{"id":"ccd533f7-f20a-4ae4-89c9-6539e4c83c84","arxiv_id":"2606.26506","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"TinyCNNDeep achieves 83.69% subject-wise accuracy on four-class EEG classification of eye states and sleep conditions by converting multi-channel signals to images and applying residual learning with channel attention.","lead":"The paper presents TinyCNNDeep, a lightweight CNN using residual connections and squeeze-and-excitation attention that converts 5-channel EEG segments into 224x224 images and reports 83.69% mean accuracy on four-class classification of eye state and sleep deprivation. A smart generalist might read it for insight into applying image-based deep learning to minimal-electrode brain-signal monitoring in fatigue or vigilance tasks.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Image conversion of 5-channel EEG segments to 224x224 padded grayscale images may discard temporal resolution and introduce artifacts that prevent genuine inter-channel modeling.","rationale":"The reader's weakest assumption directly identifies the same hinge point in the argument. Full-text availability does not remove the need for an explicit check on whether the padding-heavy image transform is responsible for the performance jump; the proposed test isolates that factor without requiring new data.","tokens_in":1813,"tokens_out":368,"duration_ms":35648,"concrete_test":"Re-train TinyCNNDeep on the identical 35-subject segments but replace the 224x224 image pipeline with a 1D residual+SE network operating directly on the raw 5-channel time series (same window length, same subject-wise evaluation); if mean accuracy falls below 70% or the gap to the RF baseline shrinks below 15 points, the image representation is the load-bearing factor and requires separate validation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that per-channel Z-score, min-max scaling, and center padding to 224x224 lets 2D convolutions jointly capture temporal dynamics and cross-channel relations without critical loss. With only five channels (Fp1,Fp2,O1,Oz,O2), the resulting image is dominated by padding rather than signal; this risks the network learning padding boundaries or normalized amplitude patterns instead of physiologically meaningful structure. The reported 36-point gap over time-frequency RF and near-chance performance of EEGNet/ShallowConvNet/DeepConvNet is consistent with either an unusually effective representation or an undetected flaw in how the image transform interacts with the residual+SE architecture. No ablation isolating the representation step is described, leaving this assumption least secure for the headline accuracy number.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces TinyCNNDeep, a lightweight CNN that integrates residual blocks with a Squeeze-and-Excitation attention module for four-class EEG classification (normal vs. sleep-deprived crossed with eyes-open vs. eyes-closed). Short segments from five channels (Fp1, Fp2, O1, Oz, O2) are converted to 224x224 grayscale images via per-channel Z-score normalization, min-max scaling, and center padding; the model is evaluated on a 35-subject dataset and reports a subject-wise mean accuracy of 83.69%, exceeding the strongest baseline (Random Forest on time-frequency features) by 36.03 points while standard EEG architectures perform near chance.","tokens_in":2011,"tokens_out":503,"duration_ms":44817,"significance":"If the central result holds after validation, the work would demonstrate that an image-based representation paired with residual+SE blocks can deliver high multiclass accuracy on sleep-related EEG using only five electrodes and a computationally light model. The subject-wise analysis and confusion-matrix breakdown of error patterns between eyes-closed conditions are useful additions. The large margin over both the RF baseline and established EEG CNNs would be noteworthy if shown to arise from the proposed representation rather than implementation differences.","major_comments":[{"comment":"Methods (EEG-to-image conversion paragraph): converting five-channel segments to 224x224 grayscale images via per-channel Z-score, min-max scaling, and center padding produces images in which the actual signal occupies only a small fraction of the canvas. This risks the 2D convolutions learning padding boundaries or normalized amplitude patterns instead of physiologically meaningful temporal or cross-channel structure; the absence of any ablation that isolates the representation step leaves the 83.69% accuracy claim dependent on an untested assumption.","section":"Methods (EEG-to-image conversion)"},{"comment":"Results (performance tables and baseline comparisons): the near-chance accuracies reported for EEGNet, ShallowConvNet, and DeepConvNet versus TinyCNNDeep's 83.69% require explicit documentation of hyperparameter search, data partitioning (subject-wise vs. pooled), and cross-validation procedure. Without these details it is impossible to determine whether the 36-point gap reflects a genuine representational advantage or differences in training protocol or data leakage.","section":"Results (performance tables and baseline comparisons)"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on