{"id":"95f44cfd-c0fa-41dc-89e4-04e9f92bd640","arxiv_id":"2509.07416","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"A saccade-guided baseline reconstruction plus wavelet decomposition reduces EOG drift and improves gaze prediction error, but the real-data evaluation compares against a simulation-only baseline.","lead":"This paper introduces a de-drifting method for electrooculography (EOG) signals that identifies saccades, reconstructs a clean baseline, and subtracts an estimated drift trend. The method reports reduced gaze prediction error on simulation and real data, but the real-data comparison relies on a simulation baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-data superiority claim is based on comparing FGD real-data error to alternative methods' simulation errors, not to the same alternatives on the same real data; the 26.53% reduction is not established.","rationale":"I read the paper in good faith: the FGD pipeline is a reasonable idea, and the simulation experiment does compare methods under injected drift. However, the strongest claim explicitly asserts improved gaze prediction in a free-movement HRC scenario and a 26.53% reduction in real data. The support for that assertion is a cross-condition comparison: Table I's real-data column has no alternative-method entries, so the real-data percentage reduction is computed against simulation baselines, not against alternatives applied to the same real signal. This is the most load-bearing weakness because it directly undermines the central comparative claim, independent of any internal assumptions about drift during calibration. The reader's weakest assumption about the drift-free initial five seconds (Sec. III-B-2) is legitimate, but it is secondary: even if the baseline reconstruction is unbiased, the real-data comparison would still be missing. The parameter-tuning-on-evaluation-data issue compounds the risk by potentially inflating the simulation gain. I therefore agree with the REJECT verdict; no verdict adjustment is needed, though a revised paper with real-data baselines and held-out parameter selection could change the outcome.","tokens_in":11039,"tokens_out":6514,"duration_ms":64724,"concrete_test":"Run Wavelet Decomposition at level 7, 5th-order Polynomial Fitting, and 0.3 Hz High-pass Filtering on the same real 90-saccade dataset that produced the FGD real-data column, using the same simulation-trained regression model and the error metric of Eq. (19). If the FGD real-data mean error of 1.033° is not lower than the best alternative's real-data mean error, the central comparative claim fails. Preferably, perform this on multiple subjects with paired significance testing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparative claim rests on Table I, but that table is not a same-condition comparison. The Polynomial Fitting, High-pass Filtering, and Wavelet Decomposition columns are simulation results from 10 injected-drift scenarios, while the Real Data column contains only the proposed FGD, evaluated on 90 saccades from a single subject. Section IV-C then compares the real-data FGD average (1.033°) against the simulation Wavelet average (1.406°) to claim a 26.53% reduction. Because real-world noise and drift differ from the simulated injections, this does not show that FGD outperforms the alternatives on real data. Additionally, the comparison parameters (polynomial order 5, cutoff 0.3 Hz, wavelet level 7, m=15, N=3) were selected by performance testing on the same evaluation data, with no held-out split. The calibration-drift assumption in Sec. III-B-2 is a further internal risk, but even if the baseline reconstruction is sound, the real-data comparison gap is the decisive weakness in the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Feature-Guided De-drifting (FGD) pipeline for single-channel horizontal EOG. It detects saccades via derivative thresholds, reconstructs a saccade-free baseline by vertically aligning floating inter-saccadic segments, approximates the drift trend with level-7 1D wavelet decomposition, and subtracts the trend from the raw signal. Gaze is then predicted by regression calibrated to PtGaze reference. The authors evaluate FGD in simulation by injecting low-frequency drift into a minimally drifting real EOG sample and on real data from one subject (90 saccades). They report mean gaze-prediction error reductions to 0.896° (36.29%) in simulation and 1.033° (26.53%) on real data compared with wavelet decomposition, and claim superior morphology preservation.","tokens_in":11363,"tokens_out":5636,"duration_ms":60223,"significance":"If the claimed accuracy and morphology preservation were established, FGD would be a practical contribution to long-term EOG gaze tracking in free-movement HRC, with the notable advantage of not requiring a fixed cue-target protocol. The algorithmic architecture is clearly described and the idea of feature-guided baseline reconstruction is reasonable. Credit is due for reporting both simulation and pilot real-data results and for specifying the signal model, electrode configuration, and parameter choices. However, the evidence currently provided does not support the central comparative claims: the real-data comparison is not against