{"id":"4cb84bcd-37f0-4c64-8ca4-5129bc7fc575","arxiv_id":"2501.01960","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"GAF-FusionNet combines time-series ECG with Gramian Angular Field images through cross-channel attention and reports 94.5%, 96.9%, and 99.6% accuracy on three benchmarks.","lead":"GAF-FusionNet turns each heartbeat recording into both a waveform and a square image, then fuses them with an attention mechanism to classify heart conditions. The authors report high accuracies on three public ECG datasets, but the results are hard to trust because the main data split and the code are not provided.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 99.6% MIT-BIH result rests on an unstated beat-level split; with random 80/20 beats the same patient appears in train and test, so the claim of consistent SOTA is unsupported until patient-independent splitting is confirmed.","rationale":"The reader's weakest_assumption is the same as mine: MIT-BIH patient identity contamination. I agree with that identification. My independent read of the full text confirms no sentence in Section 4.1 or elsewhere describes a patient-disjoint split; Table 1 only gives sample counts. Since 109,446 beats originate from 48 records (approximately 47 subjects), a split made by randomly assigning beats is not patient-disjoint. The central claim of \"significant improvements over SOTA\" in the abstract and conclusion relies on the MIT-BIH result, so until the split is clarified the empirical claim is unsupported. I would keep the reader's REJECT; the proposed check is a direct way to either substantiate the concern or retire it. I did not find a separate, equally load-bearing internal inconsistency; the architecture details are sparse, but the empirical protocol is the decisive issue.","tokens_in":8887,"tokens_out":4558,"duration_ms":42100,"concrete_test":"Retrain GAF-FusionNet on MIT-BIH using an inter-patient split: assign all beats from each record to either train or test (merge records 201 and 202, which share a subject, so no subject appears in both), e.g., an 80/20 record-level split or leave-one-record-out, and compare against Table 2. If the 99.6% accuracy drops by more than about 2 percentage points, or no longer clearly exceeds the Multi-Scale CNN baseline, the reported result is explained by beat-level leakage and the central claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that GAF-FusionNet \"consistently outperforms all baseline methods\" in Table 2. The largest and most important number is 99.6% on MIT-BIH. The paper never states the train/test protocol for MIT-BIH: Table 1 gives 87,554/21,892 beats (an 80/20 split by beat count), but the Methods and Experiments sections do not say that the 48 recordings were assigned to train or test as units, or that the same patient's beats never appear on both sides. The dataset contains 48 half-hour excerpts from about 47 subjects, so a random 80/20 beat-level split makes it essentially certain that every subject contributes beats to both train and test. Under those conditions, the model can memorize per-patient waveform patterns and the 99.6% figure reflects identity leakage rather than the claimed multimodal generalization. This is a missing experimental condition, not a disagreement with community consensus; the standard inter-patient protocol would settle it. The ECG200 and ECG5000 results use UCR fixed splits and are less suspect, but the Table 2 \"consistent outperformance\" claim inherits the MIT-BIH problem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GAF-FusionNet, a multimodal ECG classifier that processes raw ECG time series and Gramian Angular Field (GAF) images in two parallel branches, fusing them with a dual-layer cross-channel split attention module. The method is evaluated on ECG200, ECG5000, and the MIT-BIH Arrhythmia Database, and Table 2 reports accuracy of 94.5%, 96.9%, and 99.6% respectively, with the claim that GAF-FusionNet consistently outperforms all compared baselines. The paper also presents an ablation study on MIT-BIH. The central claim is empirical: the proposed fusion architecture sets a new state of the art on all three datasets.","tokens_in":9134,"tokens_out":4594,"duration_ms":43441,"significance":"If the reported results hold, GAF-FusionNet would be a useful contribution to multimodal ECG classification, showing that combining time-series and image-based representations with learned cross-modal attention can improve accuracy over single-modality models. The GAF formulation in Section 3.2 and the attention fusion equations in Section 3.3 are clearly presented, and the ablation study is a reasonable attempt to isolate the contribution of each component. However, the significance is currently undercut by missing experimental protocol details. The paper contains no theoretical derivations, no machine-checked proofs, and no released code or data split at the time of review, so the empirical claims must carry the entire contribution. The main result on MIT-BIH depends on a train/test split whose patient independence is not stated, and one of the standard dataset splits appears to be reversed. These issues are fixable