Identifying Neural Signatures from fMRI using Hybrid Principal Components Regression

Jared Rieck; Joshua L. Gowin; Julia Wrobel; Martin Paulus; Ryan Peterson; Yue Wang

REVIEW 3 major objections 5 minor 5 references

Identifying Neural Signatures from fMRI using Hybrid Principal Components Regression

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that giving principal components index-dependent LASSO penalties, rather than a uniform penalty, improves fMRI-based decoding of mental state, with reported gains up to 51.7% in cross-validated deviance R2.

desk verdict A sensible ranked-sparsity PCR extension with a genuinely interesting JSRL hybrid, but the headline gains rest on a CV protocol that may be circular; worth refereeing with a demand for nested CV and code. read the letter →

arxiv 2509.07300 v1 pith:RFQGVO3F submitted 2025-09-09 stat.ML cs.LG

Jared Rieck , Julia Wrobel , Joshua L. Gowin , Yue Wang , Martin Paulus , Ryan Peterson This is my paper · ORCID

classification stat.MLcs.LG MSC 62J07

keywords fMRImulti-voxelpatternanalysisLASSOPCRprincipalcomponentsregressionsparsityrankingpenalizedlogisticcross-validationbraindecoding

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

The paper tries to show that standard LASSO principal-components regression used in fMRI multi-voxel pattern analysis fails when task-relevant signal is concentrated in a few components, because it penalizes all components equally. It proposes sparsity-ranked LASSO PCR (SRL), which makes the L1 penalty a function of the component index, and Joint Sparsity-Ranked LASSO (JSRL), which adds individual voxels as predictors alongside components with a voxel penalty set by an information-parity calculation. Across three fMRI tasks, the authors report that SRL and JSRL match or beat standard LASSO PCR on all nine binary classification problems, with JSRL improving cross-validated deviance R2 by up to 51.7% and AUC by up to 7.3%. If these gains hold out of sample, the method offers a drop-in replacement that is both more accurate and more interpretable, localizing signal to core voxel clusters.

What carries the argument

The central object is the ranked penalty function W_k (k is the principal-component index, gamma shapes the penalty), together with the information-parity rescaling r_IP = -log(Σ_{k=1}^K k^{-2|gamma|}) / (2 log V) + 1/2 that balances prior Fisher information between the PC and voxel modalities in JSRL. The penalty function formalizes the implicit practice of preferring high-variance components; the parity formula gives a principled starting point for tuning the relative weight of voxels, with tau used as a searchable multiplier.

What would settle it

Run a strictly nested cross-validation on the same three datasets where the outer test folds are never touched while selecting gamma, r, tau, and lambda, and compare SRL and JSRL against PCL on the untouched folds; if the improvement in deviance R2 drops to near zero or reverses, the central claim fails.

Watch

Extended reading notes

Core claim

Building on the observation that standard LASSO PCR assigns equal prior weight to every principal component, the paper defines a penalty weight W_k = k^gamma for gamma >= 0 (and its mirror for gamma < 0), so the model can express a prior that early components are more or less likely to carry task signal. JSRL extends this by concatenating PCs and raw voxels, assigning voxels a constant penalty scaled by V^r, and choosing the rescaling r so that the PC and voxel modalities contribute equal prior Fisher information; a tunable multiplier tau then relaxes or tightens that parity. The paper reports that on emotion regulation, monetary incentive, and risky-gain tasks, SRL and JSRL consistently mat

Load-bearing premise

The claim that SRL and JSRL outperform PCL rests on the assumption that the reported cross-validated AUC and deviance are unbiased out-of-sample estimates; because the paper tunes gamma, r, tau, and lambda without a clearly separated outer evaluation loop and uses no independent holdout, leakage could inflate the gains.

Editorial extensions

If this is right

SRL and JSRL can be dropped into existing LASSO-PCR pipelines with only a change in penalty weights, so fMRI labs can adopt the method without new preprocessing.
The selected gamma for a task is interpretable: near-zero gamma means no benefit from ranking, while gamma > 0 indicates the task signal is concentrated in early components; future studies can use it to characterize tasks.
JSRL's voxel-level weights produce more localized, centered predictive maps, suggesting it may reduce edge and partial-volume noise automatically.
Because JSRL matched or outperformed PCL in all nine tasks, a researcher who currently uses uniform LASSO PCR has no accuracy reason to avoid ranked sparsity.