the same methods on the same data, and key parameters were tuned on the evaluation set. No code or data are supplied, so the numerical results are not independently reproducible. The paper would need re-evaluation experiments, not just re-analysis, to support its conclusions.","major_comments":[{"comment":"The central claim that FGD achieves a 26.53% reduction in real data compared with the best alternative is not supported by a same-condition comparison. In Table I, the Polynomial Fitting, High-pass Filtering, and Wavelet Decomposition columns are simulation results from the 10 injected-drift scenarios, while the Real Data column contains only the proposed FGD. The text in §IV-C nonetheless compares the real-data FGD average (1.033°) with the simulation Wavelet average (1.406°) and interprets the difference as a real-data improvement. Because the real and simulated signals contain different drift and noise characteristics, this comparison cannot establish superiority on real data. The authors should run all comparison methods on the same real recordings and report the resulting per-target and average errors, with a statistical test across saccades or subjects.","section":"§IV-C, Table I"},{"comment":"Several key parameters are selected using the same data on which the final accuracy is reported. The text states that polynomial order 5, high-pass cutoff 0.3 Hz, wavelet decomposition level 7, and baseline reconstruction sample count m=15 were chosen 'based on performance testing with our dataset'; N=3 for the derivative is also tuned to the sampling frequency. No held-out split, cross-validation, or nested validation is described. The reported reductions are therefore optimistically biased and may not generalize. The authors should either fix parameters a priori from prior work/physical reasoning or perform selection on a training set and evaluate on a disjoint test set.","section":"§IV-A parameter selection"},{"comment":"The baseline reconstructor assumes that the first five seconds of calibration are drift-free. Equation (10) computes δ1 from the original E(t) alone, and the text says this 'assum[es] that the initial five seconds of calibration represent a drift-free true baseline.' If drift is already present during calibration, δ1 is biased, and because every later segment in Eqs. (11)-(13) is adjusted relative to the preceding adjusted values, the entire reconstructed baseline and the drift estimate are corrupted. The paper provides no evidence or test that the calibration interval is drift-free in the real data. This assumption should be either justified quantitatively or removed/replaced.","section":"§III-B-2, Eq. (10)"}],"minor_comments":[{"comment":"References [12] and [18] are duplicated (same Bulling et al. paper) and should be consolidated. The acknowledgment contains 'King Abudlaziz City for Science and Technology'; the spelling should be corrected.","section":"General"},{"comment":"No quantitative detection accuracy is reported for the peak detector and saccade window detector; only selected plots are shown. Since the method's name and contribution rest on feature extraction, reporting true/false positive rates or comparing detected saccade boundaries against manual annotations would strengthen the paper.","section":"IV-B"},{"comment":"The table reports only mean errors with standard deviations. The number of saccade events per target is not stated, and no significance tests are performed. At minimum, per-target sample sizes and confidence intervals should be given.","section":"Table I"},{"comment":"The blink removal procedure is described in prose but not formalized; the threshold and 'two consecutive oppositely directed surges' criterion would be difficult to reproduce. Consider adding an equation or pseudocode.","section":"III-A"}],"recommendation":"major_revision","confidential_remarks":"The reader's report recommends rejection, and I agree that the real-data comparison flaw is severe. I am recommending major revision rather than reject because the deficiencies are experimental and fixable in principle: the authors can run the baseline methods on the same real EOG signals and perform parameter selection on a held-out set. If the authors cannot provide such comparisons, or if the existing recorded data are insufficient, rejection should be reconsidered at the revision stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The FGD method is a sensible engineering combination: detect saccades from the derivative, reconstruct a saccade-free baseline using adaptive windows, then let wavelet decomposition approximate the drift. The claimed advantage over TEMOD—no fixed trial protocol, no cue-target knowledge—is real and worth pursuing. The paper is clearly written, the figures actually show the pipeline working, and the authors are upfront about the limited dataset in their future-work paragraph.\n\nThe problem is the evaluation. Table I is not a same-condition comparison. The polynomial, high-pass, and wavelet columns are simulation results from injected drift; the real-data column contains only FGD. The 26.53% reduction is then computed by comparing real-data FGD (1.033°) against the simulation wavelet result (1.406°). That does not show FGD beats those methods on real data—it could just mean the simulated drift is harder than the real drift. The paper's own sentence in Section IV-C says FGD \"outperforms other approaches evaluated in the simulation,\" which is not the same as outperforming them on real data.