test samples, not the 164/30 split; the comparison table presents them as comparable.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A quantum inspired predictor of Parkinsons disease built on a diverse, multimodal dataset." pith.science (2026). https://pith.science/paper/3XTQDNOM

@misc{pith2026241118640,
  author       = {Pith},
  title        = {Pith review of: A quantum inspired predictor of Parkinsons disease built on a diverse, multimodal dataset},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3XTQDNOM}},
  note         = {Machine review of arXiv:2411.18640}
}

read the original abstract

Parkinsons disease, the fastest growing neurodegenerative disorder globally, has seen a 50 percent increase in cases within just two years. As speech, memory, and motor symptoms worsen over time, early diagnosis is crucial for preserving patients quality of life. While machine-learning-based detection has shown promise, relying on a single feature for classification can be error-prone due to the variability of symptoms between patients. To address this limitation we utilized the mPower database, which includes 150,000 samples across four key biomarkers: voice, gait, tapping, and demographic data. From these measurements, we extracted 64 features and trained a baseline Random Forest model to select the features above the 80th percentile. For classification, we designed a simulatable quantum support vector machine (qSVM) that detects high-dimensional patterns, leveraging recent advancements in quantum machine learning. With a novel, simulatable architecture that can be run on standard hardware rather than resource-intensive quantum computers, our model achieves an accuracy of 90 percent and an AUC of 0.98, surpassing benchmark models. By utilizing an innovative classification framework built on a diverse set of features, our model offers a pathway for accessible global Parkinsons screening.

Figures

Figures reproduced from arXiv: 2411.18640 by the authors.

**Figure 2.** Radar chart for proposed model and benchmark comparison across accuracy, precision, recall, [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** Example qubit encoding of one feature of the data [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages

[11]

Adams, W. R. Bradykinesia in Parkinson’s disease. In Diagnosis and Management in Parkinson’s Disease, Elsevier; https://doi.org/10.1016/B978-0-12-818042-0.00010-9 (2020). [12] Hopkins Medicine. Assessment of bradykinesia in Parkinson’s disease using video-based pose estimation. Johns Hopkins Clinical Connection; https://clinicalconnection.hopkinsmedicine....

work page arXiv 2020
[22]

Solenov, D., Brieler, J., & Scherrer, J. F. (2018). The Potential of Quantum Computing and Machine Learning to Advance Clinical Research and Change the Practice of Medicine. Missouri Medicine, 115(5), 463–467. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6205278/ (2018). [23] Synapse. mPower Public Researcher Portal. mPower Mobile Parkinson Disease Study;...

work page Pith review arXiv 2018
[33]

Progression of voice and speech impairment in the course of Parkinson's disease: A longitudinal study

Skodda, S., Grönheit, W., Mancinelli, N., & Schlegel, U. Progression of voice and speech impairment in the course of Parkinson's disease: A longitudinal study. *Parkinson's Disease*. 2013, Article ID 389195; https://doi.org/10.1155/2013/389195 (2013). [34] Cernak, M., Orozco-Arroyave, J. R., Rudzicz, F., Christensen, H., Vásquez-Correa, J. C., & Nöth, E. ...

arXiv 2013

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [11]

Adams, W. R. Bradykinesia in Parkinson’s disease. In Diagnosis and Management in Parkinson’s Disease, Elsevier; https://doi.org/10.1016/B978-0-12-818042-0.00010-9 (2020). [12] Hopkins Medicine. Assessment of bradykinesia in Parkinson’s disease using video-based pose estimation. Johns Hopkins Clinical Connection; https://clinicalconnection.hopkinsmedicine....

work page arXiv 2020

[2] [22]

Solenov, D., Brieler, J., & Scherrer, J. F. (2018). The Potential of Quantum Computing and Machine Learning to Advance Clinical Research and Change the Practice of Medicine. Missouri Medicine, 115(5), 463–467. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6205278/ (2018). [23] Synapse. mPower Public Researcher Portal. mPower Mobile Parkinson Disease Study;...

work page Pith review arXiv 2018

[3] [33]

Progression of voice and speech impairment in the course of Parkinson's disease: A longitudinal study

Skodda, S., Grönheit, W., Mancinelli, N., & Schlegel, U. Progression of voice and speech impairment in the course of Parkinson's disease: A longitudinal study. *Parkinson's Disease*. 2013, Article ID 389195; https://doi.org/10.1155/2013/389195 (2013). [34] Cernak, M., Orozco-Arroyave, J. R., Rudzicz, F., Christensen, H., Vásquez-Correa, J. C., & Nöth, E. ...

arXiv 2013