{"id":"eabc51c7-72ba-40f3-9694-cb3a36fe81e9","arxiv_id":"2411.13931","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The authors release a Fieldtrip implementation of three EOG-correlation-based ICA artifact removal methods with example code and single-subject illustrations.","lead":"This paper describes and releases Matlab/Fieldtrip code for three ICA-based methods that reduce eye-movement artifact contamination in EEG recordings. The tools are aimed at researchers who need to keep long trials that cannot simply be discarded.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The artifact-reduction claim rests on visual single-subject examples and a self-referential correlation metric; no independent quantitative validation is reported.","rationale":"The reader's weakest assumption concerns MSF correctness; that is a real sub-risk for Methods 2 and 3, but it is not the most load-bearing condition for the central claim because the paper explicitly presents MSF creation as replaceable and the central claim concerns artifact reduction. The necessary condition for the claim is a demonstrated reduction of ocular artifact content with preservation of brain signals. No such demonstration appears: Figure 2 is circular, and Figures 1, 3 and 4 are single-subject visual inspections. The code and data release are valuable and provide real support for reproducibility, but they do not by themselves establish effectiveness. I therefore keep the reader's CONDITIONAL verdict while locating the decisive gap in outcome validation rather than in MSF parameter choice. A concrete validation study on the provided data, with expert-annotated artifact intervals and a signal-preservation metric, would settle whether the concern actually lands.","tokens_in":11451,"tokens_out":5000,"duration_ms":54926,"concrete_test":"Run the supplied code on the released dataset (and, if possible, additional subjects recorded under the same protocol); have two annotators independently mark blink and saccade intervals from VEOG/HEOG, then compute (i) peak-to-peak amplitude or EOG-regression residual of Fp1/Fz inside annotated artifact intervals before versus after each of the three methods, and (ii) correlation or RMS difference between cleaned and raw EEG in artifact-free intervals as a signal-preservation proxy. If (i) does not decrease substantially while (ii) stays small, the artifact-reduction claim fails; if (ii) is large, over-removal is indicated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the delivered functions reduce ocular-artifact influence in EEG. The only evidence offered is Figures 1, 3 and 4 for a single subject (001) plus Figure 2, which shows the sum of |correlation| between ICs and EOG before and after. Figure 2 is tautological with respect to the claim: the method selects components precisely to minimize that correlation, so a decrease is guaranteed by construction and does not establish that ocular artifacts in the reconstructed EEG channels are reduced or that non-artifact brain activity is preserved. Section 5 even notes amplitude decreases in non-artifact regions after Method 1, which is consistent with over-removal. The automatic MSF used in Methods 2 and 3 (ft_artifact_zvalue, cut-off 0.5, 2-15 Hz, padding 0.1; Section 4.5) is likewise never compared with expert annotation, so partial removal may excise non-artifact samples. Because these gaps concern the evidence for the claimed effect, not merely the exposition, the claim to reduce the influence of ocular artifacts is not yet supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript describes and releases Matlab/Fieldtrip implementations of three methods, originally from [14], for reducing ocular artifacts in EEG: (1) complete removal of ICA components most correlated with EOG channels; (2) partial removal of those components only at samples flagged by an artifact membership function (MSF); and (3) partial removal using an unmixing matrix estimated from artifact-free data. It also provides a function to build the MSF using ft_artifact_zvalue, sample scripts (Boxes 1-3), and example results for one subject (001). The central claim is that these implementations lessen the influence of ocular artifacts and that the code and data allow exact replication.","tokens_in":11663,"tokens_out":4212,"duration_ms":40741,"significance":"The paper addresses a real need identified in [13]: lack of open-source artifact-removal implementations. Its strengths are the detailed function-by-function description, the release of data and code at a persistent handle, and the attempt to keep the MSF format simple and replaceable. If the methods were quantitatively validated, the contribution would be a useful reference implementation for practitioners. However, the current evidence for effectiveness is limited to visual inspection of two channels in one subject and a correlation metric that is partly self-referential; the paper therefore does not yet substantiate its central claim.","major_comments":[{"comment":"The only quantitative evidence is the reduction in the sum of absolute correlations between ICs and EOG before and after processing. Because the algorithm selects components precisely to minimize this quantity, the