Sleep Brain and Cardiac Activity Predict Cognitive Flexibility and Conceptual Reasoning Using Deep Learning

Boshra Khajehpiri; Eric Granger; Fiona C. Baker; Massimiliano de Zambotti; Mohamad Forouzanfar

REVIEW 5 major objections 3 minor 18 references

Sleep Brain and Cardiac Activity Predict Cognitive Flexibility and Conceptual Reasoning Using Deep Learning

T0 review · 5 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Five hours of sleep EEG and ECG separate high from low executive function at 80.3% accuracy.

desk verdict New application with public code, but the 80.3% headline rests on full-data thresholding and architecture selection on the same folds. read the letter →

arxiv 2506.00279 v1 pith:FK56CYCT submitted 2025-05-30 cs.AI cs.LG

Boshra Khajehpiri , Eric Granger , Massimiliano de Zambotti , Fiona C. Baker , Mohamad Forouzanfar This is my paper

classification cs.AIcs.LG

keywords sleepEEGECGheartratevariabilitycognitiveflexibilityPCETtransformerdeeplearningpolysomnography

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 ordinary overnight sleep recordings contain enough information to predict a person's executive function, specifically cognitive flexibility and conceptual reasoning. It introduces CogPSGFormer, a hybrid CNN-Transformer that takes one EEG channel and one ECG channel plus derived spectral and heart-rate-variability features, and claims 80.3% accuracy in sorting 817 adults into low versus high PCET Concept Level Responses groups. The result matters because it points toward a non-invasive, physiology-based screen for cognitive performance in a general adult population, beyond the age-related impairment focus of prior work.

What carries the argument

The load-bearing mechanism is CogPSGFormer's multi-scale convolutional embedding: one shared 1D convolution whose filters are dynamically sliced to kernel sizes 3 and 7, applied segment-by-segment to raw EEG (30-second epochs) and ECG (2-minute epochs), then global-average-pooled, positionally encoded, and fed to Transformer encoders. Parallel linear-embedded Transformer paths process three feature time series (six EEG power bands, six time-domain HRV, five frequency-domain HRV), and the last-step representations of all five streams are concatenated for classification. This design captures both fine-grained sleep events such as spindles and slow patterns across the night while keeping EEG and ECG information separate.

What would settle it

Hold out a cohort, set the median split only on training labels, and freeze the architecture before evaluating; if accuracy on the untouched test cohort falls to chance (near 50%), the claim of predictive markers is not supported. A label-permutation test would give the same answer.

Watch

Extended reading notes

Core claim

The paper's central claim is that a multi-scale convolutional-transformer model can extract stable markers of cognitive flexibility from sleep signals: trained on 5 hours of C3-M2 EEG and ECG with HRV and EEG power-band features, CogPSGFormer classifies STAGES participants into low versus high Penn Conditional Exclusion Test performers with 80.3% accuracy and F1 0.89 under subject-based 10-fold cross-validation. This beats the vanilla Transformer (76.16%) and stacked LSTM (73.10%) on the same mixed raw-plus-feature inputs, and the ablation study identifies the best fusion strategy: separate processing paths for EEG and ECG with a shared convolutional layer across two kernel sizes.

Load-bearing premise

The reported accuracy assumes that choosing the low/high cutoff and the model architecture with the full dataset in view does not inflate the 10-fold score, so the 80.3% reflects how well sleep signals predict a new person's executive function.

Editorial extensions

If this is right

Executive function could potentially be screened from a single night's sleep physiology without cognitive testing.
Combining raw physiological signals with engineered spectral and HRV features is better than either alone for this prediction task.
Transformer self-attention models long-range overnight dependencies more effectively than stacked LSTMs for sleep-based cognition.
Keeping EEG and ECG streams separate while sharing multi-scale convolutional weights yields the best accuracy, offering a design principle for multimodal sleep models.

