PaperlayerView exact revision on arXiv

Source-linked AI summary

Quantum Versus Classical Machine Learning for Neuropsychiatric Disorders: A Feature-Matched Multi-Cohort Evaluation

Utsav Poudel, Anuj Nepal, Vyas Viswanath, Subramaniyaswamy Vairavasundaram

arXiv:2610.21344v1cs.AI

TL;DR

Resting-state functional-connectivity classification has seen quantum models proposed, but evidence under leave-site-out validation and matched classical comparisons remains limited. This paper compares six quantum model families with matched classical models across four cohorts using identical correlation-derived features. Most quantum models match their classical counterparts, but none consistently outperforms them.

  • Problem

    Whether non-linear connectivity improves generalization to unseen acquisition sites remains open, and quantum models have rarely been evaluated against matched classical baselines under leave-site-out validation.

    Source: §I · p. 1

  • Method

    The study compares six quantum model families with matched classical models using identical correlation-derived features across four cohorts.

    Source: §I · p. 2, A. Models · p. 10

  • Results

    Most quantum models matched their classical counterparts, but none consistently outperformed them; the quantum kernel was statistically indistinguishable from a matched RBF kernel across all four cohorts.

    Source: §XII · p. 18

  • Takeaways & Limitations

    The evaluated results do not demonstrate a reproducible quantum advantage across cohorts, and fidelity-based graphs face register-width concentration and post-encoding cancellation constraints.

    Source: §XII · p. 18

    All classifiers use correlation-derived inputs, so quantum similarity between raw regional BOLD time series was not tested.

    Source: §XI · p. 17

Abstract

from arXiv · show

Resting-state functional connectivity is widely used in machine learning-based classification of neuropsychiatric disorders. Quantum machine learning models have been proposed for functional connectivity data, but few studies have evaluated them under leave-site-out validation or against matched classical models. We compared six quantum model families with classical ones using identical correlation-derived features across four public cohorts (4,308 subjects). Leave-site-out validation was used for the Autism Brain Imaging Data Exchange (ABIDE I), ADHD-200 (attention-deficit/hyperactivity disorder), and REST-meta-MDD (major depressive disorder) cohorts, while the UCLA Consortium for Neuropsychiatric Phenomics (UCLA-CNP) was evaluated using stratified 10-fold cross-validation. The quantum kernel was statistically indistinguishable from a matched classical kernel across all four cohorts, with the average accuracy difference within 1.4 percentage points at 95% confidence. On ABIDE I, a hybrid quantum convolutional network achieved similar accuracy to its classical counterpart (0.645 vs. 0.641), while a trainable quantum kernel exceeded its classical counterpart by 0.064 in area under the curve (AUC). Quantum convolutional networks also achieved higher AUC than their matched classical networks on this cohort and showed smaller train-test gaps on all four cohorts, without an accuracy gain. Overall, most quantum models matched their classical counterparts, but none consistently outperformed them. This study, among the first quantum evaluations of ABIDE I under leave-site-out validation, provides a feature-matched benchmark of quantum and classical models comprising 284 matched comparisons. We further analyse fidelity-based graph construction and account for the observed quantum-classical parity through effective-dimension and eigenspectrum analyses.

I. INTRODUCTION

The paper asks whether non-linear connectivity and quantum similarity improve generalization to unseen sites, where matched classical comparisons remain limited. It situates a feature-matched, multi-cohort benchmark within classical adaptive graph learning and quantum neuroimaging research.

I. INTRODUCTION

  • Pearson correlation captures linear pairwise dependence, leaving open whether non-linear connectivity improves generalization to unseen acquisition sites.The paper contrasts placing non-linearity in graph edges with recovering it downstream in GNNs, where repeated aggregation over dense, noisy graphs can cause over-smoothing.Source: §I · p. 1
  • The paper combines classical adaptive graph learning with quantum neuroimaging, areas that had largely developed separately and lacked matched leave-site-out comparisons.Its fidelity-initialized adaptive graph uses correlation-derived features identically across quantum and classical models, so differences reflect the similarity function rather than input representation.Source: §I · pp. 1–2
  • The study benchmarks six quantum model families against matched classical models across four cohorts, using leave-site-out validation where cohort structure permits.The authors describe this as among the first such evaluations to their knowledge.Source: §I · p. 2
  • The authors identify a systematic comparison of quantum and classical approaches, including 284 matched comparisons and a spectral analysis of quantum-fidelity and RBF kernels.The evaluation includes prediction-degeneracy screening and multiple-testing correction.Source: §I · p. 2