decrease is expected by construction and does not independently show that ocular artifacts in EEG channels are reduced or that brain signals are preserved. I recommend adding validation against expert-labeled artifact intervals, a downstream EEG analysis endpoint, or comparison with an established method such as ADJUST, MARA, or ASR, with quantitative metrics such as artifact attenuation, signal-to-error ratio, or preservation of known evoked responses.","section":"Section 5, Figure 2"},{"comment":"The text states that in some areas not influenced by artifacts there is a decrease in amplitude with respect to the original shape of the signal. This indicates possible over-removal, but it is not quantified or compared across methods. Since the paper claims lessening artifact influence, it must report the trade-off between artifact energy removed and non-artifact signal distortion, for example by computing RMS error or spectral error on artifact-free segments, or by showing preservation of known neural responses.","section":"Section 5, Method 1"},{"comment":"The automatic MSF created by fta_create_msf_fieldtrip uses ft_artifact_zvalue with a cut-off of 0.5, a 2-15 Hz band-pass, and padding of 0.1. These parameters are asserted without validation. Methods 2 and 3 depend entirely on this MSF, so the paper should report the sensitivity of the results to these choices and, ideally, compare the MSF against expert annotation or another artifact detector, such as those in [20-22], in terms of detection accuracy.","section":"Section 4.5"},{"comment":"All examples are from a single subject (001) and a single stimulus ('truck'). No across-subject statistics, error bars, or reproducibility checks are provided. For a software/tool paper, a small multi-subject demonstration with quantitative summary measures is necessary to support the claim that the implementations generally lessen artifact influence rather than merely working on one selected example.","section":"Section 5"}],"minor_comments":[{"comment":"The phrase 'describes and implementation of scripts of code' should be corrected to 'describes an implementation of scripts of code' or 'describes implementations of scripts of code'.","section":"Abstract"},{"comment":"The units '10 KQ' should be '10 kΩ', and the same correction should be applied throughout the equipment description.","section":"Section 2.1"},{"comment":"The term 'modulus of correlation coefficients' should be replaced with 'absolute value' or 'magnitude' for clarity and consistency with standard statistical terminology.","section":"Section 4.1, step 4"},{"comment":"The example code contains typos and inconsistencies with the function signatures described in the text, such as 'clean.data' instead of 'cleaned_data', 'cfg.hpfilttype -' instead of 'cfg.hpfilttype =', and 'Corretion' instead of 'Correction'. These should be fixed so that the examples run exactly as printed.","section":"Boxes 1-3"},{"comment":"Function names are rendered inconsistently, for example 'fta_select_compjrom_coeff_correlation' versus 'fta_select_compjrom_coeffcorrelation' and 'fta_channels_jrom_ica_reject_components' versus 'fta_channels_from_ica_reject_components'. The text should use exactly the names that appear in the released code to avoid confusion for readers attempting to replicate the pipeline.","section":"Sections 4.1-4.9"},{"comment":"The captions for Figures 1 and 3 use 'Fpl', but the standard and presumably intended label is 'Fp1'.","section":"Figure captions"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable software-description contribution, but the central artifact-reduction claim needs quantitative validation before it can be accepted as stated. The most serious issue is the self-referential nature of Figure 2; the authors should be required to add independent evaluation, either against expert annotations, a downstream task, or a benchmark method. If such validation is not feasible, the authors should explicitly downgrade the claim to a software implementation with illustrative examples, and the title and abstract should be adjusted accordingly. The code and data availability statement should also be verified at revision time."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an engineering companion to the authors' earlier method paper [14]. The genuine contribution is a set of documented Matlab/Fieldtrip functions for three variants of EOG-correlation-based ICA artifact removal, plus an automatic artifact membership function builder, all on a public repository. That fills a real practical gap: EEG researchers who keep long trials often need open implementations rather than ad hoc scripts. The paper does a good job of step-by-step API documentation and makes the code and one subject's data available.\n\nWhat it does not do is provide a quantitative validation of the artifact-reduction claim. The evidence is visual inspection of one subject (subject 001) and Figure 2, which shows the sum of |correlation| between ICs and EOG before/after removal. That metric is exactly what the algorithm minimizes, so the drop is by construction. There is no independent error measure, no expert annotation, no downstream-task evaluation, no baseline comparison with, say, ADJUST or MARA. The paper itself acknowledges amplitude decreases in non-artifact regions after Method 1, which is consistent with over-removal. Methods 2 and 3 rely on the MSF built by ft_artifact_zvalue with cut-off 0.5, 2-15 Hz band-pass, and padding 0.1; those parameters are never validated against ground truth. So the effectiveness claim is plausible but not established.