but are load-bearing for the paper's central claim.","major_comments":[{"comment":"The MIT-BIH train/test protocol is not specified. Table 1 reports an 87,554/21,892 beat-level split, but the text never states that recordings from the same patient are confined to one side of the split, nor does it mention the standard inter-patient protocol for this database. With 48 recordings from roughly 47 subjects, a random beat-level 80/20 split makes it nearly certain that beats from every subject appear in both training and test sets, so the 99.6% accuracy in Table 2 may reflect memorization of patient-identity-specific waveform patterns rather than generalization. Please report the exact patient-to-split assignment and rerun all MIT-BIH results under an inter-patient protocol (e.g., training on a subset of patients and testing on held-out patients), stating the number of patients in each split.","section":"§4.1, Table 1; §4.2, Table 2"},{"comment":"The ECG5000 split appears to be reversed. The UCR ECG5000 archive uses 500 training samples and 4,500 test samples, but Table 1 lists a 4,500/500 train/test split. If the authors trained on 4,500 samples and tested on 500, the reported 96.9% accuracy is not comparable with results obtained under the standard benchmark split. Please confirm the actual split and correct the table, or explain the discrepancy.","section":"§4.1, Table 1"},{"comment":"The implementation description is internally inconsistent. Section 3.3 defines a temporal branch using a 1D CNN followed by BiLSTM and a spatial branch using a 2D CNN, but the implementation paragraph says 'We use Resnet34, pre-trained by ImageNet, as backnone of the feature extraction layer' without specifying which branch this replaces or how a 1D ECG signal is adapted for a 2D ImageNet-pretrained network. Additionally, Eq. (21) defines a square-root decay learning-rate schedule while the text states a cosine annealing schedule. These details are essential for reproducibility and must be resolved.","section":"§4.1, Implementation Details; §3.3"},{"comment":"All reported results are single runs without error bars, confidence intervals, or statistical significance tests. The ECG200 test set contains only 100 samples, so the claimed 2.0-point accuracy improvement over the best baseline corresponds to two additional correct predictions. To support the claim that GAF-FusionNet 'consistently outperforms all baseline methods,' the authors should report results over multiple random seeds (or a paired test over the same test folds) and provide variance or significance information.","section":"§4.2, Table 2"}],"minor_comments":[{"comment":"The code link is given as 'will soon be available,' but no code is currently provided; please make the code and data split publicly available at the time of publication so the experiments can be reproduced.","section":"Abstract"},{"comment":"Figure 1 is too coarse to verify the details of the dual-layer attention module; please annotate the tensor shapes and show explicitly where 'Split 1' through 'Split r' and 'Global pooling' correspond to the operations in Eqs. (12)–(17).","section":"Figure 1"},{"comment":"The concluding sentence mentions 'aiding in the understanding and treatment of psychiatric disorders'; this appears unrelated to the ECG classification task and should be corrected (probably to cardiovascular conditions).","section":"§5, Conclusion"},{"comment":"Several in-text citation names do not match the reference list entries (e.g., 'Wei et al.' vs. Guo et al. [7], 'Satria et al.' vs. Mandala et al. [15], 'Michal et al.' vs. Heldeweg et al. [9], 'Madeline et al.' vs. Kent et al. [11]). Please align all citations and reference entries.","section":"References"},{"comment":"The segmentation formula uses an overlap parameter o, but it is not stated how the window length w and overlap o are chosen for each dataset, nor whether the final window is padded when the signal ends; please specify these choices.","section":"§3.1, Eq. (3)"},{"comment":"The text says training was performed on 'an NVIDIA RTX 4090 GPU with 128GB memory'; the RTX 4090 has 24GB of memory, so this hardware description is inaccurate and should be corrected.","section":"§4.1, Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The central claim is empirical and currently rests on an unspecified and potentially contaminated MIT-BIH split, plus a likely reversed ECG5000 split. I would not recommend acceptance until the authors rerun the experiments under a patient-independent MIT-BIH protocol and under the standard UCR splits, and until the implementation details (ResNet34 vs. the described 1D/2D CNN branches, learning-rate schedule) are reconciled. If the authors cannot provide the data split or code, the reported numbers should not be considered reproducible. The architectural idea is reasonable, but the paper's contribution is mostly empirical, so these experimental-protocol issues are decisive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the 99.6% MIT-BIH result is the whole ballgame, and the paper never tells you how the split was made. With 48 recordings and an 80/20 beat-level split, the same patient almost certainly shows up on both sides, so the headline is likely identity leakage. The stress-test note lands.