Reading between the lines

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

Because the low/high cutoff is the cohort median computed on all 817 participants, the 80.3% figure is a ranking accuracy within this sample, not a calibrated absolute assessment; a new cohort with a different score distribution could see lower performance.
A stricter evaluation with nested cross-validation and a fixed threshold defined only on training data would test whether the physiological markers generalize to an absolute standard.
If the markers are real, consumer sleep trackers that record single-channel EEG and PPG-derived HRV could be tested for lightweight cognitive monitoring, though signal quality and home conditions differ from in-lab polysomnography.
Attention-weight analysis on the Transformer could identify which sleep segments carry the predictive signal, turning the classifier into a hypothesis generator for sleep-cognition research.

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

First, the useful part: this is a new application—predicting PCET-based executive function from multimodal EEG/ECG sleep PSG in a general adult cohort—and the closest prior work (Song et al.) targeted cognitive impairment in older adults with single-channel EEG. The shared-weight multi-scale CNN embedding is a modest but reasonable design choice, and the ablation is clearly reported. Subject-based 10-fold cross-validation is the right frame, and the code is public, which earns real credit. The paper also shows, sensibly, that mixing raw signals with engineered features beats either alone.

The soft spots are all in the evaluation, and they matter. The binary label uses the median PCET score computed over all 817 participants, so each training fold sees labels that depend on test-fold scores. Table III then reports an ablation over eight architectures evaluated on the same folds, and the best one is presented as the headline 80.3%. Selecting the max over models on the same test folds inflates the estimate even with a correct subject split. Z-score normalization is described without specifying that statistics were fit on training folds only; if full-data statistics were used, that is standard feature leakage. There are also no confidence intervals and no demographic-controlled baseline, so the 80.3% has no clear uncertainty band. Individually these are common omissions; together they make the result 'suggestive' rather than 'demonstrated.'

The paper's own limitations paragraph is honest about missing demographic integration and the binary outcome, but it doesn't flag the median-threshold or model-selection issues, which are more load-bearing.

Bottom line: this deserves a serious referee. The application is new, the pipeline is plausible, and the code is public. But the predictive claim should be re-evaluated with a nested or hold-out protocol—compute the median inside training folds, select the architecture inside folds, and report CIs. If that rerun keeps accuracy near 80%, it's a solid result. As it stands, I'd trust the direction, not the number.

Referee Report

5 major / 3 minor

Summary. The paper introduces CogPSGFormer, a multi-scale convolutional-transformer model that combines one channel of EEG (C3-M2), one channel of ECG, and derived spectral and HRV features to classify 817 STAGES participants into low versus high PCET Concept Level Responses groups. Under subject-based 10-fold cross-validation, the model achieves 80.3% accuracy and F1 0.89, outperforming a vanilla Transformer (76.16%) and a stacked LSTM (73.10%). The authors interpret this as evidence that sleep EEG and ECG contain stable, learnable markers of executive function in a general adult population, and they provide publicly available code.

Significance. If the predictive claim is robust, this work is among the first to show that deep learning on raw PSG signals plus engineered features can predict executive-function scores in a non-clinical adult cohort, extending prior work that focused on memory or on older adults with cognitive impairment. The paper's strengths include a public code repository, a subject-based cross-validation protocol, a systematic ablation over input representations and architectural choices, and the use of multimodal (EEG+ECG) data. However, the significance of the central claim depends on whether the reported accuracy is an unbiased estimate of generalization; several evaluation choices, as detailed in the major comments, currently undermine that estimate.

major comments (5)