\n\nThat said, the authors are fairly careful in language: they say 'illustrative examples' and 'expected results,' and they explicitly invite users to substitute other artifact detectors and selection criteria. The code is the deliverable, and for a software descriptor paper that's a reasonable framing. The lack of quantitative evaluation is a soft spot, not a fatal flaw, as long as readers treat the figures as demonstrations rather than proof.\n\nWho's this for? Applied EEG researchers who want a working Fieldtrip starting point for EOG-artifact reduction and are comfortable adapting parameters. Not for methodologists. The paper deserves a serious referee because the software release is checkable and useful, but the referee should push for either clearer effectiveness claims or an evaluation protocol.","headline":"A useful open-source Fieldtrip implementation of EOG-artifact ICA methods that honestly demonstrates rather than validates itself, with the effectiveness claim resting on visual inspection and a tautological metric.","tokens_in":12170,"tokens_out":1904,"would_cite":false,"duration_ms":19935,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper provides ready-to-run scripts that reduce ocular artifacts in EEG by removing or attenuating independent components correlated with EOG signals.","keywords":["independent component analysis","ocular artifact removal","electro-oculogram","artifact membership function","EEG preprocessing","Matlab","Fieldtrip"],"falsifier":"Take an EEG recording with simultaneous video-based eye tracking or expert-annotated blink and saccade intervals, build the membership function with the provided defaults, and compare its marked samples against the annotations; if many real artifacts go unmarked or many non-artifact segments are marked, the partial-removal methods will either leave artifacts in or attenuate brain responses. A quantitative version would add simulated ocular components with known timing into clean EEG and check whether the partial-removal methods attenuate exactly those components.","tokens_in":11268,"feed_emoji":"👁️","tokens_out":9497,"duration_ms":90053,"temperature":0.7,"pith_summary":"This paper reports a set of ready-to-run scripts that lessen the influence of ocular (eye-movement and blink) artifacts in EEG recordings without requiring trials to be discarded. The implementations translate three previously proposed ICA-based methods into reusable functions: removing independent components that correlate strongly with vertical and horizontal EOG channels, removing those components only at artifact-marked time samples, and deriving the ICA unmixing matrix from artifact-free segments before partial removal. The authors argue these tools matter because long-trial EEG experiments cannot always avoid eye artifacts by rejecting contaminated trials, and because open, reproducible code lowers the barrier to using and extending ICA-based artifact handling.","feed_headline":"Ready-to-run code removes eye artifacts from EEG recordings","feed_subtitle":"Scripts damp the EEG components that track eye-movement channels, so long recordings can be kept instead of discarded.","key_machinery":"The load-bearing object is the artifact membership function (MSF), a logical vector of the same length as the data in which artifact samples are set to 1 and all others to 0. The paper provides a function that builds the MSF automatically using thresholded z-scores on the EOG channels after a 2-15 Hz band-pass filter, with a cut-off of 0.5 and 0.1 seconds of padding; it also uses the sum of absolute correlation coefficients between ICA components and the vertical and horizontal EOG channels to select which components to remove. The MSF makes the pipeline modular, because any artifact detector that outputs such a marker vector can be substituted.","core_discovery":"The paper's central claim is that its provided functions, built around an artifact membership function and EOG-correlation measures, successfully reduce ocular artifact influence in EEG channels. In the complete-removal variant, ICA components whose summed absolute correlation with the vertical and horizontal EOG signals is an outlier are discarded and the channels are reconstructed from the rest. In the two partial-removal variants, the same component selection is applied but only samples marked by the artifact membership function are zeroed, so non-artifact brain activity outside those windows is meant to be preserved; the third variant recomputes the ICA unmixing matrix on artifact-free samples before the partial removal. The paper shows example results on frontal channels, where large artifact peaks disappear and the summed correlation between independent components and EOG channels drops after processing.","pith_inferences":["An implication the paper leaves implicit is that the default