Exploratory Integration of EEG Spectral Features and Gaze Variability for Mild Cognitive Impairment Discrimination

Hisatomo Kowa; Mamoru Hiroe; Minoru Nakayama; Takashi Nagamatsu; Takeru Mukunoki; Yujia Zheng; Yuma Sonoda

REVIEW 3 major objections 4 minor 2 references

This paper claims that combining a small set of EEG spectral features with gaze variability substantially improves discrimination of mild cognitive impairment, raising AUC from 0.64 to 0.78 in a preliminary study.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 05:54 UTC pith:UXEWI6LQ

load-bearing objection Genuinely exploratory multimodal pilot whose headline AUCs are undermined by feature-selection leakage—the authors disclose it, but the gaze-complementarity claim is not yet established. the 3 major comments →

arxiv 2607.29493 v1 pith:UXEWI6LQ submitted 2026-07-31 cs.HC

Exploratory Integration of EEG Spectral Features and Gaze Variability for Mild Cognitive Impairment Discrimination

Takeru Mukunoki , Mamoru Hiroe , Minoru Nakayama , Yujia Zheng , Yuma Sonoda , Hisatomo Kowa , Takashi Nagamatsu This is my paper

classification cs.HC

keywords mild cognitive impairmentEEG spectral featuresgaze variabilityeye trackingLASSOROC AUCfixation stabilitymultimodal biomarkers

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that combining EEG spectral power features with gaze variability during a fixation task improves discrimination between mild cognitive impairment (MCI) and healthy aging. In a stepwise comparison on 38 participants, all EEG features alone gave an AUC of 0.52, two LASSO-selected EEG features gave 0.64, and adding two gaze variability features raised the AUC to 0.78. The authors interpret this jump as evidence that neural and behavioral variability measures carry complementary information. However, because feature selection was performed on the full dataset before cross-validation, the authors themselves caution that the reported AUCs are preliminary upper-bound estimates.

Core claim

On the paper's own terms, the central claim is that dimensionality reduction of EEG features followed by integration of gaze variability yields a substantial gain in MCI discrimination. Specifically, delta-band power at electrode O1 and low-beta power at T6, selected via LASSO and concatenated with the standard deviations of the x-coordinate of gaze for the right and left eyes, produced a leave-one-out cross-validated AUC of 0.78, compared to 0.64 for the EEG features alone and 0.52 for the full EEG feature set. The authors suggest that the occipital delta and temporal low-beta rhythms reflect visual processing and attentional control, while gaze instability reflects a related but distinct b

What carries the argument

The central machinery is a stepwise classification pipeline: (1) high-dimensional EEG spectral power features across six frequency bands and all 10-20 electrodes; (2) L1-regularized logistic regression (LASSO) applied to the full dataset to select the two most contributive EEG features; (3) a logistic regression classifier with balanced class weights, trained under leave-one-out cross-validation; and (4) aggregation of held-out probabilities to compute the ROC AUC. The gaze features are the standard deviations of gaze position in the x-direction for the right and left eyes (R_std_x, L_std_x), computed after excluding saccade intervals, serving as fixation-stability measures.

Load-bearing premise

The load-bearing premise is that the leave-one-out cross-validated AUCs are valid out-of-sample estimates even though the two EEG features were selected using the full dataset before cross-validation; the paper itself states these should be interpreted as preliminary upper-bound estimates.

What would settle it

Re-run the analysis with nested cross-validation or a completely held-out test set, performing LASSO feature selection only on the training folds each time. If the AUC for EEG-plus-gaze drops to near 0.64 or lower, the claim that gaze variability adds complementary information would be falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

If the result holds, multimodal indicators combining neural and behavioral variability could improve non-invasive MCI screening over either modality alone.
The specific selected features (occipital delta, temporal low-beta, gaze x-variability) point to concrete physiological targets for future validation.
The study suggests that even minimal gaze features, captured during a simple fixation task, can add value to EEG-based classification.
The stepwise improvement from 0.52 to 0.64 to 0.78 implies that feature selection and multimodal integration are both beneficial, not just one.
The approach is portable: the same pipeline could be applied to other cognitive tasks or larger cohorts to test generalizability.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the gaze gain survives proper nested cross-validation, the complementarity between neural and behavioral variability might generalize to other cognitive assessments (e.g., memory or attention tasks), not just fixation.
The paper reports only x-axis gaze standard deviations in the final model even though both x and y were computed; y-axis variability might add further discrimination or substitute when x is unavailable.
The unseparated feature selection makes the 0.64-to-0.78 jump uncertain; a re-analysis with selection performed inside each cross-validation fold is the direct next test.
The proposed mechanism — that neural and gaze variability reflect different aspects of a common visual-attention process — is testable by correlating the two feature sets with independent cognitive scores.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Genuinely exploratory multimodal pilot whose headline AUCs are undermined by feature-selection leakage—the authors disclose it, but the gaze-complementarity claim is not yet established.