[Section II-B] The binary label is defined by the median PCET score computed over the full 817-participant dataset, and the cross-validation folds are then created from these labels. Because test-fold participants' scores contribute to the threshold, the training labels in every fold carry distributional information about the test set; this is label leakage and makes the 80.3% accuracy a cohort-relative, not a fixed-threshold, estimate. For an unbiased evaluation, the median (or another threshold) must be computed within each training fold only, or a pre-specified external threshold should be used.
[Table III / Section III-A] The main result (MS-SC-SHS, 80.3%) is selected as the best among at least eight architectures (three vanilla Transformer variants, LSTM, single-scale, MS-SHC-SHS, MS-SC-SHS, MS-SC-SS) all evaluated on the same 10-fold splits, and the reported number is the maximum accuracy over that set. Selecting the best model on the test folds inflates the expected generalization performance, so the 80.3% figure is not an unbiased estimate for a pre-specified model. A nested cross-validation or a held-out validation set for architecture selection is required to support the predictive claim.
[Section II-C] The text states that 'all raw and feature sequences were normalized using z-score standardization' but does not specify whether the mean and standard deviation used for normalization were computed within each training fold (or on the training portion only). If the normalization statistics are estimated from the full dataset, including test folds, this constitutes feature leakage. The authors should clarify the procedure and, if necessary, rerun the evaluation with per-fold normalization.
[Section II-C vs. Section II-D1, Eqs. (4)-(5)] There is an internal inconsistency in the definition of HRV segment lengths. Section II-C states that time-domain HRV parameters are computed on 2-minute segments and frequency-domain HRV parameters on 5-minute segments, but Section II-D1 and Eqs. (4)-(5) say the opposite: '6 time HRV parameters extracted per 5-minute window' and '5 frequency HRV parameters extracted per 2-minute window.' This contradiction affects the exact input dimensions and the reproducibility of the method; the authors should correct one of the two descriptions and align the notation.
[Section III-B / Table III] The paper reports point estimates of accuracy and F1 without confidence intervals, standard deviations across folds, or statistical significance tests. Given that the differences among several architectures are only 1-4 percentage points (e.g., single-scale 79.81% vs. MS-SC-SHS 80.30%), it is unclear whether these differences are within noise. Additionally, the paper does not include a baseline that uses only demographic variables (age, sex, BMI, OSA status), which would help establish whether the sleep-derived signals add predictive value beyond easily available covariates. The Discussion (Section IV) acknowledges demographic influence but does not address this as a quantitative baseline.

minor comments (3)

[Section II-D, Eq. (9)] In the sentence after Eq. (9), '2. dconv' appears to be a typo for '2 · dconv' or '2×dconv'.
[Section II-C] The text refers to 'Y ASA sleep analysis toolbox' with an unusual spacing; it should be 'YASA' (Yet Another Spindle Algorithm) and a proper citation should be provided. Similarly, 'Neurokit' is mentioned without a reference.
[Figure 1] Figure 1 is dense and the small labels (e.g., for the HRV feature blocks) are difficult to read; increasing font size or splitting into subfigures would improve clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the central claim is an empirical prediction from external PCET labels, not a quantity defined by the EEG/ECG inputs.

full rationale

The paper's only formal chain is an empirical pipeline: raw EEG/ECG and derived features are fed into a CNN-Transformer, and the output is a binary PCET class. There is no equation-level derivation in which the target is defined in terms of the input features, no fitted parameter that is renamed as a prediction, and no load-bearing self-citation. The median-split PCET threshold in Section II-B is computed on the full 817-subject dataset, so labels in each training fold depend on test-fold scores; and Table III reports the best of several architectures evaluated on the same 10-fold splits. These are model-selection and target-definition leaks that risk inflating the 80.3% figure, but they do not make the prediction mathematically equivalent to the inputs: the PCET score is an external cognitive test, and the model could fail. The limitations section acknowledges feature/channel restrictions, black-box interpretability, and missing demographic and sleep-stage annotations, but does not claim a first-principles derivation. No self-citation is load-bearing; reference [14] is an external baseline and [15]-[18] are standard dataset/test citations. Hence no circular step is exhibited, and the correct score is 0. The generalization concerns should be pursued as correctness/statistical-validity issues, not circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