Reading between the lines

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

If ranked sparsity generalizes, the same index-weighted penalty could be applied to other dimension-reduction methods, such as PLS or sparse PCA, and to non-neuroimaging high-dimensional data where variance-ordered components coexist with raw features.
The information-parity formula suggests a principled way to set group penalties between any two feature modalities, not just PCs versus voxels; one could test whether parity-based initialization improves tuning stability in other multimodal settings.
The reported tau values below one across datasets imply that, in practice, voxels should receive weaker penalties than parity suggests; a testable extension is whether tau tends to correlate with the true spatial sparsity of the signal.
A fully independent replication with an external holdout dataset would clarify whether the 51.7% deviance-R2 gain is real, a step the authors flag as future work.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Here's the short version: the ranked-sparsity idea is a sensible extension of LASSO PCR and the JSRL hybrid is genuinely interesting, but the empirical payoff is not yet credible because the evaluation protocol as written doesn't separate hyperparameter tuning from performance reporting. The 51.7% and 7.3% headline numbers could be the minimum of a noisy CV criterion rather than true out-of-sample gains.

What's new: the W_k = k^gamma weighting on PC indices is a clean, interpretable way to encode the prior that earlier components carry more signal, and the information-parity derivation for r_IP is a nice touch. The JSRL model that combines PCs and raw voxels under one penalty is a plausible way to recover localized signal that PCs smear out. The paper is also transparent about its limitations—no independent holdout, possible PCA leakage, qualitative interpretability.

Where the soft spots are: the stress-test note is on target. Section 2.5 says gamma is chosen by cross-validated deviance over 27 values, lambda by 10-fold CV, and tau by a 'final cross-validation step,' with the final model being the one that minimized CV deviance. There's no sign of an outer loop that leaves the evaluation folds untouched. With n=68–82 and a large configuration grid, that's a recipe for optimistic bias, and since SRL/JSRL have many more configurations than PCL, the bias would favor them. The paper's own discussion admits the evaluation is only CV and PCA may leak. So the central empirical claim is conditional at best.

That said, this is fixable. The method is not broken; the protocol as described is ambiguous. If the authors resubmit with a properly nested CV, or an independent test set, and the gains persist, this would be a useful drop-in for MVPA. Code and data availability also need to come with it.

Bottom line: worth sending to reviewers, but they should be told to focus on the CV design. I'd give it a maybe for reading group—the method is worth chewing on, but the evaluation will dominate the conversation.

Referee Report

3 major / 5 minor

Summary. The paper proposes two modifications to LASSO principal components regression for fMRI decoding: Sparsity-Ranked LASSO (SRL), which assigns index-based penalty weights to principal components, and Joint Sparsity-Ranked LASSO (JSRL), which adds voxel-level predictors with voxel penalties calibrated by an 'information parity' formula. The methods are compared against principal-component LASSO, voxel LASSO, and voxel elastic net on 9 binary classification tasks from three fMRI studies. The authors report that JSRL improves cross-validated deviance R² by up to 51.7% and AUC by up to 7.3% relative to PCL, and that sparsity-ranked models often match or outperform the alternatives. The paper includes qualitative brain-activation maps and region annotations for emotion-regulation tasks.

Significance. If the reported gains are genuine, the proposed methods offer a simple, drop-in modification to a widely used MVPA tool, with potential practical value in decoding mental states from fMRI. The ranking idea is intuitive and the information-parity formula provides a principled starting point for the voxel penalty. However, the central claim is purely empirical, and the evidence is compromised by an ambiguous evaluation protocol: hyperparameters are selected by minimizing cross-validated deviance, and the same deviance is then reported as the performance metric. With no clearly separated outer loop or independent holdout, the headline improvements may be optimistically biased. The paper does not currently provide code or data, further limiting verification.

major comments (3)