read the letter

Quick take: this is a 3-page ETRA pilot combining EEG spectral features with gaze variability to discriminate MCI from healthy controls (n=38). The specific combination is new relative to the papers it cites, and the writing is clear and appropriately modest. The catch is that the headline AUC sequence—0.52, 0.64, 0.78—is computed with feature-selection leakage. The authors ran LASSO on the full dataset to pick two EEG features, then did LOOCV on those fixed features, so the left-out participant's label was visible during feature selection. They disclose this in §3.4, calling the AUCs 'preliminary upper-bound estimates.' Credit for honesty, but it doesn't repair the statistic: the incremental gain attributed to gaze integration (0.64→0.78) is not a clean out-of-sample result.

What's genuinely good: the empirical setup is new as far as the cited references show, and the interpretive frame—neural variability plus gaze variability may tap complementary processes—is plausible. The task is simple (4 seconds of fixation from a 20-second trial), so the approach could translate to a low-cost screening tool if the effect is real. The authors also keep their claims in check; the discussion lists small sample size, lack of external validation, and restricted gaze features as limitations.

The soft spot is the one above, and it's load-bearing. The solution is standard: perform feature selection inside each CV fold (or use a nested CV), and ideally validate on a separate cohort. A secondary, minor issue is that the gaze features R_std_x/L_std_x are chosen without explanation, while §3.3.2 computes both x and y standard deviations; that choice is also not cross-validated. The paper also omits some details needed for reproduction (LASSO penalty, exact band boundaries, channel count), which matters more in a 3-page format.

Who should read this? Anyone working on digital biomarkers for MCI, mainly as a prompt to think about multimodal integration and about how easily CV can leak. It is not a reliable source for effect sizes.

My recommendation: send it to peer review with a specific request for a nested-CV re-analysis. The qualitative hypothesis is worth testing, the authors have been transparent, and a re-analysis could turn it into a solid pilot. As is, the numbers should be treated as preliminary upper bounds—which is exactly what the authors say.

Referee Report

3 major / 4 minor

Summary. The paper reports an exploratory comparison of three classification models for discriminating mild cognitive impairment (MCI) from healthy aging using EEG spectral power and gaze variability during a fixation task. Model (a) uses all EEG spectral-band features; model (b) uses two EEG features selected by LASSO (delta power at O1, low-beta power at T6); model (c) adds two gaze variability features (R_std_x, L_std_x). Classification is logistic regression with balanced class weights, evaluated by leave-one-out cross-validation (LOOCV) and AUC. Reported AUCs are 0.52, 0.64, and 0.78, respectively, and the authors interpret the stepwise increase, especially 0.64→0.78, as evidence that neural and behavioral variability measures carry complementary information.

Significance. If the reported effect is real, the combination of simple EEG spectral features and gaze variability could offer a low-cost, objective digital biomarker for early MCI screening. The idea that neural and behavioral variability during fixation tap complementary processes is plausible and clinically relevant. However, the central quantitative claim is not established by the current analysis because the headline AUCs are computed under feature-selection leakage, and the paper explicitly labels them as upper-bound estimates. The manuscript is honest about this limitation, but the admission does not repair the statistic. A properly nested cross-validation or feature-selection-within-training-folds design, plus uncertainty quantification, would be needed before the complementarity claim can be assessed.

major comments (3)