No new physical entities, particles, forces, or dimensions are postulated. The only new element is the neural network architecture, which is a computational model, not an invented entity in the scientific sense. The central claim rests on preprocessing choices, label construction, and architecture selection decisions that are not fully independent of the evaluation data.

free parameters (4)

PCET median split threshold = not reported
Defines low vs high labels from the full STAGES cohort median; threshold is not disclosed and test-fold scores contribute to it in cross-validation.
Uniform sleep segment length = 5 hours
Selected so all sequences have the same length; arbitrary choice may discard later-night REM and NREM periods relevant to cognition.
Model hyperparameters = lr=1e-4; 3 raw and 4 feature transformer layers; 4 heads; raw FFN 92, feature FFN 8; raw dim 64, feature dim 8…
Selected on a validation subset per fold, but the architecture itself was chosen after comparing test-fold results, so the reported configuration is optimistic.
Preprocessing thresholds = EEG 0.3-35 Hz, ECG 0.5-35 Hz; heart rate 40-100 bpm; YASA artifact mask
Hand-chosen filter and exclusion criteria can bias the cohort; their influence on accuracy is not analyzed.

assumptions (6)

domain assumption STAGES PSG recordings from six centers are comparable after channel re-referencing and filtering.
Used to pool data across sites (Section II-A); site effects could confound sleep-cognition relationships.
domain assumption PCET Concept Level Responses is a valid measure of cognitive flexibility and conceptual reasoning, and a median split into low/high is a meaningful target.
The binary target is defined in Section II-B; if the median split or test validity fails, the classification task is not well posed.
domain assumption A single EEG channel (C3-M2) and a single ECG channel capture the sleep physiology most relevant to executive function.
Chosen based on prior literature (Section II-A); other channels may carry predictive information and their exclusion could bias results.
domain assumption Cognitive performance measured either before or after the PSG night is stable enough to be predicted from that night's sleep.
PCET was administered the evening before or morning after the sleep study (Section II-B); if the night affects only the morning group, pooling mixes predictive directions.
ad hoc to paper A uniform 5-hour segment preserves the predictive sleep microstructure.
Section II-C selects 5 hours for sequence-length consistency without analyzing whether the discarded portion of the night matters for cognition.
domain assumption The convolutional-transformer training procedure, including subject-based splits and z-score normalization, avoids information leakage beyond the median-split issue.
Subject-based splitting is standard, but the full-dataset median threshold creates a separate leakage path.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Sleep Brain and Cardiac Activity Predict Cognitive Flexibility and Conceptual Reasoning Using Deep Learning." pith.science (2026). https://pith.science/paper/FK56CYCT

@misc{pith2026250600279,
  author       = {Pith},
  title        = {Pith review of: Sleep Brain and Cardiac Activity Predict Cognitive Flexibility and Conceptual Reasoning Using Deep Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FK56CYCT}},
  note         = {Machine review of arXiv:2506.00279}
}

read the original abstract

Despite extensive research on the relationship between sleep and cognition, the connection between sleep microstructure and human performance across specific cognitive domains remains underexplored. This study investigates whether deep learning models can predict executive functions, particularly cognitive adaptability and conceptual reasoning from physiological processes during a night's sleep. To address this, we introduce CogPSGFormer, a multi-scale convolutional-transformer model designed to process multi-modal polysomnographic data. This model integrates one-channel ECG and EEG signals along with extracted features, including EEG power bands and heart rate variability parameters, to capture complementary information across modalities. A thorough evaluation of the CogPSGFormer architecture was conducted to optimize the processing of extended sleep signals and identify the most effective configuration. The proposed framework was evaluated on 817 individuals from the STAGES dataset using cross-validation. The model achieved 80.3\% accuracy in classifying individuals into low vs. high cognitive performance groups on unseen data based on Penn Conditional Exclusion Test (PCET) scores. These findings highlight the effectiveness of our multi-scale feature extraction and multi-modal learning approach in leveraging sleep-derived signals for cognitive performance prediction. To facilitate reproducibility, our code is publicly accessible (https://github.com/boshrakh95/CogPSGFormer.git).