\n\nAlso, the key parameters (wavelet level j=7, polynomial order 5, high-pass cutoff 0.3 Hz, m=15) were chosen by performance testing on the evaluation dataset. No held-out split, no pre-registration. One subject, 90 saccades, no significance tests. The baseline reconstructor assumes the initial five seconds are drift-free; if that fails, the whole drift estimate is biased. These are load-bearing flaws for the headline numbers.\n\nThat said, the core idea is not broken. The adaptive saccade window and segment-wise baseline adjustment are reasonable, and the method could work with a proper same-condition comparison: run all methods on the same real data, choose parameters on a training set, report variance, and include more subjects. The paper is short and reads like a workshop submission with a promising pipeline but inadequate validation.\n\nIt deserves a serious referee because the idea is useful and the evaluation is fixable. I would not cite the quantitative results as they stand, but I'd keep the paper in mind if they redo the comparison. For a reading group, it's a decent example of how a plausible method can be undermined by an unfair baseline comparison.","headline":"The method is plausible but the central quantitative claim compares real-data FGD to simulation-only alternatives, so the reported superiority is not established.","tokens_in":11804,"tokens_out":1441,"would_cite":false,"duration_ms":18262,"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":"Eye-movement features guide a drift-removal method that cuts EOG gaze error by up to 36 percent.","keywords":["electrooculography","baseline drift","saccade detection","wavelet decomposition","gaze estimation","human-robot collaboration","signal de-drifting"],"falsifier":"Take a simulated EOG signal generated as in the paper's simulation setup but add a low-frequency drift component (e.g., a 0.05 Hz ramp) that begins during the first five seconds of calibration. Run FGD and plain wavelet decomposition, and compare their reconstructed drift trends against the known injected drift. If FGD's trend estimate is no closer to the true drift than the wavelet baseline, the initial drift-free calibration assumption fails.","tokens_in":10965,"feed_emoji":"👁️","tokens_out":4504,"duration_ms":49213,"temperature":0.7,"pith_summary":"This paper claims that EOG gaze signals contaminated by low-frequency baseline drift can be de-drifted accurately by exploiting the timing of saccades rather than by treating the whole signal as drift. The proposed Feature-Guided De-drifting method detects saccades, removes them, stitches the remaining segments into a continuous baseline, and estimates the drift trend from that reconstructed baseline with multilevel wavelet decomposition. In simulation, mean gaze-prediction error drops to 0.896 degrees, a 36.29% reduction over the best conventional alternative; on real data collected from five trials it drops to 1.033 degrees, a 26.53% reduction. The method matters because it removes drift without the fixed trial protocol that previous feature-based de-drifting required, preserving the signal's morphology for free-movement human-robot collaboration.","feed_headline":"Eye-movement features cut EOG gaze error by 36 percent","feed_subtitle":"A wavelet-based drift remover that first removes saccades keeps gaze prediction accurate in free-movement trials.","key_machinery":"The load-bearing mechanism is the Baseline Reconstructor: it converts the saccade-excluded signal into a continuous, drift-only baseline by adding a per-segment displacement computed from the signal mean around each saccade's start and end. This removes saccade morphology from the drift estimate before the 1D multilevel wavelet decomposition (level 7) fits the low-frequency trend, so the final subtraction corrects drift without flattening the eye-movement features.","core_discovery":"The central discovery is that a drift estimate should be computed from a signal from which saccades have been actively subtracted and whose remaining segments have been reconnected, not from the raw EOG trace. Treating E(t)=f(t)+d(t)+w(t), the authors identify saccades from derivative peaks, mark each saccade's start and end with adaptive thresholds, reconstruct the baseline by shifting each non-saccade segment by a displacement that matches its endpoints, and then apply 1D multilevel wavelet decomposition to this reconstructed baseline to approximate the drift trend. Subtracting that trend from the raw signal yields a de-drifted signal whose gaze predictions are more accurate than polynomia","pith_inferences":["If the calibration-period-drift assumption is common in the target deployments, adding a drift estimate for the first five seconds (e.g., from a longer pre-recording or from electrode impedance checks) would make the method more reliable than the paper's current evaluation shows.","The same detect-and-stitch baseline reconstruction could transfer to other biosignals with transient high-frequency