\n\nWhat's actually new: the dual-layer cross-channel split attention is a real although modest combination—intra-modality self-attention followed by cross-attention, with residual connections and layer norm. The ablation table is honest and shows each component earns its keep. The UCR fixed splits for ECG200 and ECG5000 are standard, so those numbers are less suspect; if the MIT-BIH split were patient-independent, the architecture could be a useful benchmark.\n\nSoft spots beyond the split: no code is actually available despite the abstract promising it; no error bars or repeated runs; implementation details conflict (ResNet34 pretrained on ImageNet as backbone vs. a small 2D CNN in the figure, for GAF images whose size makes ImageNet pretraining odd); the learning-rate schedule in Eq. (21) is not the cosine annealing described in Section 4.1; the baselines appear to be reimplemented without citations to their reported numbers. The claim that GAF-to-ECG is 'largely unexplored' is inaccurate—there's prior work, some cited in their own paper. Minor: the conclusion drifts into 'psychiatric disorders,' which suggests careless editing.\n\nWho it's for: a cautionary example of beat-level leakage, and maybe a starting point if the authors fix the split. As is, the central claim is unsupported. If I were the editor, I'd desk reject and invite a resubmission with a proper inter-patient split, error bars, and code.","headline":"The MIT-BIH result is uninterpretable without the split; a solid incremental architecture is wasted on a likely-leaky benchmark.","tokens_in":9646,"tokens_out":3238,"would_cite":false,"duration_ms":29995,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GAF-FusionNet fuses ECG waveforms with Gramian Angular Field images and split attention, and reports accuracies of 94.5%, 96.9%, and 99.6% on ECG200, ECG5000, and MIT-BIH Arrhythmia.","keywords":["ECG classification","Gramian Angular Field","Split attention","Multimodal fusion","Arrhythmia detection","Time-series imaging","Deep learning"],"falsifier":"Re-run GAF-FusionNet on MIT-BIH with a truly patient-disjoint split—train on one set of subjects, test on a different set—and compare accuracy to the reported 99.6% beat-level figure; a large drop would show that same-patient beats, not learned cardiac patterns, carried the result.","tokens_in":8695,"feed_emoji":"❤️","tokens_out":11023,"duration_ms":88899,"temperature":0.7,"pith_summary":"GAF-FusionNet is an attempt to show that ECG classification improves when the raw heartbeat waveform and an image encoding of the same signal are learned jointly rather than treated as one modality. The paper maps each ECG segment to a Gramian Angular Field image, runs a 1D CNN plus BiLSTM on the waveform and a 2D CNN on the image, and merges the two branches with a dual-layer cross-channel split attention module. On ECG200, ECG5000, and the MIT-BIH Arrhythmia Database, the reported accuracies are 94.5%, 96.9%, and 99.6%, respectively, beating the listed baselines on every dataset and metric. The intended takeaway is that adaptive cross-modal fusion of waveform and image features is a productive direction for ECG classification.","feed_headline":"Waveform plus image encoding hits 99.6% on MIT-BIH arrhythmia","feed_subtitle":"The fusion model also beats five baselines on ECG200 and ECG5000.","key_machinery":"The central mechanism is the dual-layer cross-channel split attention module, a two-stage attention block between the temporal and spatial branches of the network. The temporal branch is a 1D CNN feeding a BiLSTM; the spatial branch is a 2D CNN applied to the Gramian Angular Field matrix, whose entries are $\\cos(\\phi_j + \\phi_k)$ for angular encodings $\\phi_j$ and $\\phi_k$ of rescaled ECG samples. In the first attention layer, each branch self-attends over its own features; in the second, each branch attends to the other branch's projected features, and the two attended representations are added, normalized, concatenated, and passed to an MLP classifier. The module's role is to let the network decide per sample how much weight to give the waveform versus the image view, which the ablations identify as the main source of accuracy gain.","core_discovery":"On the paper's own terms, the central discovery is that a dual representation of ECG—raw time series plus a Gramian Angular Field image—combined through learned cross-modal attention beats every baseline it is compared against. The reported margins are 2.0 percentage points over the best baseline on ECG200, 1.2 points on ECG5000, and 0.8 points on the MIT-BIH Arrhythmia Database, where the model reaches 99.6% accuracy with a macro F1 of 99.5%. The ablation study attributes the gain to the dual-layer split attention module: removing it drops MIT-BIH accuracy to 97.8%, removing cross-channel interaction drops it to 98.1%, and either single-modality branch stays below 97.5%. The intended lesson is that learned, context-dependent weighting of waveform and image features matters more than simply concatenating them.","pith_inferences":["Editorial inference: