{"id":"e2ed2a72-baf1-410b-9b39-ec1eb66787c8","arxiv_id":"2411.13903","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"AmpliNetECG12, a 280k-parameter CNN with a softmax-based activation, reports an F1 score of 80.7% and ROC-AUC of 97% on the CPSC2018 12-lead ECG arrhythmia benchmark using 50 Hz data.","lead":"This paper introduces AmpliNetECG12, a compact convolutional network for classifying 12-lead ECG signals into nine arrhythmia types, using a new softmax-based activation called aSoftMax. The authors report competitive benchmark results with only 280,000 parameters, which is relevant for portable cardiac monitoring, but the manuscript contains several internal inconsistencies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The architecture description is internally inconsistent: Table 2's output shapes violate convolution arithmetic, and Section 4 describes a different model with three backbones and an attention module, so the reported F1/AUC cannot be attributed to a well-defined architecture.","rationale":"The reader's weakest_assumption focused on the missing evaluation protocol (split ratio, epochs, seeds, preprocessing), which is indeed a reproducibility barrier. My stress-test identifies a more fundamental attribution problem: the paper does not describe one coherent architecture at all. Table 2 is dimensionally inconsistent at the ConvB 1x9/s1 layer, and Section 4 explicitly describes a different system (three backbones, lead-wise attention). This means the reported F1/AUC cannot be confidently assigned to the architecture the paper claims to introduce. That is not a disagreement with the reader's verdict; it reinforces REJECT. I mark agreement as 'partial' because the reader's stated weakest assumption was the protocol rather than the architectural incoherence, though their rationale did mention 'the architecture table is internally inconsistent' and 'the Discussion section describes a different model.' The concrete test I propose is a minimal first step: instantiate the table and check whether the forward pass is even possible. If it fails, no amount of hyperparameter disclosure can rescue the claim; a corrected architecture and code release would be necessary. The paper does have some credible pieces—a real public dataset (CPSC2018), a plausible lightweight design idea, and standard metrics—but those do not offset the inability to pin down what was actually evaluated. On the central claim, the load-bearing condition is 'the evaluated model equals the described model,' and that condition is currently violated. Therefore the reader's REJECT verdict should stand unchanged.","tokens_in":9994,"tokens_out":2224,"duration_ms":23438,"concrete_test":"Implement the Table 2 architecture in TensorFlow with exactly the listed filter shapes, strides, and 'same' or 'valid' padding choices, and run a forward pass on a 12x1500x1 input. If the ConvB 1x9/s1 layer cannot produce a 12x1500x128 output from a 12x750x128 input (it cannot without unspecified transpose or asymmetric padding), the architecture table is self-contradictory. Then independently count backbones and attention modules in Section 2.3/Table 2 versus Section 4; if Section 4's three-backbone, attention-based design is absent from Table 2, the reported results are not tied to a single architecture. Re-running the model after resolving these contradictions and confirming the F1/AUC values would settle the attribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the specific architecture in Table 2, using aSoftMax and kernel weight sharing, achieves an F1 of 80.74% and AUC of 97.0% on CPSC2018 with 280k parameters. This attribution requires that the evaluated model is exactly the described model. That requirement fails twice. First, Table 2 lists a ConvB 1x9/s1 layer following a ConvB 1x3/s2 layer; the preceding output is 12x750x128, but the table gives the next output as 12x1500x128. With a stride of 1 and no stated padding, this is dimensionally impossible; with padding, the padding and any asymmetry are unspecified, so the architecture is not precisely defined. Second, Section 4 states that the system processes 'three input ECG leads independently using three distinct backbones' and introduces a 'Lead-wise Attention module,' neither of which appears in Table 2 or Section 2.3. A reader cannot tell which architecture produced the numbers; the abstract's AUC (96.00%) also differs from Section 3's micro-average AUC (0.97) and the conclusion's 97.0%, and the F1 varies