[§2.5 (Model Comparison)] The evaluation protocol does not separate hyperparameter tuning from performance estimation. For SRL, γ is chosen by minimizing cross-validated deviance; for JSRL, γ and τ are likewise selected by cross-validated deviance, and the 'final model selected for prediction' is the one minimizing that same criterion. No outer loop is described that keeps evaluation folds untouched during tuning. The reported AUC and deviance thus appear to be the minimum over many candidate configurations, which is expected to be optimistically biased, especially with n=68–82 and grids of 27 γ values and multiple τ values. The Discussion's admission that evaluation is only through CV without an independent holdout reinforces this concern. The authors should either describe a fully nested CV scheme in which tuning occurs inside each training fold, or evaluate on an independent holdout, and report the correspondi
[§2.2.1, Eq. (4)] The 'information parity' derivation is not load-bearing for the empirical claim because the resulting r_IP is only a starting point; the final r = τ·r_IP is tuned by cross-validation. This is acceptable, but the text should be clearer that Eq. (4) is a heuristic initialization, not a guarantee of optimal performance. As written, calling it 'information parity' may overstate its role. The derivation itself appears algebraically correct, but its connection to predictive accuracy is unestablished.
[Discussion, limitations paragraph] The authors acknowledge that 'model performance was evaluated solely through cross-validation, without validation in an independent holdout dataset' and that PCA 'introduces a potential avenue for data leakage.' These admissions directly bear on the validity of the abstract's quantitative claims. Since the abstract presents the 51.7% and 7.3% improvements as the main result, the manuscript should either provide the nested-CV or holdout evidence needed to substantiate those numbers, or substantially temper the claims.

minor comments (5)

[§2.5, Figure 1 caption] The text says 'A total of five modeling methods were applied' but then 'The construction of these six models'—the count is inconsistent. Also, 'left panel of Figure 1' appears after Figure 1 has already been referenced; check cross-references.
[§2.2.1] Typo: 'voxel indeces' should be 'voxel indices.' Minor.
[§2.5] The description of the repeated 10-run procedure is vague: 'each run using a different random seed to determine training/testing splits during 10-fold cross-validation.' It is unclear whether tuning is performed separately within each training split or once on the full data. This ambiguity is part of the major comment, but clarifying the exact loop structure would help.
[§3, Table 1] The Table 1 caption calls τ a 'multiplier' but the text calls it 'rescaling factor.' Consistent terminology would improve readability. Also, reporting only averages over seeds loses information about variability across seeds; consider including SDs or ranges.
[Data and Code Availability] Code availability is deferred to acceptance. For a methods paper, sharing code at submission is important for reproducibility; consider making it available in a repository or as supplementary material.

Circularity Check

1 steps flagged · score 6.0 of 10

Tuning and evaluation share the same CV deviance, so reported gains are the selection minimum

fitted input called prediction [Section 2.5 (Model Comparison), JSRL tuning paragraph; corroborated by Discussion, first limitation]
"A final cross-validation step was performed to determine the optimal value for the voxel-mode penalty from a set of candidate τ values ... The final model selected for prediction used the parameter values r, γ, λfrom the model which minimized the cross-validated deviance in this step."

The headline claims (up to 51.7% improvement in CV deviance R², 7.3% in AUC) are presented as out-of-sample performance, but the final model's hyperparameters (r, γ, λ) are explicitly chosen as the minimizer of the cross-validated deviance on the same data. No outer loop is described that would keep evaluation folds untouched during this selection; the Discussion concedes 'model performance was evaluated solely through cross-validation, without validation in an independent holdout dataset.' Thus the reported 'prediction' is the tuning objective itself: the minimum of a noisy CV criterion over a grid. Because SRL/JSRL have more tuning dimensions (γ and τ grids) than PCL (only λ), this selection bias is asymmetric and can manufacture the reported improvements even under the null. This is fit

full rationale

The methodological derivation—the weighted LASSO objective and the Fisher-information parity formula for r_IP—is self-contained algebra and does not reduce to its inputs. However, the central empirical claim is evaluated using cross-validated deviance that is also used to select the hyperparameters (γ, τ, λ). The paper states the final model is the one that minimized cross-validated deviance in the tuning step, and there is no independent holdout. Therefore the headline improvements are the fitted CV minimum rather than out-of-sample predictions. This is a genuine evaluation circularity (pattern: fitted input called prediction). It is not a self-citation issue: the Peterson & Cavanaugh citations are motivational and not load-bearing for the derivation, and no uniqueness theorem or ansatz is smuggled via citation. Score 6 reflects that the reported 'prediction' reduces by construction to the selection criterion, but the underlying method itself is not mathematically circular.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central empirical claim rests on a handful of tuned hyperparameters (gamma, lambda, tau) and on the assumption that the cross-validation procedure gives unbiased estimates. The information-parity formula is a derivation, but the actual voxel penalty is tau*r_IP with tau fit to the data. No new physical or mathematical entities are postulated.