artifact-detection parameters are shown on one subject and one stimulus; testing them across many recordings would show whether they generalize or need per-recording tuning.","A testable extension would replace the sum-of-absolute-correlation selection with component topographies or equivalent dipole locations, which often separate ocular sources more cleanly than raw correlation.","The same membership-function machinery could be pointed at other artifact types with reference channels, such as muscle or cardiac artifacts, though the paper does not demonstrate this.","Allowing the membership function to carry continuous weights instead of zeros and ones would enable graded attenuation, a possibility the authors mention but do not implement."],"forward_implications":["Researchers can apply the functions to their own EEG recordings and keep contaminated trials in the analysis instead of discarding them.","Because the artifact membership function is a simple binary marker vector, alternative artifact detectors can be plugged in without changing the component selection or reconstruction code.","In the two partial-removal methods, only samples identified as artifacts are altered, so genuine brain activity in clean segments should survive the cleaning.","The complete-removal method offers a baseline that removes whole components and may discard some non-artifact signal, which the partial methods are designed to avoid."],"supporting_citations":[{"why":"It supplies the three ICA/EOG-correlation methods that this paper implements.","marker":"[14]"},{"why":"It provides the preprocessing and component-analysis functions on which the scripts are built.","marker":"[16]"},{"why":"It supplies the artifact-detection routine used to build the automatic artifact membership function.","marker":"[17]"},{"why":"It documents the lack of open-source artifact-removal code and data that motivates releasing this implementation.","marker":"[13]"},{"why":"It introduces the detection idea behind the artifact membership function format.","marker":"[19]"}],"fun_headline_variants":["Ready-to-run EEG scripts strip eye artifacts","Drop EOG-linked components to clean EEG data","Keep long EEG trials: code nulls ocular noise","ICA tool removes eye artifacts from EEG channels","Code reduces EEG artifact influence from eye blinks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The automatic artifact detector that builds the membership function is assumed to mark exactly the eye-artifact samples, so that zeroing them removes ocular noise without also deleting genuine brain activity.","fun_headline_variants_meta":{"raw":{"variants":["Ready-to-run EEG scripts strip eye artifacts","Drop EOG-linked components to clean EEG data","Keep long EEG trials: code nulls ocular noise","ICA tool removes eye artifacts from EEG channels","Code reduces EEG artifact influence from eye blinks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001196,"raw_usage":{"total_tokens":4868,"prompt_tokens":819,"completion_tokens":4049,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":435,"completion_tokens_details":{"reasoning_tokens":3979}},"tokens_in":435,"tokens_out":4049,"duration_ms":28686,"temperature":1.0,"reasoning_tokens":3979,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:43:22.509606+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an EEG recording with simultaneous video-based eye tracking or expert-annotated blink and saccade intervals, build the membership function with the provided defaults, and compare its marked samples against the annotations; if many real artifacts go unmarked or many non-artifact segments are marked, the partial-removal methods will either leave artifacts in or attenuate brain responses. A quantitative version would add simulated ocular components with known timing into clean EEG and check whether the partial-removal methods attenuate exactly those components.","supporting_citations":[{"cited_title":"Preprocessing for Lessening the Influence of Eye Artifacts in EEG Analysis","cited_arxiv_id":null,"evidence_quote":"It supplies the three ICA/EOG-correlation methods that this paper implements."},{"cited_title":"FieldTrip: Open Source Software for Advanced Analysis of MEG, EEG, and Invasive Electrophysiological Data","cited_arxiv_id":null,"evidence_quote":"It provides the preprocessing and component-analysis functions on which the scripts are built."},{"cited_title":"Fieldtrip Matlab Software Toolbox for MEG, EEG and iEEG Analysis; Donders Institute for Brain, Cognition and Behaviour: Nijmegen, The Netherlands, 2024","cited_arxiv_id":null,"evidence_quote":"It supplies the artifact-detection routine used to build the automatic artifact membership function."},{"cited_title":"Review of Challenges Associated with the EEG Artifact Removal Methods","cited_arxiv_id":null,"evidence_quote":"It documents the lack of open-source artifact-removal code and data that motivates releasing this implementation."},{"cited_title":"An Improved Algorithm for Automatic Detection of Saccades in Eye Movement Data and for Calculating Saccade Parameters","cited_arxiv_id":null,"evidence_quote":"It introduces the detection idea behind the artifact membership function format."}],"review_version":1}