between 80.71%, 80.74%, and 0.8070 across text and table. These are not merely missing hyperparameters; they are contradictions about the object of evaluation. If the evaluated model is not the described model, the headline performance claim is unattributable and untestable, regardless of whether the evaluation protocol later becomes reproducible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AmpliNetECG12, a lightweight convolutional neural network for 12-lead ECG arrhythmia classification on the CPSC2018 dataset. The two claimed innovations are a new activation function called aSoftMax, defined as A(x) = x ⊙ softmax(|x|), and a kernel weight-sharing mechanism applied across ECG leads. The authors report an accuracy of 84%, an F1-score of about 80.7%, and an ROC-AUC of about 96–97% with 280,000 trainable parameters, using 50 Hz downsampled data. The paper also claims interpretability through SHAP values, UMAP visualizations, and feature-map analysis. The central quantitative claims are not verifiable from the manuscript because the architecture is internally inconsistent, the reported metrics are mutually contradictory, and the experimental protocol is under-specified.","tokens_in":10347,"tokens_out":3487,"duration_ms":34768,"significance":"If the reported performance were reproducible, the result would be practically significant: a model with 280k parameters operating on 50 Hz 12-lead ECG that approaches or exceeds lightweight baselines could enable wearable and low-resource deployment. The paper uses a public benchmark, compares against several lightweight baselines, and attempts an explainability analysis. However, the manuscript in its current form does not provide a well-defined architecture, consistent numbers, or a reproducible protocol, so the significance of the claimed contribution cannot be assessed. The absence of any ablation study also leaves the specific role of aSoftMax unverified.","major_comments":[{"comment":"The output shape of the fourth ConvB layer is dimensionally impossible. The preceding ConvB 1x3/s2 layer produces an output of 12 x 750 x 128, and the next ConvB layer is a 1x9 convolution with stride 1. With 'same' padding, a stride-1 convolution preserves the temporal length, so the output cannot become 12 x 1500 x 128; without padding or with explicit upsampling, the spatial dimensions would shrink. As written, the architecture cannot be instantiated, so the model that generated the reported F1/AUC numbers is not defined by the table.","section":"§2.3, Table 2"},{"comment":"Section 4 describes a system that 'processes three input ECG leads independently using three distinct backbones' and introduces a 'Lead-wise Attention module,' neither of which appears in the architecture described in Section 2.3 or in Table 2. These are two different models, and the paper does not state which one produced the experimental results. The headline performance claim therefore cannot be attributed to the proposed architecture with any confidence.","section":"§4 vs. §2.3, Table 2"},{"comment":"The reported headline metrics are mutually inconsistent. The abstract reports F1 = 80.71% and AUC = 96.00%; Section 3 reports micro-averaged AUC = 0.97, precision 82%, recall 80%, and F1 = 81%; Table 3 gives an average F1 of 0.8074; Table 4 gives F1 = 0.8070; and Section 6 reports F1 = 80.74% and AUC = 97.0%. Likewise, Table 4 reports 0.20B FLOPs and a model size of 1.01 MB, while the text in Section 3.1 states 0.76 billion FLOPs and 1.2 MB. A reader cannot determine the actual claimed performance or computational cost.","section":"§1, §3, §6, Table 4"},{"comment":"The experimental protocol is under-specified to the point of non-reproducibility. No split ratio, number of training epochs, batch size, random seed, or exact preprocessing pipeline is given. The statement that records were 'stratified and split' does not specify the fraction or the stratification variable. Additionally, the paper describes CPSC2018 as multi-label but then says labels were 'one-hot encoded to transform the problem into a multi-class classification task'; if the dataset contains records with more than one label, this transformation is not well defined without further explanation. These omissions make the comparison in Table 4 unverifiable.","section":"§2.7, Table 4"},{"comment":"The central novelty, aSoftMax, is not isolated or validated. The activation is written as A(x) = x ⊙ softmax(|x|), but the text refers to it as 'Eq. X' and the promised plot and first-order