free parameters (3)

gamma = varies by task, e.g., 0.030 to 1.000 (Table 1)
Exponent in the PC penalty weighting function W_k = k^gamma; selected by cross-validation to minimize deviance.
lambda = not reported, tuned per gamma/tau
Overall L1 penalty strength selected via 10-fold CV.
tau = varies by task, e.g., 0.115 to 0.900 (Table 1)
Multiplier on the information-parity r_IP for the voxel penalty; selected by cross-validation.

assumptions (5)

domain assumption Ranked sparsity framework of Peterson & Cavanaugh (2022) is valid and appropriate for PCR
The weight function and objective in Eq. (1) and (2) are taken from that framework; the present paper extends it to PCs.
domain assumption Task-relevant fMRI signal is more likely concentrated in high-variance principal components
This prior motivates W_k = k^gamma for gamma>0; the paper itself notes the opposite can be true for subtle tasks.
standard math The Laplace prior interpretation of LASSO and the Fisher information calculation for prior parity are correct
Used to derive r_IP in Eq. (4).
domain assumption PCA performed on the full dataset is unsupervised and does not leak outcome information
The authors acknowledge in the Discussion that PCA could leak and influence out-of-sample fits, making this a fragile assumption.
domain assumption Repeated 10-fold cross-validation with different seeds yields unbiased performance estimates despite model selection on the same data
The paper does not explicitly describe a nested outer loop; the validity of the headline improvements depends on this.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Identifying Neural Signatures from fMRI using Hybrid Principal Components Regression." pith.science (2026). https://pith.science/paper/RFQGVO3F

@misc{pith2026250907300,
  author       = {Pith},
  title        = {Pith review of: Identifying Neural Signatures from fMRI using Hybrid Principal Components Regression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RFQGVO3F}},
  note         = {Machine review of arXiv:2509.07300}
}

abstract

Recent advances in neuroimaging analysis have enabled accurate decoding of mental state from brain activation patterns during functional magnetic resonance imaging scans. A commonly applied tool for this purpose is principal components regression regularized with the least absolute shrinkage and selection operator (LASSO PCR), a type of multi-voxel pattern analysis (MVPA). This model presumes that all components are equally likely to harbor relevant information, when in fact the task-related signal may be concentrated in specific components. In such cases, the model will fail to select the optimal set of principal components that maximizes the total signal relevant to the cognitive process under study. Here, we present modifications to LASSO PCR that allow for a regularization penalty tied directly to the index of the principal component, reflecting a prior belief that task-relevant signal is more likely to be concentrated in components explaining greater variance. Additionally, we propose a novel hybrid method, Joint Sparsity-Ranked LASSO (JSRL), which integrates component-level and voxel-level activity under an information parity framework and imposes ranked sparsity to guide component selection. We apply the models to brain activation during risk taking, monetary incentive, and emotion regulation tasks. Results demonstrate that incorporating sparsity ranking into LASSO PCR produces models with enhanced classification performance, with JSRL achieving up to 51.7\% improvement in cross-validated deviance $R^2$ and 7.3\% improvement in cross-validated AUC. Furthermore, sparsity-ranked models perform as well as or better than standard LASSO PCR approaches across all classification tasks and allocate predictive weight to brain regions consistent with their established functional roles, offering a robust alternative for MVPA.

Figures

Figures reproduced from arXiv: 2509.07300 by the authors.

**Figure 2.** Classification performance for each method is compared in terms of cross-validated AUC (A) [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗

**Figure 3.** Distribution of beta coefficient values in the brain space for PCL, SRL, and JSRL models are [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references · 4 canonical work pages

[1]

Abraham, A., Pedregosa, F., Eickenberg, M., Gervais, P., Mueller, A., Kossaifi, J., Gramfort, A., Thirion, B., & Varoquaux, G. (2014). Machine learning for neuroimaging with scikit-learn.Frontiers in Neuroinformatics,8,

work page 2014
[14]