our manuscript. We address each major comment below and outline the revisions we will make to strengthen the work.","responses":[{"response":"We acknowledge the referee's valid concern that the central padding leaves the EEG signal occupying only a limited central region of the 224x224 canvas. The conversion procedure (per-channel Z-score normalization followed by min-max scaling to [0,255] and center padding with zeros) was chosen to produce a fixed-size input compatible with standard 2D CNN architectures while preserving the original temporal resolution and amplitude relationships across the five channels. The padding is constant (zero-valued) and does not vary with the data, reducing the likelihood of the model learning spurious boundary patterns; the residual blocks and SE module are intended to focus on the central signal content. Nevertheless, because no ablation isolating the representation step was performed, we will add such an ablation study in the revised manuscript (comparing center-padded images against zero-padded full-canvas versions, resized non-padded images, and a 1D CNN baseline on the raw segments) to demonstrate that the reported accuracy is driven by the signal content rather than padding artifacts.","revision_made":"yes","referee_comment":"[Methods (EEG-to-image conversion)] Methods (EEG-to-image conversion paragraph): converting five-channel segments to 224x224 grayscale images via per-channel Z-score, min-max scaling, and center padding produces images in which the actual signal occupies only a small fraction of the canvas. This risks the 2D convolutions learning padding boundaries or normalized amplitude patterns instead of physiologically meaningful temporal or cross-channel structure; the absence of any ablation that isolates the representation step leaves the 83.69% accuracy claim dependent on an untested assumption."},{"response":"We agree that the current manuscript lacks sufficient detail on the training protocols for the baseline models. All experiments, including those for EEGNet, ShallowConvNet, DeepConvNet, and the Random Forest baseline, were performed under a strict subject-wise partitioning (no subject appears in both training and test sets) with 5-fold cross-validation across the 35 subjects. Hyperparameters for the deep models were selected via grid search on a held-out validation split within each training fold, using the same search space and early-stopping criteria applied to TinyCNNDeep. In the revised manuscript we will add a dedicated subsection (and supplementary table) that explicitly documents the hyperparameter grids, the exact subject-wise split procedure, the number of folds, and the training settings (optimizer, learning rate schedule, batch size, and regularization) for every model. This documentation will allow readers to verify that the performance gap is not attributable to differences in training protocol or data leakage.","revision_made":"yes","referee_comment":"[Results (performance tables and baseline comparisons)] Results (performance tables and baseline comparisons): the near-chance accuracies reported for EEGNet, ShallowConvNet, and DeepConvNet versus TinyCNNDeep's 83.69% require explicit documentation of hyperparameter search, data partitioning (subject-wise vs. pooled), and cross-validation procedure. Without these details it is impossible to determine whether the 36-point gap reflects a genuine representational advantage or differences in training protocol or data leakage."}],"tokens_in":1544,"tokens_out":691,"duration_ms":26489,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main point is that this paper claims 83.69% subject-wise accuracy on joint eye-state and sleep-deprivation classification from five channels, beating a time-frequency random forest by 36 points while standard EEG nets sit near chance. That gap is the result worth noting.\n\nThey turn short EEG segments into 224x224 grayscale images with per-channel Z-score, min-max scaling, and center padding, then feed them to a residual CNN with an SE attention block. Reporting per-subject means and a confusion matrix that pins most errors to the eyes-closed classes is a step above many EEG classification papers.\n\nThe soft spot is the representation itself. Five channels padded out to 224x224 leaves most of the image as uniform background, so the 2D convolutions have limited actual signal to work with. It is not obvious how temporal dynamics survive that transform without loss or how the network avoids latching onto padding edges instead of physiology. No ablation on the image step appears, and the near-chance results from EEGNet, ShallowConvNet, and DeepConvNet are consistent with either a genuine advance or an input format that simply does not suit those baselines. The stress-test concern holds up on the details given.\n\nThis is for people who need lightweight models for applied EEG work with minimal electrodes, especially around sleep and vigilance. A reader who wants concrete numbers on a four-class setup and per-subject variability can pull useful pieces from it.