events superimposed on low-frequency drift (e.g., ECG baseline wander), provided an event detector with comparable reliability exists.","The reported error reductions combine drift removal with a fixed regression model; a reader should treat the percentages as end-to-end gaze-error gains, not as isolated measures of drift-estimation accuracy.","A direct ablation that compares FGD against wavelet decomposition on the same reconstructed baseline without the saccade-removal step would quantify how much of the gain comes from feature guidance versus the wavelet settings."],"forward_implications":["EOG-based gaze tracking can be used in free-movement scenarios where gaze targets are not pre-scripted, since FGD needs no prior trial-protocol knowledge.","Saccade morphology survives de-drifting, so downstream analyses that rely on saccade amplitudes and timing are not corrupted by the drift removal step.","The method's lower sampling-frequency requirement makes drift-free EOG feasible on cheaper hardware for wearable and robotic interfaces.","The same FGD output can feed sensor-fusion pipelines for human-robot collaboration with a cleaner baseline, improving gaze prediction for intent inference.","With further work, the offline wavelet step can be replaced by low-latency trend estimation, pointing to real-time closed-loop control."],"supporting_citations":[{"why":"Previous baseline-recognition de-drifting that requires a fixed trial protocol; FGD is positioned as removing that requirement.","marker":"[13]"},{"why":"Supplies the eye-movement taxonomy, blink-removal window, and EOG background used in signal acquisition.","marker":"[12]"},{"why":"Documents sources and behavior of EOG baseline drift and prior mitigation techniques, motivating the drift model.","marker":"[20]"},{"why":"Provides the camera-based gaze tracker whose predictions serve as the reference gaze angle for error evaluation.","marker":"[17]"},{"why":"Establishes the linear gaze range used to design the target-guide layout and EOG setup.","marker":"[9]"},{"why":"Polynomial fitting approach used as one of the comparison baselines for de-drifting performance.","marker":"[22]"}],"fun_headline_variants":["EOG drift fix uses eye movements to slash gaze error 36%","New method removes EOG drift, boosting gaze accuracy 36%","Eye-movement-based drift removal cuts EOG error by a third","Saccade-guided de-drift sharpens EOG gaze tracking"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The method assumes the first five seconds of recording are free of baseline drift; if drift is already present during that calibration window, the reconstructed baseline is shifted, and the drift trend estimated from it is wrong.","fun_headline_variants_meta":{"raw":{"variants":["EOG drift fix uses eye movements to slash gaze error 36%","New method removes EOG drift, boosting gaze accuracy 36%","Eye-movement-based drift removal cuts EOG error by a third","Saccade-guided de-drift sharpens EOG gaze tracking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1087,"prompt_tokens":719,"completion_tokens":368,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":292}},"tokens_in":463,"tokens_out":368,"duration_ms":4138,"temperature":1.0,"reasoning_tokens":292,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T22:12:18.532160+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a simulated EOG signal generated as in the paper's simulation setup but add a low-frequency drift component (e.g., a 0.05 Hz ramp) that begins during the first five seconds of calibration. Run FGD and plain wavelet decomposition, and compare their reconstructed drift trends against the known injected drift. If FGD's trend estimate is no closer to the true drift than the wavelet baseline, the initial drift-free calibration assumption fails.","supporting_citations":[{"cited_title":"Temod: Target- enabled model-based de-drifting of the eog signal baseline using a battery model of the eye,","cited_arxiv_id":null,"evidence_quote":"Previous baseline-recognition de-drifting that requires a fixed trial protocol; FGD is positioned as removing that requirement."},{"cited_title":"Eye move- ment analysis for activity recognition using electrooculography,","cited_arxiv_id":null,"evidence_quote":"Supplies the eye-movement taxonomy, blink-removal window, and EOG background used in signal acquisition."},{"cited_title":"A comparison of eog baseline drift mitigation techniques,","cited_arxiv_id":null,"evidence_quote":"Documents sources and behavior of EOG baseline drift and prior mitigation techniques, motivating the drift model."},{"cited_title":"Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation,","cited_arxiv_id":null,"evidence_quote":"Provides the camera-based gaze tracker whose predictions serve as the reference gaze angle for error evaluation."},{"cited_title":"Eye-tracking capabilities of low-cost eog system,","cited_arxiv_id":null,"evidence_quote":"Establishes the linear gaze range used to design the target-guide layout and EOG setup."},{"cited_title":"Algorithm for automatic analysis of electro- oculographic data,","cited_arxiv_id":null,"evidence_quote":"Polynomial fitting approach used as one of the comparison baselines for de-drifting performance."}],"review_version":1}