the attention module is modality-agnostic, so the same GAF-plus-split-attention design should transfer to EEG, electromyography, or other one-dimensional biosignals without architectural change.","Editorial inference: because the cross-channel attention computes pairwise branch interactions, the design extends naturally to three or more input views—for instance, multiple ECG leads or an added spectrogram—by chaining additional cross-attention steps.","Editorial inference: the paper reports only aggregate metrics; a per-class confusion matrix on MIT-BIH, especially for rare arrhythmia classes, would clarify whether the 99.5% macro F1 is earned evenly or dominated by common beat types."],"forward_implications":["Adding the GAF image branch and split-attention fusion should improve over waveform-only ECG models: the ablation shows the full model is 2.6 points above the time-series-only variant on MIT-BIH (99.6% vs 97.0%).","If the MIT-BIH result is patient-disjoint, 99.6% on 15 heartbeat classes would be a new benchmark among the listed baselines, whose best is Multi-Scale CNN at 97.8%.","The dual-layer attention module is the main source of gain: removing it costs 1.8 points, while removing only cross-channel attention costs 1.5 points.","The fusion recipe transfers across dataset sizes and recording lengths, from 200-sample ECG200 to the 109,446-beat MIT-BIH collection."],"supporting_citations":[{"why":"Introduces Gramian Angular Fields, the time-series-to-image transform that creates the spatial branch input.","marker":"[23]"},{"why":"Supplies the MIT-BIH Arrhythmia Database, the source of the 15-class beat data and the 99.6% result.","marker":"[16]"},{"why":"Supplies the ECG200 and ECG5000 time-series datasets used for the other two benchmark results.","marker":"[3]"},{"why":"Defines the DNN baseline that GAF-FusionNet must outperform on all three datasets.","marker":"[8]"},{"why":"Defines the attention-based CNN baseline and motivates the use of attention within ECG classification.","marker":"[5]"},{"why":"Defines the Multi-Scale CNN baseline, the strongest listed competitor on MIT-BIH.","marker":"[21]"},{"why":"Defines the LSTM-FCN baseline combining recurrent and convolutional feature extraction.","marker":"[10]"},{"why":"Defines the Informer transformer-style baseline in the comparison table.","marker":"[27]"},{"why":"Provides the self-attention formulation that the dual-layer split attention module adapts.","marker":"[20]"}],"fun_headline_variants":["GAF-FusionNet: 99.6% on MIT-BIH via wave+image fusion","Cross-modal attention fusion hits 99.6% on ECG benchmark","Fusing ECG waveforms with images: 99.6% on MIT-BIH","Dual-modality split attention lifts ECG accuracy to 99.6%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported MIT-BIH accuracy presupposes that the 87,554 training beats and 21,892 test beats come from different patients; if the same patient's beats appear in both sets, the model can memorize individual heartbeats rather than learn generalizable arrhythmia patterns.","fun_headline_variants_meta":{"raw":{"variants":["GAF-FusionNet: 99.6% on MIT-BIH via wave+image fusion","Cross-modal attention fusion hits 99.6% on ECG benchmark","Fusing ECG waveforms with images: 99.6% on MIT-BIH","Dual-modality split attention lifts ECG accuracy to 99.6%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000578,"raw_usage":{"total_tokens":2692,"prompt_tokens":881,"completion_tokens":1811,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":1724}},"tokens_in":497,"tokens_out":1811,"duration_ms":10992,"temperature":1.0,"reasoning_tokens":1724,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:34:44.507642+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run GAF-FusionNet on MIT-BIH with a truly patient-disjoint split—train on one set of subjects, test on a different set—and compare accuracy to the reported 99.6% beat-level figure; a large drop would show that same-patient beats, not learned cardiac patterns, carried the result.","supporting_citations":[{"cited_title":"In: Workshops at the Twenty-Ninth AAAI Conference on Artificial Intelligence (2015),https://api","cited_arxiv_id":null,"evidence_quote":"Introduces Gramian Angular Fields, the time-series-to-image transform that creates the spatial branch input."},{"cited_title":"IEEE engineering in medicine and biology magazine20(3), 45–50 (2001)","cited_arxiv_id":null,"evidence_quote":"Supplies the MIT-BIH Arrhythmia Database, the source of the 15-class beat data and the 99.6% result."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ECG200 and ECG5000 time-series datasets used for the other two benchmark results."},{"cited_title":"Nature Medicine25(1), 65–69 (2019)","cited_arxiv_id":null,"evidence_quote":"Defines the DNN baseline that GAF-FusionNet must outperform on all three datasets."},{"cited_title":"Scientific Reports 7(1), 10543 (2017)","cited_arxiv_id":null,"evidence_quote":"Defines the attention-based CNN baseline and motivates the use of attention within ECG classification."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the LSTM-FCN baseline combining recurrent and convolutional feature extraction."}],"review_version":1}