\n\nIt deserves a serious referee because the task is narrow but underexplored and the empirical claim is specific enough to check. Send it to review, but flag the image conversion and ask for ablations plus clearer cross-validation and significance details.","headline":"TinyCNNDeep gets a big reported accuracy lift on the four-class EEG task but the five-channel image padding step looks like the weakest link.","tokens_in":2508,"tokens_out":418,"would_cite":false,"duration_ms":34588,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"TinyCNNDeep classifies four EEG states of sleep condition and eye state at 83.69% subject-wise accuracy from five channels.","keywords":["EEG classification","sleep deprivation","eye state","convolutional neural network","attention mechanism","lightweight model","sleep-related EEG","four-class classification"],"falsifier":"Accuracy falling below 70% on an independent set of subjects recorded under different conditions or with altered electrode placements would show that the performance does not generalize beyond the original dataset.","tokens_in":2706,"feed_emoji":"🧠","tokens_out":687,"duration_ms":28256,"temperature":0.7,"pith_summary":"The paper introduces TinyCNNDeep to jointly classify normal versus sleep-deprived states and eyes-open versus eyes-closed states from short EEG segments. It converts five-channel recordings into 224x224 grayscale images via per-channel normalization and padding, then applies a residual CNN with squeeze-and-excitation attention. On data from 35 subjects, the model reaches 83.69% mean accuracy, exceeding the best traditional baseline by 36 percentage points while other standard EEG networks perform near chance. The work shows that this image-based approach with attention yields an efficient solution for multiclass sleep-related EEG tasks under minimal electrode setups.","feed_headline":"TinyCNNDeep reaches 83.69% accuracy on four-class EEG sleep task","feed_subtitle":"Lightweight residual SE-attention model classifies sleep deprivation and eye states from five channels, beating strongest baseline by 36 poi","key_machinery":"TinyCNNDeep, a lightweight convolutional neural network that combines residual learning with a Squeeze-and-Excitation (SE) attention module applied to 224x224 grayscale images derived from five EEG channels.","core_discovery":"TinyCNNDeep attains a subject-wise mean accuracy of 83.69% on the four-class EEG classification task involving sleep condition and eye state, outperforming the strongest baseline by 36.03 percentage points while using only five channels and an image-based input representation with residual learning and channel attention.","pith_inferences":["The method's success suggests that standard 2D convolutions can capture both temporal dynamics and inter-channel relationships in EEG when the signal is rendered as an image.","The same conversion and architecture could be applied to other few-channel EEG problems such as fatigue detection or cognitive workload classification.","If the accuracy holds across sessions, the approach may support portable monitoring devices that require only five electrodes."],"forward_implications":["The image-based representation paired with residual feature extraction and channel attention supplies an accurate and computationally efficient framework for multiclass sleep-related EEG classification under a minimal electrode configuration.","Per-subject analysis reveals substantial inter-subject variability in classification performance.","Confusion-matrix inspection shows that residual misclassifications concentrate between eyes-closed states across the two sleep conditions.","Three established EEG architectures operate near chance level on this four-class task."],"fun_headline_variants":["TinyCNNDeep scores 83.69% on four-class EEG sleep-eye task","TinyCNNDeep hits 83.69% on EEG four-class sleep task","TinyCNNDeep achieves 83.69% on four-class EEG sleep task","83.69% accuracy by TinyCNNDeep in four-class EEG eye states"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Converting short multi-channel EEG segments into 224x224 grayscale images through per-channel Z-score normalization, min-max scaling, and center padding allows 2D convolutions to jointly model inter-channel and temporal structure without critical information loss.","fun_headline_variants_meta":{"raw":{"variants":["TinyCNNDeep scores 83.69% on four-class EEG sleep-eye task","TinyCNNDeep hits 83.69% on EEG four-class sleep task","TinyCNNDeep achieves 83.69% on four-class EEG sleep task","83.69% accuracy by TinyCNNDeep in four-class EEG eye states"]},"model":"grok-4.3","cost_usd":0.00915,"raw_usage":{"total_tokens":4120,"prompt_tokens":705,"num_sources_used":0,"completion_tokens":82,"cost_in_usd_ticks":91499500,"prompt_tokens_details":{"text_tokens":705,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3333,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":705,"tokens_out":82,"duration_ms":51518,"temperature":1.0,"reasoning_tokens":3333,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T04:19:35.645814+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Accuracy falling below 70% on an independent set of subjects recorded under different conditions or with altered electrode placements would show that the performance does not generalize beyond the original dataset.","supporting_citations":[],"review_version":1}