[§3.4 and §3.3.1] Feature-selection leakage invalidates the reported LOOCV AUCs. LASSO was applied to the full dataset to select the two EEG features before any cross-validation split, so the label of the left-out participant was visible at feature-selection time. The LOOCV evaluation of the fixed two-feature model therefore partially measures the fit of feature selection to the test participant, not out-of-sample generalization. The paper's own §3.4 acknowledges this ('preliminary upper-bound estimates'), but the magnitude of the bias is unknown, so the stepwise comparison (0.52→0.64→0.78) is not a reliable estimate of discrimination. The authors should rerun the analysis with feature selection nested inside each training fold (or use a fully pre-specified feature set) and report AUC with confidence intervals.
[§3.4 and §3.3.2] The choice of gaze features is not justified and is not cross-validated. §3.3.2 computes standard deviations in both x and y directions for each eye, but §3.4 selects only R_std_x and L_std_x for model (c). No rationale is given for discarding the y-direction features, and there is no indication whether this choice was made before seeing the outcome. If the gaze-feature subset was chosen after inspecting full-data results, the same leakage concern applies. At minimum, the feature set should be pre-specified in the methods, and the selection rule should be applied within cross-validation folds.
[§4] No statistical inference accompanies the AUC comparisons. With only 38 participants, the difference between AUC = 0.64 and AUC = 0.78 may be within sampling variability, especially given the small number of MCI cases (n = 17). The authors should report confidence intervals for each AUC and a test comparing models (e.g., DeLong's test or bootstrap). Without these, the phrase 'suggesting a substantial improvement in classification performance' is unsupported.

minor comments (4)

[§3.3.1] The paper states LASSO identified the 'two most contributive features' but does not specify how the regularization strength was chosen. This matters because the number and identity of selected features depends on the path of λ; please report the criterion (e.g., cross-validated λ, fixed λ) and the coefficients.
[§4] A gaze-only model is mentioned with AUC = 0.55, but this model is not described in the Methods. Its feature set, evaluation procedure, and whether it was subject to the same leakage issue should be specified if it is included in the stepwise comparison.
[§3.2] The choice to analyze only the first 4 seconds of the 20-second trial is not justified. If this was an a priori decision, it should be stated; if it was data-driven, the implications for the reported AUCs should be discussed.
[§3.3.1] Re-referencing by subtracting regional reference electrodes for different topographies is described briefly; please clarify whether this is a standard procedure or an ad hoc choice, and whether it affects comparability across channels.

Circularity Check

1 steps flagged

LASSO feature selection on the full dataset makes the LOOCV AUCs—especially the EEG 0.64 and EEG+gaze 0.78—partly in-sample estimates; the paper discloses this but does not correct it.

specific steps

fitted input called prediction [§3.3.1 EEG Features; §3.4 Classification Analysis; §4 Results]
"Subsequently, L1 regularization (LASSO) was applied as an exploratory analysis using the full dataset to identify the two most contributive features. These selected features were used in models (b) and (c). ... Feature selection using L1 regularization (LASSO) was conducted exploratorily using the full dataset, and the selected features were used in models (b) and (c). Because feature selection and performance evaluation were not fully separated, the reported AUC values should be interpreted as preliminary upper-bound estimates."

The features used in models (b) and (c) were selected by LASSO on the complete dataset before LOOCV splits. Thus in each fold the left-out participant's label was already available at feature-selection time. The LOOCV AUCs of 0.64 and 0.78 therefore measure, in part, how well the feature-selection step fit the whole sample, not purely out-of-sample generalization. The paper's own 'upper-bound estimates' caveat admits this leak. Consequently the reported improvement from 0.64 to 0.78 attributed to gaze integration is not a clean out-of-sample result, so the central complementarity claim is not quantitatively established by these numbers.

full rationale

The only substantive circular step is the evaluation loop created by applying LASSO to the full dataset and then reporting LOOCV AUCs for the preselected features. This is a selection-leakage/specification-search form of fitted-input-called-prediction: the 'test' participant's label influenced which two EEG features entered the model, so the later cross-validated probabilities are not fully out-of-sample. The paper explicitly acknowledges 'feature selection and performance evaluation were not fully separated' and asks that the AUCs be read as 'preliminary upper-bound estimates'; that honesty does not remove the circularity. No self-citation chain, uniqueness argument, or renamed known result is load-bearing. The qualitative hypothesis of EEG/gaze complementarity may survive, but the reported stepwise AUC comparison (0.52 → 0.64 → 0.78) cannot be taken as clean evidence for it.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper's numbers rest on a small number of fitted or unstated choices and ordinary domain assumptions. The two EEG features are fitted by LASSO on the full labeled dataset (leaked); the gaze features R_std_x/L_std_x are chosen without a stated criterion; the 4 s analysis window and the band definitions are unstated choices; MoCA is assumed as ground truth and a single 20 s trial is assumed to yield stable estimates. No invented entities are introduced; 'complementarity' is an interpretation of measured features, not a postulated mechanism.

free parameters (5)