A. Classical Adaptive Graph Learning for Functional Connectivity

  • Classical adaptive graph learning makes connectivity learnable, with AGMGC reporting 84.65% accuracy on ABIDE and DAGMNet reporting 91.59% accuracy and 96.80% AUC on ABIDE I.AGMGC uses classically learned projections; DAGMNet fuses structural MRI, fMRI, and phenotypic data.Source: A. Classical Adaptive Graph Learning… · p. 2

B. Quantum Machine Learning in Neuroimaging and Healthcare

  • Prior quantum neuroimaging studies include circuit classifiers for time series, images, or EEG, while QG-WRN reports 68.49% accuracy on ABIDE I using a random 80/20 split.The paper notes that the split does not decouple inter-site scanner variance, and that the cited approaches do not build an inter-regional adjacency matrix.Source: B. Quantum Machine Learning in… · p. 2

C. Quantum Architectures Beyond Static Fidelity

The paper positions its static fidelity-based approach against other quantum architectures and explains the connectivity and encoding choices that frame that comparison. It also distinguishes prior network findings from model outputs.

C. Quantum Architectures Beyond Static Fidelity

  • The paper does not evaluate the quantum time-series transformer, which processes temporal dependencies directly and is identified as a natural next comparison.The cited approach is reported to target the static-overlap bottleneck without widening the register, with gains most pronounced in small samples.Source: C. Quantum Architectures Beyond Static… · p. 2

D. Positioning of This Work

  • Unlike QG-WRN, this work holds out whole acquisition sites, uses the 200-region CC200 parcellation, and matches quantum models to classical models on features, capacity, and training budget.Its architectural distinction from other quantum-neuroimaging classifiers is an explicit fidelity-initialized, classically weighted adjacency feeding a graph-attention network.Source: D. Positioning of This Work · pp. 2–3

E. Network-Level Findings Reported for These Cohorts

  • Table I and Figure 1 summarize previously reported disrupted networks, not this study’s model results; near-chance performance also limits network-level interpretation.The figure colors Yeo networks by their rank in the literature; the UCLA-CNP panel shows the CC200/Yeo parcellation.Source: E. Network-Level Findings Reported for… · p. 3

A. Graph Representation of Functional Connectivity

  • A functional-connectivity graph represents brain regions as nodes and their connections as weighted edges; Pearson correlation is a common adjacency measure but primarily captures linear, zero-lag, pairwise dependence.It may miss non-linear or higher-order relationships involving multiple regions.Source: A. Graph Representation of Functional… · p. 3

B. Quantum Feature Maps and Hilbert Spaces

  • Angle encoding maps each feature to a qubit’s Y-rotation angle, while CNOT gates couple qubits so the joint state cannot be factored into independent single-qubit descriptions.The paper scales features before encoding because rotations are periodic, and adds CNOT gates to capture dependencies between features.Source: B. Quantum Feature Maps and Hilbert… · p. 3

C. Quantum Fidelity as a Connectivity Metric

Quantum-state fidelity supplies a non-linear similarity for selecting candidate graph connections. The graph-attention mechanism then learns edge weights over that selected support.

C. Quantum Fidelity as a Connectivity Metric

  • Fidelity is the squared overlap between two encoded states and is equivalent to a quantum kernel that defines a non-linear similarity over region features.The paper compares its relationship to linear and RBF kernels through matched performance, effective dimension, and eigenspectrum analyses.Source: C. Quantum Fidelity as a Connectivity… · p. 3
  • In high-dimensional Hilbert spaces, concentration can make fidelity graphs sparser than dense correlation graphs, but it may suppress informative relationships as well as noise.Sparsity alone does not establish denoising.Source: C. Quantum Fidelity as a Connectivity… · p. 3

D. Adaptive Graph Attention Mechanism

  • Fidelity selects candidate edges without diagnostic labels, while a graph attention network learns their weights; in the principal experiments, the selected support stays fixed during training.The attention combines latent-vector similarity, a learned pairwise function, and standard GAT attention, with learnable mixing weights.Source: D. Adaptive Graph Attention Mechanism · pp. 3–4

A. Dataset and Preprocessing

The study evaluates four separate cohorts with cohort-specific validation, then constructs matched models from correlation-derived features and a fidelity-supported graph pipeline. Preprocessing choices can materially affect held-out performance.