Figures

Figures reproduced from arXiv: 2506.00279 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages

[1]

Sleep and cognition,

M. C. Deak and R. Stickgold, “Sleep and cognition,” Wiley Inter- disciplinary Reviews: Cognitive Science, vol. 1, no. 4, pp. 491–500, 2010

work page 2010
[2]

Impairment of executive functions due to sleep alterations: An integrative review on the use of P300,

N. C. Lima et al., “Impairment of executive functions due to sleep alterations: An integrative review on the use of P300,” Frontiers in Neuroscience, vol. 16, p. 906492, 2022

work page 2022
[3]

Symbolic dynamics of sleep heart rate variability is associated with cognitive decline in older men,

S. A. Immanuel et al., “Symbolic dynamics of sleep heart rate variability is associated with cognitive decline in older men,” in Proc. 45th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), 2023, pp. TBD

work page 2023
[4]

Cognitive performance, sleepiness, and mood in partially sleep deprived adolescents: the need for sleep study,

J. C. Lo et al., “Cognitive performance, sleepiness, and mood in partially sleep deprived adolescents: the need for sleep study,” Sleep, vol. 39, no. 3, pp. 687–698, 2016

work page 2016
[5]

Rapid eye movement sleep, sleep continuity, and slow wave sleep as predictors of cognition, mood, and subjective sleep quality in healthy men and women, aged 20–84 years,

C. Della Monica et al., “Rapid eye movement sleep, sleep continuity, and slow wave sleep as predictors of cognition, mood, and subjective sleep quality in healthy men and women, aged 20–84 years,” Frontiers in Psychiatry, vol. 9, p. 255, 2018

work page 2018
[6]

Macro and micro sleep architecture and cognitive performance in older adults,

I. Djonlagic et al., “Macro and micro sleep architecture and cognitive performance in older adults,” Nature Human Behaviour, vol. 5, no. 1, pp. 123–145, 2021

work page 2021
[7]

Relationships between sleep stages and changes in cognitive function in older men: the MrOS Sleep Study,

“Relationships between sleep stages and changes in cognitive function in older men: the MrOS Sleep Study,” Sleep, vol. 38, no. 3, pp. 411– 421, 2015

work page 2015
[8]

Memory and sleep: How sleep cognition can change the waking mind for the better,

K. A. Paller et al., “Memory and sleep: How sleep cognition can change the waking mind for the better,” Annual Review of Psychology, vol. 72, pp. 123–150, 2021

work page 2021

Show all 18 references

[9]

Overnight sleep loss and ’executive’ decision mak- ing—subtle findings,

J. Horne, “Overnight sleep loss and ’executive’ decision mak- ing—subtle findings,” Sleep, vol. 36, no. 6, pp. 823–824, 2013

2013
[10]

Sleep duration and executive function in adults,

A. Sen and X. Y . Tai, “Sleep duration and executive function in adults,” Current Neurology and Neuroscience Reports, 2023

2023
[11]

The effects of sleep quality on response inhibi- tion,

N. H. Abbas et al., “The effects of sleep quality on response inhibi- tion,” Young Anthropology, vol. 2, pp. 10–16, 2020

2020
[12]

Neural effects of sleep deprivation on inhibitory control and emotion processing,

J. R. Magnuson et al., “Neural effects of sleep deprivation on inhibitory control and emotion processing,” Behavioural Brain Research, vol. 426, p. 113845, 2022

2022
[13]

Sleep quality and aging: a systematic review on healthy older people, mild cognitive impairment and Alzheimer’s disease,

M. Casagrande et al., “Sleep quality and aging: a systematic review on healthy older people, mild cognitive impairment and Alzheimer’s disease,” International Journal of Environmental Research and Public Health, vol. 19, no. 14, p. 8457, 2022

2022
[14]