derivative are absent. More importantly, no ablation replaces aSoftMax with ReLU, Swish, or standard softmax under the same architecture; without such an experiment, the paper provides no evidence that the activation contributes to the reported results. The claim in the abstract that aSoftMax has 'stochastic characteristics' is also unsupported, since the given formula is deterministic.","section":"§2.5"}],"minor_comments":[{"comment":"The text cites 'lightX3ECG [46]' and 'ECGNet' without a reference number in the list; the reference list ends at [28], so the comparendum sources are incomplete.","section":"§3.1"},{"comment":"The preprocessing description is unclear about how 1500 samples are selected from recordings of 6–60 seconds: the text says 'excess values been dropped,' but the exact cropping window and alignment procedure are not specified.","section":"§2.2"},{"comment":"There are numerous typos and grammatical errors that impede readability, including 'increasse' in the abstract, 'Comparision' in Table 4, 'ConvB for with Conv2D' in Table 2, and 'captures the summary' mismatched with a plural subject.","section":"Throughout"},{"comment":"The kernel weight-sharing description says leads 'are not stacked as channels but treated as separate dimensions,' but the input tensor is written as 12 x 1500 x 1 and the convolution operations appear to treat the 12 as a channel-like dimension; the relationship between the prose and the tensor shapes is not made explicit.","section":"§2.4"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to describe two different architectures and to report inconsistent performance numbers; the central empirical claim cannot be tied to a well-defined model. This is a load-bearing problem that a normal revision could not fix without substantially redoing the experiments and rewriting the architecture description. If the authors can supply the exact architecture, code, and protocol, a fresh submission might be worth considering, but the present manuscript is not publishable in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the core idea is worth a look, but this manuscript is not. The aSoftMax activation—x times softmax of |x|—is simple and interesting for ECG, and cross-lead kernel sharing is a sensible way to cut parameters. If the reported F1 of 0.81 at 50 Hz with 280k parameters were reproducible, it would be a useful option for low-power screening. But it isn't reproducible, and the paper doesn't even describe a single coherent model. Table 2 has a stride-2 block that supposedly turns 12x750x128 into 12x1500x128 through a 1x9 stride-1 convolution; that's dimensionally impossible. Section 4 describes three backbones and a Lead-wise Attention module that appear nowhere in the architecture description. And the headline numbers drift: F1 80.71 vs 80.74 vs 0.8070, AUC 96% vs 0.97, FLOPs 0.20 B vs 0.76 B, size 1.01 vs 1.2 MB. No split ratio, epochs, batch size, or seed is given; no code. So the central claim—that this specific architecture achieves these numbers—can't be attributed to any well-defined model. The per-class table looks plausible, and the weak PAC result (F1 0.60) is honestly reported, but that doesn't fix the attribution problem. The reference list is standard and related work is fairly covered, so I don't see a citation-pattern issue. This paper is for someone who wants a quick idea to try in their own lightweight ECG pipeline, not for anyone wanting a citable result. As it stands I wouldn't send it to referees; the internal contradictions are visible on first read. I'd desk-reject with an invitation to resubmit if the author can provide a consistent architecture, full training protocol, and code. A clean ablation of aSoftmax would be the minimum evidence I'd want to see.","headline":"Interesting aSoftMax idea, but internal contradictions and a missing protocol make the reported results unattributable.","tokens_in":10892,"tokens_out":2862,"would_cite":false,"duration_ms":27918,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Using 280,000 trainable parameters and 50 Hz data, a convolutional network with the new aSoftMax activation and cross-lead kernel sharing reports 80.74% average F1 and 97.0% ROC-AUC on nine-class CPSC2018 arrhythmia classification.","keywords":["12-lead ECG","arrhythmia classification","lightweight CNN","aSoftMax activation","kernel weight sharing","CPSC2018","explainable AI","ECG downsampling"],"falsifier":"Re-run the model on CPSC2018 with a fully documented