A. Dataset and Preprocessing

  • The four cohorts total 4,308 subjects and are analyzed separately because parcellation widths and preprocessing pipelines differ; three use leave-site-out validation, while UCLA-CNP uses stratified 10-fold cross-validation.The cohorts were selected because they provide diagnostic labels and site or scanner identifiers; UCLA-CNP has only two scanners.Source: A. Dataset and Preprocessing · p. 4
  • ABIDE I preprocessing produces 200-region time-series data, while supervised feature selection ranks correlation connections within each training fold to avoid held-out-label leakage.Matched models use the top 8 connections; unrestricted models use 2,000, and all matched comparisons receive Pearson correlation coefficients as inputs.Source: A. Dataset and Preprocessing · pp. 4–5
  • Changing how a region is described and where its PCA basis is fitted moved held-out AUC from 0.459 to 0.739, farther than any measured quantum–classical difference.The comparison involved PCA choices and did not involve a graph.Source: A. Dataset and Preprocessing · p. 4

B. System Architecture Overview

  • The system uses a simulated quantum circuit for state preparation and fidelity computation, followed by classical edge learning and graph classification.The reported configuration uses eight qubits per region; fidelity supplies candidate edges and k-nearest-neighbour masking retains the top 20 connections per node.Source: B. System Architecture Overview · pp. 5–6
  • The quantum encoder’s per-qubit Pauli-Z expectations form node features, while its state vector is retained for fidelity; in principal graph runs the encoder and topology remain fixed during training.The classifier is a two-layer GAT with four attention heads in its first layer, global mean pooling, and a softmax read-out.Source: B. System Architecture Overview · pp. 5–6

C. Training Setup

The hybrid graph pipeline computes quantum regional representations and fidelity-based support before training; training then updates only its classical components.

C. Training Setup

  • Quantum states and fidelity-selected support are computed once per region and cached before edge learning and graph-attention training.The support retains each region’s 20 highest-fidelity neighbors.Source: C. Training Setup · pp. 6–7
  • The principal graph runs keep the quantum encoder and fidelity-derived topology fixed while Adam updates the edge-learning MLP, GAT, and edge weights.The encoder and topology remain fixed throughout training; the model-suite circuits are trained separately.Source: C. Training Setup · p. 6
  • Each training step applies the GAT to cached regional latents and retained-edge weights, then computes cross-entropy loss and updates classical parameters.The circuit preparation and support initialization occur once, without gradients.Source: C. Training Setup · p. 7

D. Experimental Design and Evaluation Metrics

The evaluation pairs quantum models with matched classical twins and tests generalization across sites where possible, using fold-paired metrics and statistical checks.

D. Experimental Design and Evaluation Metrics

  • Each quantum model is compared with its classical twin using the same features, folds, optimizer, and training budget.The matching is intended to avoid comparing models with unequal feature access or training resources.Source: D. Experimental Design and Evaluation… · p. 7
  • Leave-site-out validation holds out an acquisition site for testing, while UCLA-CNP uses stratified 10-fold validation.The site-based split is intended to reduce exploitation of scanner-specific characteristics.Source: D. Experimental Design and Evaluation… · p. 7
  • The protocol records accuracy, F1, AUC, sensitivity, specificity, and MCC, then compares per-fold paired differences with Wilcoxon tests and screens degenerate predictions.A model is marked degenerate if pooled sensitivity or specificity is below 0.30 or |MCC| < 0.02; sweep results also use Bonferroni correction.Source: D. Experimental Design and Evaluation… · p. 7

V. EXPERIMENTAL RESULTS

Across the evaluated cohorts, matched quantum and classical kernels show statistical parity, while ABIDE I results highlight both model-family differences and robustness checks.