LASSO regularization strength (λ)
Not reported in §3.4; determines which two EEG features are selected, hence the entire structure of models (b) and (c).
Selected EEG features (O1 delta, T6 low-beta) = delta at O1, low-beta at T6
Chosen by LASSO fit on the full dataset including test-fold outcomes (§3.4); the selection is a fitted quantity that leaks into the LOOCV AUCs.
Gaze feature set (R_std_x, L_std_x) = R_std_x, L_std_x
Only the x-direction gaze standard deviations are used in model (c) (§3.4), although §3.3.2 computed both x and y; no criterion is given for this choice.
Analysis window (first 4 s of 20 s trial) = 4 s
§3.2: choice stated without justification; affects all spectral and gaze statistics.
Frequency band boundaries
§3.3.1 lists delta, theta, alpha, low-beta, high-beta, gamma without cutoff frequencies; bands are standard but the exact boundaries are unspecified modeling choices.

axioms (4)

domain assumption MoCA threshold is accepted as ground truth for MCI vs. control group assignment
§3.1: 'Group classification was based on the Montreal Cognitive Assessment (MoCA) score'; no clinical adjudication or neuropsychological battery described.
domain assumption A single 20 s fixation trial (first 4 s analyzed) yields EEG PSD and gaze-variance estimates stable enough for cross-sectional discrimination
§3.2/§3.3; no test-retest or reliability analysis; single trial per participant.
domain assumption LOOCV logistic regression on n=38 yields interpretable AUC estimates
§3.4; violated as stated by the authors because feature selection is performed on the full dataset before folding.
standard math Welch's PSD and the regional re-referencing scheme as implemented produce comparable features
§3.3.1; standard signal processing, but the regional reference subtraction (Fz/C3/C4/Pz) is an unusual scheme with no validation cited.

pith-pipeline@v1.3.0-daily-deepseek · 3437 in / 18315 out tokens · 189036 ms · 2026-08-03T05:54:46.722905+00:00 · methodology

0 comments

read the original abstract

Early detection of mild cognitive impairment (MCI) is an important challenge in aging societies. Electroencephalography (EEG) and eye-tracking have independently been explored as potential biomarkers; however, their integrative effects remain insufficiently examined. This exploratory study investigated whether combining EEG spectral features with gaze variability may provide complementary information for MCI discrimination. EEG signals were recorded using the 10--20 system, and spectral power features were extracted. We compared three models: (a) high-dimensional EEG features, (b) L1-regularized feature selection (LASSO), and (c) integration of the selected EEG features with gaze variability. Performance was evaluated using leave-one-out cross-validation and area under the ROC curve (AUC). Model (a) yielded limited discrimination (AUC = 0.52). Feature selection increased AUC (0.64), and additional integration of gaze variability further increased AUC (0.78). These preliminary findings suggest potential complementarity between neural and behavioral variability measures.

Figures

Figures reproduced from arXiv: 2607.29493 by Hisatomo Kowa, Mamoru Hiroe, Minoru Nakayama, Takashi Nagamatsu, Takeru Mukunoki, Yujia Zheng, Yuma Sonoda.

discussion (0)

Reference graph

Works this paper leans on

2 extracted references

[2023]

Psychol.14 (July 2023), 1197567

Eye-tracking paradigms for the assessment of mild cognitive impairment: a systematic review.Front. Psychol.14 (July 2023), 1197567. https://doi.org/10.3389/ fpsyg.2023.1197567

arXiv 2023
[2025]

Fusion122, 103202 (Oct

Eye tracking based detection of mild cognitive impairment: A review.Inf. Fusion122, 103202 (Oct. 2025), 103202. https://doi.org/10.1016/j.inffus.2025.103202 Alexandra Wolf, Kornkanok Tripanpitak, Satoshi Umeda, and Mihoko Otake-Matsuura

arXiv 2025

[1] [2023]

Psychol.14 (July 2023), 1197567

Eye-tracking paradigms for the assessment of mild cognitive impairment: a systematic review.Front. Psychol.14 (July 2023), 1197567. https://doi.org/10.3389/ fpsyg.2023.1197567

arXiv 2023

[2] [2025]

Fusion122, 103202 (Oct

Eye tracking based detection of mild cognitive impairment: A review.Inf. Fusion122, 103202 (Oct. 2025), 103202. https://doi.org/10.1016/j.inffus.2025.103202 Alexandra Wolf, Kornkanok Tripanpitak, Satoshi Umeda, and Mihoko Otake-Matsuura

arXiv 2025