stratified split and identical 50 Hz preprocessing, and compare per-class F1 with LightX3ECG and the other baselines; if the average F1 is not close to 80.7% or does not beat them, the efficiency claim fails.","tokens_in":9763,"feed_emoji":"🫀","tokens_out":9170,"duration_ms":82498,"temperature":0.7,"pith_summary":"The paper sets out to show that accurate 12-lead ECG arrhythmia classification does not require large models or high sampling rates. It introduces a convolutional architecture, AmpliNetECG12, built around a new activation function called aSoftMax, defined as $A(x) = x \\odot \\operatorname{softmax}(|x|)$, which amplifies the amplitude of ECG deflections in a normalized way. The model shares convolutional kernels across the 12 leads, treating leads as separate spatial dimensions rather than channels, which cuts trainable parameters to 280,000. On the CPSC2018 dataset, with signals downsampled to 50 Hz, the paper reports an average F1 of 80.74%, a ROC-AUC of 97.0%, and an accuracy of about 84%, outperforming several lightweight baselines at a fraction of their compute and storage. If these numbers hold, the architecture makes portable, resource-limited arrhythmia screening more practical.","feed_headline":"A 280K-parameter ECG model hits 80.7% F1 on 9 arrhythmias","feed_subtitle":"New aSoftMax activation plus cross-lead weight sharing make 12-lead arrhythmia screening portable at 50 Hz.","key_machinery":"The machinery has three parts. The first is aSoftMax, an activation defined by $A(x) = x \\odot \\operatorname{softmax}(|x|)$; by applying softmax to absolute values and multiplying back by the original signal, it acts as a relativistic amplifier that boosts prominent ECG deflections while suppressing weak ones, and it is used after every convolutional block. The second is kernel weight sharing across leads: ECG leads are stacked as separate dimensions rather than channels, and the same 1D kernels are convolved over each lead, so the network learns features common to all leads instead of training one extractor per lead. The third is a compact body of four convolutional blocks with layer normalization, lead-wise min and max pooling whose outputs are concatenated, followed by a SoftMax classifier; training uses Adamax and categorical focal loss with label smoothing.","core_discovery":"The central claim is that a deliberately small CNN can match or beat larger lightweight models on nine-class arrhythmia classification when two design choices are combined: aSoftMax activation, which computes $A(x) = x \\odot \\operatorname{softmax}(|x|)$ so that the sign of each sample is preserved while its magnitude is reweighted relative to the whole tensor, and kernel weight sharing across leads, which uses the same filters for every lead and thereby captures features that are consistent across the 12-lead layout. On CPSC2018 the model reports average F1 80.74%, ROC-AUC 97.0%, precision 81.90%, recall 80.33%, with 280,000 trainable parameters and a disk footprint of 1.01 MB. The authors interpret the result as evidence that compressed 50 Hz data retains enough morphology for reliable detection and that the stochastic character of aSoftMax improves interpretability by highlighting the ECG segments that drive each arrhythmia decision.","pith_inferences":["A test the paper itself does not run is swapping aSoftMax for ReLU, Swish, or plain softmax at equal parameter count; if the F1 gap shrinks, the activation is the active ingredient.","Because aSoftMax reweights magnitudes rather than signs, the mechanism could transfer to other multilead physiological signals such as EEG or EMG, where deflection polarity is clinically meaningful.","At 50 Hz, the Nyquist limit discards everything above 25 Hz; comparing the same model at 50, 100, and 250 Hz would show where the diagnostic information lives, which the paper leaves open.","The interpretability claim could be tested by checking whether the highlighted regions align with clinician-annotated P, QRS, and T waves; the paper does not quantify this alignment."],"forward_implications":["A 280,000-parameter model that works on 50 Hz data can run on wearable and portable devices with limited memory, battery, and transmission bandwidth.","Downsampling 12-lead ECGs by a factor of 10 reduces storage and data-transfer costs without, on the reported numbers, sacrificing classification quality.","Shared kernels across leads imply that a single feature extractor can serve all leads, which should improve generalization