A. Classification Performance

  • On ABIDE I, the quantum convolutional network reached 0.615 accuracy versus 0.589 for its matched classical counterpart, while graph configurations remained near chance.The largest performance difference was between graph-based and correlation-feature classifiers, rather than between quantum and classical methods; the table reports graph-model leave-site-out performance.Source: A. Classification Performance · p. 8, Table II · p. 8

B. Matched-Feature Comparison

  • Across 62 folds, the quantum kernel was statistically indistinguishable from its matched classical kernel on all four cohorts.The matched comparison found no statistically significant differences; the quantum kernel tracked the same-feature classical kernel, while both trailed the full-feature model.Source: Figure 4 · p. 8, B. Matched-Feature Comparison · p. 8
  • On ABIDE I, the matched-feature kernel comparison differs from the full-feature comparison because the latter has access to 2,000 features rather than the register-limited eight.The authors attribute losses to the full-feature models on ABIDE and REST-meta-MDD to the register limit rather than to the quantum kernel.Source: B. Matched-Feature Comparison · p. 8

C. Permutation, Seed and Noise Robustness Checks

  • Every observed ABIDE I kernel accuracy and AUC exceeded all 104 within-site label permutations, while seed and noise checks found no corrected significant matched difference.Three of four matched differences held their sign across seeds, and no tested noisy setting fell significantly below the matched RBF kernel.Source: C. Permutation, Seed and Noise… · pp. 8–9

D. Why the Two Kernels Perform Alike

The kernel analysis links predictive parity to similar effective dimensions and eigenspectra, and the model-suite section defines the broader set of matched quantum comparisons.

D. Why the Two Kernels Perform Alike

  • At matched settings, quantum fidelity and RBF kernels have nearly identical leading eigenspectra, offering a plausible account of their comparable predictive performance.On 400 ABIDE subjects, the leading 40 normalized eigenvalues correlate at r = 0.997; the authors caution that this does not prove the kernels belong to the same mathematical family.Source: D. Why the Two Kernels Perform Alike · p. 9

VI. QUANTUM MODEL SUITE

  • The quantum model suite tests five additional quantum models against their naturally matched classical comparators on identical features and folds.The suite was introduced to test whether the conclusions depended on the fidelity-kernel choice.Source: §VI · p. 9

A. Models

  • The suite includes convolutional and variational classifiers, fixed and Pegasos fidelity-kernel SVMs, and a trainable embedding kernel with a trained classical comparator.The trainable kernel comparator uses an RBF kernel on a learned linear metric under the same objective, optimizer, and step count.Source: A. Models · pp. 9–10

B. Results

Quantum-model results varied by model and cohort: the sole significant accuracy win did not replicate, while QCNN showed the most consistent direction of comparison. Kernel comparisons also varied across cohorts, and runtime favored classical models.

B. Results

  • TQEK’s only significant accuracy win was +0.028 on ABIDE I (p = 0.049), but it did not replicate on REST-meta-MDD or UCLA-CNP.Across twenty accuracy comparisons, two uncorrected results were significant in opposite directions and neither survived Bonferroni correction.Source: B. Results · p. 10
  • QCNN led its matched classical network on ABIDE I (+0.044), UCLA-CNP (+0.023), and ADHD-200 (+0.030), matched it on REST-meta-MDD, and did not trail on any cohort.Its ABIDE I difference approached significance (p = 0.058).Source: B. Results · p. 10
  • On REST-meta-MDD, classical runtimes were faster by up to two orders of magnitude; for example, QCNN took 32.6 seconds per fold versus 8.6 seconds for the classical CNN.These timings used an exact simulator without quantum shot noise or device overhead.Source: B. Results · p. 10

A. Quantum Kernel Versus Classical Kernels

B. Evaluation of Replication Across Cohorts

  • On UCLA-CNP, the quantum kernel beat the linear kernel on 8 of 10 folds (p = 0.037), whereas on REST-meta-MDD the linear kernel won on 18 of 25 folds (p = 0.011).UCLA-CNP AUCs were at chance for every model in this comparison, and neither of the two significant tests across the three further cohorts survived Bonferroni correction.Source: B. Evaluation of Replication Across… · p. 11
  • On REST-meta-MDD, the full-dimensional classical RBF reached F1 = 0.671 and AUC = 0.684, versus F1 = 0.571 and AUC = 0.544 for the quantum kernel on eight principal components.Source: B. Evaluation of Replication Across… · p. 11

A. Register Width Determines Whether Fidelity Is Usable

Fidelity graphs become less usable as register width increases, while wider QCNNs show a small single-site accuracy gain at substantial computational cost. The graph-attention pipeline remains at chance, and identical post-encoding unitary layers cannot influence fidelity graphs.