A transformer model for predicting cognitive impairment from sleep,

T.-A. Song et al., “A transformer model for predicting cognitive impairment from sleep,” bioRxiv, 2022

2022
[15]

The National Sleep Research Resource: towards a sleep data commons,

G.-Q. Zhang et al., “The National Sleep Research Resource: towards a sleep data commons,” Journal of the American Medical Informatics Association, vol. 25, no. 10, pp. 1351–1358, 2018

2018
[16]

The relationship between heart rate variability and electroencephalography functional connectivity variability is as- sociated with cognitive flexibility

Alba, Guzmán, et al. "The relationship between heart rate variability and electroencephalography functional connectivity variability is as- sociated with cognitive flexibility." Frontiers in Human Neuroscience 13 (2019): 428262

2019
[17]

The Penn Conditional Exclusion Test: a new measure of executive-function with alternate forms for repeat admin- istration,

M. M. Kurtz et al., “The Penn Conditional Exclusion Test: a new measure of executive-function with alternate forms for repeat admin- istration,” Archives of Clinical Neuropsychology, vol. 19, no. 2, pp. 191–201, 2004

2004
[18]

An EEG study on the somatotopic organisation of sensorimotor cortex activation during action execution and observation in infancy,

C. de Klerk et al., “An EEG study on the somatotopic organisation of sensorimotor cortex activation during action execution and observation in infancy,” Developmental Cognitive Neuroscience, vol. 15, pp. 1–10, 2015

2015

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Sleep and cognition,

M. C. Deak and R. Stickgold, “Sleep and cognition,” Wiley Inter- disciplinary Reviews: Cognitive Science, vol. 1, no. 4, pp. 491–500, 2010

work page 2010

[2] [2]

Impairment of executive functions due to sleep alterations: An integrative review on the use of P300,

N. C. Lima et al., “Impairment of executive functions due to sleep alterations: An integrative review on the use of P300,” Frontiers in Neuroscience, vol. 16, p. 906492, 2022

work page 2022

[3] [3]

Symbolic dynamics of sleep heart rate variability is associated with cognitive decline in older men,

S. A. Immanuel et al., “Symbolic dynamics of sleep heart rate variability is associated with cognitive decline in older men,” in Proc. 45th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), 2023, pp. TBD

work page 2023

[4] [4]

Cognitive performance, sleepiness, and mood in partially sleep deprived adolescents: the need for sleep study,

J. C. Lo et al., “Cognitive performance, sleepiness, and mood in partially sleep deprived adolescents: the need for sleep study,” Sleep, vol. 39, no. 3, pp. 687–698, 2016

work page 2016

[5] [5]

Rapid eye movement sleep, sleep continuity, and slow wave sleep as predictors of cognition, mood, and subjective sleep quality in healthy men and women, aged 20–84 years,

C. Della Monica et al., “Rapid eye movement sleep, sleep continuity, and slow wave sleep as predictors of cognition, mood, and subjective sleep quality in healthy men and women, aged 20–84 years,” Frontiers in Psychiatry, vol. 9, p. 255, 2018

work page 2018

[6] [6]

Macro and micro sleep architecture and cognitive performance in older adults,

I. Djonlagic et al., “Macro and micro sleep architecture and cognitive performance in older adults,” Nature Human Behaviour, vol. 5, no. 1, pp. 123–145, 2021

work page 2021

[7] [7]

Relationships between sleep stages and changes in cognitive function in older men: the MrOS Sleep Study,

“Relationships between sleep stages and changes in cognitive function in older men: the MrOS Sleep Study,” Sleep, vol. 38, no. 3, pp. 411– 421, 2015

work page 2015

[8] [8]

Memory and sleep: How sleep cognition can change the waking mind for the better,

K. A. Paller et al., “Memory and sleep: How sleep cognition can change the waking mind for the better,” Annual Review of Psychology, vol. 72, pp. 123–150, 2021

work page 2021