A., & Haney, M

https://doi.org/10.3389/fninf.2014.00014 Bedi, G., Lindquist, M. A., & Haney, M. (2015). An fmri-based neural signature of decisions to smoke cannabis.Neuropsychopharmacology,40(12), 2657–2665. https://doi.org/10.1038/npp.2015. 135 Boulesteix, A.-L., De Bin, R., Jiang, X., & Fuchs, M. (2017). IPF-LASSO: Integrative l1-penalized regression with penalty fac...

arXiv 2014
[173]

org / 10

https : / / doi . org / 10 . 1186 / s12888-022-03780-y Kirk-Provencher, K. T., Hakimi, R. H., Andereas, K., Penner, A. E., & Gowin, J. L. (2024). Neural response to threat and reward among young adults at risk for alcohol use disorder.Addiction Biology,29(2), e13378. https://doi.org/10.1111/adb.13378 Knyazev, G., Savostyanov, A., Bocharov, A., Saprigyn, A...

work page arXiv 2024
[1099]

L., Stewart, J

https://doi.org/10.1016/j.biopsych.2009.07.014 Gowin, J. L., Stewart, J. L., May, A. C., Ball, T. M., Wittmann, M., Tapert, S. F., & Paulus, M. P. (2014). Attenuated insular processing during risk predicts relapse in early abstinent methamphetamine- dependent individuals.Neuropsychopharmacology,39(6), 1379–1387. https://doi.org/10.1038/ npp.2013.335 James...

work page arXiv 2009
[1775]

P., Keysers, C., Barrios, J

https://doi.org/10.1038/s41598-024-52203-3 Speer, S. P., Keysers, C., Barrios, J. C., Teurlings, C. J., Smidts, A., Boksem, M. A., Wager, T. D., & Gazzola, V. (2023). A multivariate brain signature for reward.NeuroImage,271, 119990. https: //doi.org/https://doi.org/10.1016/j.neuroimage.2023.119990 Tang, J., LeBel, A., Jain, S., & Huth, A. G. (2023). Seman...

work page arXiv 2023

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Abraham, A., Pedregosa, F., Eickenberg, M., Gervais, P., Mueller, A., Kossaifi, J., Gramfort, A., Thirion, B., & Varoquaux, G. (2014). Machine learning for neuroimaging with scikit-learn.Frontiers in Neuroinformatics,8,

work page 2014

[2] [14]

A., & Haney, M

https://doi.org/10.3389/fninf.2014.00014 Bedi, G., Lindquist, M. A., & Haney, M. (2015). An fmri-based neural signature of decisions to smoke cannabis.Neuropsychopharmacology,40(12), 2657–2665. https://doi.org/10.1038/npp.2015. 135 Boulesteix, A.-L., De Bin, R., Jiang, X., & Fuchs, M. (2017). IPF-LASSO: Integrative l1-penalized regression with penalty fac...

arXiv 2014

[3] [173]

org / 10

https : / / doi . org / 10 . 1186 / s12888-022-03780-y Kirk-Provencher, K. T., Hakimi, R. H., Andereas, K., Penner, A. E., & Gowin, J. L. (2024). Neural response to threat and reward among young adults at risk for alcohol use disorder.Addiction Biology,29(2), e13378. https://doi.org/10.1111/adb.13378 Knyazev, G., Savostyanov, A., Bocharov, A., Saprigyn, A...

work page arXiv 2024

[4] [1099]

L., Stewart, J

https://doi.org/10.1016/j.biopsych.2009.07.014 Gowin, J. L., Stewart, J. L., May, A. C., Ball, T. M., Wittmann, M., Tapert, S. F., & Paulus, M. P. (2014). Attenuated insular processing during risk predicts relapse in early abstinent methamphetamine- dependent individuals.Neuropsychopharmacology,39(6), 1379–1387. https://doi.org/10.1038/ npp.2013.335 James...

work page arXiv 2009

[5] [1775]

P., Keysers, C., Barrios, J

https://doi.org/10.1038/s41598-024-52203-3 Speer, S. P., Keysers, C., Barrios, J. C., Teurlings, C. J., Smidts, A., Boksem, M. A., Wager, T. D., & Gazzola, V. (2023). A multivariate brain signature for reward.NeuroImage,271, 119990. https: //doi.org/https://doi.org/10.1016/j.neuroimage.2023.119990 Tang, J., LeBel, A., Jain, S., & Huth, A. G. (2023). Seman...

work page arXiv 2023