A. Register Width Determines Whether Fidelity Is Usable

  • Mean fidelity fell from 0.0652 at 4 qubits to 0.0027 at 16, while the share of pairs above 0.01 dropped from 48.4% to 4.4%.At a 0.01 floor, the k = 20 graph used 11%, 48%, and 61% below-floor edges at 8, 12, and 16 qubits; pair counts alone are necessary but not sufficient, and the crossover depends on the floor and setup.Source: Figure 7 · p. 11, A. Register Width Determines Whether… · p. 11

B. Register Width, Accuracy and Computational Cost in Variational Models

  • At 16 qubits, QCNN test accuracy was 0.520 versus 0.503 at 8, but runtime was 685.5 s versus 4.2 s (163×); this was a single-site diagnostic.The overfit gap narrowed from +0.149 to +0.110, while training accuracy never exceeded 0.65.Source: B. Register Width, Accuracy and… · p. 12

C. The Variational Layer Is Invisible to the Graph

  • Perturbing variational parameters by O(1) changed the fidelity Gram matrix by only 1.2 × 10−6, while encoding-gain gradients through the graph were 51.5.The shared unitary cancels in fidelity, leaving encoding gains as the only circuit parameters entering the graph; principal runs kept the encoder and selected top-k support fixed.Source: C. The Variational Layer Is Invisible… · p. 12
  • The cancellation also applies to identical post-encoding unitaries before state-overlap comparisons; fidelity graphs can train edge weights, but their top-k topology has no gradient.The paper distinguishes this exact zero gradient from barren-plateau gradients and notes that cancelled layers still add depth, simulation time, and hardware gate noise.Source: C. The Variational Layer Is Invisible… · p. 12

D. Performance Limitations of the Fidelity-Graph Pipeline

  • Accuracies of 0.504, 0.512, and 0.513 left all tested fidelity-graph attention configurations at chance; the study reports this arm as a negative result.The study did not isolate whether sparsification, topology quality, attention learning, or graph over-smoothing explains this outcome.Source: D. Performance Limitations of the… · p. 12

IX. ARCHITECTURAL INTERVENTIONS ON THE FEATURE BUDGET

Across the tested architectural interventions, added capacity improved training fit but not held-out performance, and gains were no greater for quantum models than for matched classical twins. Projection-fed quantum convolution and block ensembles remained statistically comparable to their classical counterparts, while the variational classifier showed a deficit.

IX. ARCHITECTURAL INTERVENTIONS ON THE FEATURE BUDGET

A. A Learned Projection

  • H-VQC was worse than corresponding classical heads in the evaluated configuration, with three of nine significant pairwise results surviving correction and eight of nine favoring classical heads.On REST-meta-MDD, H-VQC lost to H-Linear by 0.049 (p = 0.004) and H-MLP by 0.048 (p = 0.048); further read-out ablations are needed to locate the source.Source: A. A Learned Projection · p. 13

B. Ensembles Over Disjoint Feature Blocks

  • On ABIDE I, block ensembles scored 0.618 for quantum members versus 0.622 for matched classical members; on REST-meta-MDD, they scored 0.540 versus 0.542, with no significant pairwise difference.On ABIDE I both ensembles exceeded the single classical model given all 32 features (0.603), but on REST-meta-MDD the single model led (0.564).Source: B. Ensembles Over Disjoint Feature… · p. 13

C. Data Re-uploading

  • Data re-uploading did not improve on plain QCNN: depth-one median changes were 0.000 on ABIDE I, ADHD-200, and REST-meta-MDD, and −0.043 on UCLA-CNP; no result survived Bonferroni correction.Each circuit remained at eight qubits while re-uploading increased the order of expressible Fourier terms; among small-cohort comparisons, several quantum models beat C-CNN, but classical models did too.Source: C. Data Re-uploading · pp. 13–14

D. Capacity Converts to Overfitting, Not Accuracy

  • Across evaluated cohorts, models with higher training accuracy did not achieve higher held-out accuracy; across six re-uploading models, mean training accuracy predicted the REST-meta-MDD overfit gap (r = +0.983, p = 0.0004).Source: D. Capacity Converts to Overfitting,… · p. 14
  • Across six architectural families, interventions that helped quantum models helped matched classical twins at least as much, and added capacity did not improve held-out performance.Widening the register, learning a projection, ensembling, and re-uploading improved training fit but not held-out performance in the evaluated configurations.Source: D. Capacity Converts to Overfitting,… · p. 14

X. DISCUSSION: WHEN DOES QUANTUM HELP?

Quantum models had smaller train–test gaps than neural comparators across cohorts, but this pattern did not establish a quantum-specific advantage. Among positive matched comparisons, many involved non-classifying classical comparators; the remaining signals were concentrated in AUC and were not broad superiority over the strongest classical models.