across lead layouts and reduce overfitting on lead-specific noise.","The aSoftMax activation's amplification of deflections gives a built-in explanation channel: the same weights that classify also indicate which parts of the ECG waveform mattered, supporting clinician-facing interpretability.","The model establishes a new lightweight operating point on CPSC2018, at a fraction of the parameters of the lightweight baselines it is compared with."],"supporting_citations":[{"why":"It supplies the CPSC2018 dataset of 6,877 12-lead recordings with nine diagnostic labels, and all reported accuracy, F1, and AUC numbers are measured on it.","marker":"[22]"},{"why":"It is the Swish activation from which aSoftMax is derived, providing the functional family the new activation modifies.","marker":"[21]"},{"why":"It defines the Adam optimizer, and the Adamax optimizer used for training is a variant of it.","marker":"[23]"},{"why":"It introduces focal loss, which is used as categorical focal cross-entropy to handle class imbalance during training.","marker":"[25]"},{"why":"It provides the 1D-ResNet34 baseline in the comparison and the Shapley-value interpretability approach applied to predictions.","marker":"[20]"},{"why":"It provides the 1D-SEResNet34 baseline in the comparison table.","marker":"[26]"},{"why":"It provides the TI-ResNet18 baseline, an attention-based time-incremental CNN that the paper compares against.","marker":"[27]"},{"why":"It provides the InceptionTime baseline in the comparison table.","marker":"[28]"},{"why":"It is the LightX3ECG lightweight baseline cited in the comparison table, although no matching entry appears in the reference list.","marker":"[46]"}],"fun_headline_variants":["280K-param CNN with aSoftMax matches big ECG models on 9 classes","1 MB ECG model: 280K params, 80.7% F1, 96% AUC","aSoftMax activation lets tiny CNN screen arrhythmias at 50 Hz","Lightweight lead-shared CNN hits 96% AUC on 12-lead ECG","New activation aSoftMax for portable 12-lead arrhythmia detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the model's reported F1 and AUC reflect the same train/test split and preprocessing as the baselines, but the paper never states the split ratio, number of training epochs, batch size, or random seed.","fun_headline_variants_meta":{"raw":{"variants":["280K-param CNN with aSoftMax matches big ECG models on 9 classes","1 MB ECG model: 280K params, 80.7% F1, 96% AUC","aSoftMax activation lets tiny CNN screen arrhythmias at 50 Hz","Lightweight lead-shared CNN hits 96% AUC on 12-lead ECG","New activation aSoftMax for portable 12-lead arrhythmia detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1783,"prompt_tokens":1044,"completion_tokens":739,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":629}},"tokens_in":660,"tokens_out":739,"duration_ms":6974,"temperature":1.0,"reasoning_tokens":629,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:45:02.330527+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the model on CPSC2018 with a fully documented stratified split and identical 50 Hz preprocessing, and compare per-class F1 with LightX3ECG and the other baselines; if the average F1 is not close to 80.7% or does not beat them, the efficiency claim fails.","supporting_citations":[{"cited_title":"Bhattacharyya, S","cited_arxiv_id":null,"evidence_quote":"It supplies the CPSC2018 dataset of 6,877 12-lead recordings with nine diagnostic labels, and all reported accuracy, F1, and AUC numbers are measured on it."},{"cited_title":"Gupta, M","cited_arxiv_id":null,"evidence_quote":"It defines the Adam optimizer, and the Adamax optimizer used for training is a variant of it."},{"cited_title":"Singh, S.K","cited_arxiv_id":null,"evidence_quote":"It introduces focal loss, which is used as categorical focal cross-entropy to handle class imbalance during training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the 1D-ResNet34 baseline in the comparison and the Shapley-value interpretability approach applied to predictions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the 1D-SEResNet34 baseline in the comparison table."},{"cited_title":"Yildirim, U.B","cited_arxiv_id":null,"evidence_quote":"It provides the TI-ResNet18 baseline, an attention-based time-incremental CNN that the paper compares against."},{"cited_title":"Additionally, our model excelled in terms of storage requirements","cited_arxiv_id":null,"evidence_quote":"It provides the InceptionTime baseline in the comparison table."}],"review_version":1}