A. Where the Arms Differ

  • Quantum arms had smaller overfitting gaps than both neural comparators on all four cohorts, but the pattern is descriptive and does not establish a quantum-specific mechanism or advantage.The margin over MLP was 0.034, 0.043, 0.043, and 0.051 across cohorts of 2,428, 1,035, 619, and 226 subjects; the linear classifier overfit least on every cohort.Source: Figure 8 · p. 14, A. Where the Arms Differ · p. 14

B. Analysis of Positive Quantum–Classical Differences

  • The quantum arm led in 115 of 284 matched comparisons (40.5%) and was significantly ahead in 30 (10.6%), but these counts span accuracy, F1, AUC, and specificity.The sweep compared matched quantum/classical pairs across all runs and metrics.Source: B. Analysis of Positive… · p. 14, Table VI · p. 15
  • Eighteen of 30 significant wins were screened out for failed classification (sensitivity or specificity below 0.30 or |MCC| < 0.02 on pooled-fold matrices) or comparator training failure; remaining cases offer no strong evidence of quantum-specific behaviour.Flagged classical arms included C-CNN on UCLA-CNP (MCC = −0.005) and the suite’s 1-D CNN on ADHD-200 (−0.015).Source: B. Analysis of Positive… · p. 15
  • QCNN exceeded its matched CNN by +0.037 AUC (p = 0.033) on ABIDE I and by +0.059 AUC (p = 0.047) on ADHD-200, where that comparator did not separate classes.On ABIDE I, the QCNN accuracy difference was +0.044 at p = 0.058; the re-uploading models also had ABIDE I AUC leads over C-MLP at depths zero, one, and two.Source: B. Analysis of Positive… · p. 15
  • On ABIDE I, TQEK exceeded its trainable classical kernel by +0.064 AUC (p = 0.0012), but the effect reversed on REST-meta-MDD (−0.013) and UCLA-CNP (−0.045).The ABIDE I result was the largest observed single-cohort effect and was against a functional comparator.Source: B. Analysis of Positive… · p. 15
  • Against the best classical AUC, quantum models tied on ABIDE I (0.679 versus 0.679) and trailed on the other three cohorts; no arm gained accuracy from training-set Youden’s J calibration.The quantum arm’s AUC-to-accuracy pattern was not explained by threshold calibration; averaged accuracy changes were −0.000 for quantum, +0.009 for C-CNN, +0.002 for C-MLP, and −0.009 for C-Linear.Source: B. Analysis of Positive… · p. 15

C. Assessing Evidence for Quantum Advantage

The evidence for quantum advantage was limited by multiple comparisons, cohort-specific results, and dependence among metrics. In a direct low-data test, quantum models retained smaller train–test gaps but did not gain accuracy at the smallest training sizes; several design and evaluation boundaries further qualify positive findings.

C. Assessing Evidence for Quantum Advantage

  • The quantum arm led in only 115 of 284 comparisons, and none survived Bonferroni correction; correlated metrics mean 30 significant results cannot be read as 16 independent effects.Almost every clean win was on ABIDE I, the cohort with the strongest overall classification performance, and TQEK’s strongest result reversed on both other measurable cohorts.Source: C. Assessing Evidence for Quantum… · pp. 15–16
  • Across cohorts, the quantum kernel’s mean accuracy difference from matched RBF was +0.002 (95% CI −0.009 to +0.014), bounding average improvement below 1.4 percentage points under the fitted model.The corresponding bounds were 4.1 points for QCNN versus CNN and 0.4 points for hybrid QCNN versus hybrid CNN; ADHD-200’s seven folds contributed little precision.Source: Figure 10 · p. 16, C. Assessing Evidence for Quantum… · p. 16

D. The Low-Data Regime, Tested Directly

  • On ABIDE I, the quantum arm’s train–test gap stayed smaller than the matched MLP’s at all four training sizes, but the relative gap did not widen as sample size fell.The MLP gap was 1.46× to 2.02× the quantum gap across m = 100 to 800.Source: D. The Low-Data Regime, Tested Directly · p. 16
  • At m = 100 and 200, the linear classifier—not the quantum arm—led accuracy; only one comparison was significant, favoring quantum by +0.010 over MLP at m = 400.The quantum arm was best at m = 400 (0.586) and m = 800 (0.613), not at the two smallest sizes.Source: D. The Low-Data Regime, Tested Directly · p. 16

E. Sources of Apparent Quantum Advantage

  • On REST-meta-MDD, compressing 19,900 correlations to eight components cost more than any measured quantum–classical difference; widening the register increased pairwise-fidelity concentration.The study identifies an unmatched feature budget as a source of apparent quantum advantage.Source: E. Sources of Apparent Quantum Advantage · p. 16
  • Positive results on one cohort did not replicate elsewhere; on ADHD-200, quantum wins over C-CNN were qualified by C-Linear beating it by a larger margin.TQEK’s +0.028 ABIDE I gain reversed on REST-meta-MDD and UCLA-CNP.Source: E. Sources of Apparent Quantum Advantage · p. 16
  • A post-hoc fidelity-kernel bandwidth can span near-constant to near-identity effective dimensions, while its default leaves the kernel memorizing training data.The study’s controls therefore included multiple classical arms, all cohorts, and both uncorrected and corrected significance.Source: E. Sources of Apparent Quantum Advantage · p. 16

F. Comparison with Published Results

Published comparisons place the pipeline as competitive on ABIDE I but below the strongest published ADHD-200 and REST-meta-MDD results. Across the evaluated cohorts, quantum models mostly matched classical counterparts, while the findings have defined data, simulation, and validation boundaries.

F. Comparison with Published Results

  • On ABIDE I, the full-correlation linear SVM reached 0.671 accuracy, matching the protocol-matched ~0.67 anchor; results were below the strongest published results on ADHD-200 and REST-meta-MDD.Table VII compares this work with published results on the same cohorts and reports each validation protocol.Source: F. Comparison with Published Results · pp. 16–17, Table VII · p. 17
  • ABIDE I accuracy comparisons require matching validation, modalities and covariates, and parcellation; DAGMNet’s 0.9159 accuracy used ten-fold multimodal data, different atlases, and site identity.The reported protocol and inputs mean that the 0.9159 result is not a like-for-like comparison with this study’s leave-site-out connectivity-only setting.Source: F. Comparison with Published Results · p. 17

XI. LIMITATIONS

  • Whole-brain models lack regional priors, near-chance performance limits attention-weight interpretation, and Pearson-derived inputs do not test raw-signal nonlinear dependence.Testing nonlinear dependence in raw regional BOLD signals would require encoding the time series directly.Source: §XI · p. 17
  • The main results use exact statevector simulation; hardware gate noise and decoherence remain untested, while usable fidelity range narrowed as register width increased in the evaluated datasets.The authors leave open whether the fidelity trend holds on hardware that does not incur simulation cost.Source: §XI · p. 17
  • Intervals are computed across folds, not initialisations; the largest reported standard deviation of mean accuracy across four ABIDE I seeds was 0.017 for QSVM-Pegasos.Differences smaller than 0.017 between stochastic models warrant caution; the permutation null covers only the ABIDE I kernel arm.Source: §XI · p. 17
  • Graph findings are bounded by directed, unsymmetrised adjacency with fixed k=20 and baselines chosen for matched comparisons rather than representing the latest adaptive, dynamic, or hypergraph architectures.The effects of adjacency direction and k on node distinctiveness and accuracy were not measured.Source: §XI · p. 17

XII. CONCLUSIONS

  • Most quantum models matched classical counterparts without consistently outperforming them; at eight qubits, the quantum kernel and matched classical RBF were indistinguishable across four cohorts, with the 95% interval excluding an average advantage above 1.4 percentage points.The authors state that the results do not demonstrate a reproducible quantum advantage across the evaluated cohorts.Source: §XII · p. 18
  • Higher training accuracy did not yield higher held-out accuracy in the evaluated comparisons, and the matched benchmark is specific to simulated circuits, fidelity-based topology, and tested model families.The authors provide a matched evaluation protocol and released per-fold results for future studies.Source: §XII · p. 18

Related papers

Loading 2610.21344v1…