{"id":"114350ee-82e2-4ffe-b363-14868b99cc2d","arxiv_id":"2606.24974","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":3.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Spectral entropy is proposed to measure noise introduced by post-hoc XAI methods in ECG arrhythmia classification models.","lead":"This paper proposes spectral entropy as a metric to quantify noise added by explainability techniques in deep learning models for ECG arrhythmia classification. A smart generalist might read it to see a potential way to improve trust in AI explanations for medical data.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Spectral entropy may reflect overall signal complexity rather than isolating XAI-introduced noise from model signal","rationale":"The reader's weakest assumption matches the load-bearing methodological gap exactly; even with full text access the proposal lacks an explicit separation test, leaving the metric's specificity unverified. No other internal inconsistency or stronger objection appears in the central argument.","tokens_in":1579,"tokens_out":274,"duration_ms":11096,"concrete_test":"Compute spectral entropy on (1) raw ECG inputs, (2) model activations/outputs, and (3) XAI explanation maps for the same samples; if explanation entropy does not show a consistent, statistically significant elevation attributable only to the XAI step (after subtracting base-signal entropy), the isolation assumption fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that spectral entropy differences across XAI methods (e.g., post-hoc techniques on ECG arrhythmia classifiers) specifically quantify heuristic-induced noise separable from the underlying model attributions or ECG frequency content. Without a controlled validation—such as entropy computed on raw model outputs, synthetic signals with injected XAI-like perturbations, or comparison against a ground-truth noise metric—the observed entropy values could simply track the base signal's spectral properties or model decision regions rather than the claimed XAI noise component.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes spectral entropy as a quantitative measure of noise introduced by heuristics in post-hoc XAI explanations. It claims to demonstrate the measure's usefulness for distinguishing signal from noise when applying different XAI techniques to a deep learning model for arrhythmia classification on ECG data.","tokens_in":1662,"tokens_out":333,"duration_ms":15528,"significance":"A validated, separable noise metric for XAI outputs would be useful in safety-critical domains such as ECG analysis. The approach is conceptually straightforward and could complement existing attribution methods if it can be shown to isolate XAI-specific perturbations rather than reflecting base-signal complexity or model decision boundaries.","major_comments":[{"comment":"Abstract: the claim that spectral entropy 'demonstrate[s] its usefulness' is unsupported; the text supplies no methods, dataset description, XAI techniques tested, quantitative results, baselines, or statistical comparisons.","section":"Abstract"},{"comment":"The central assumption that spectral entropy differences across XAI methods isolate heuristic-induced noise from the underlying ECG frequency content or model attributions is not tested. No controlled validation (e.g., entropy on raw model outputs, synthetic signals with known perturbations, or comparison to a ground-truth noise metric) is described.","section":null}],"minor_comments":[{"comment":"Clarify the precise definition of 'signal' versus 'noise' in the XAI output and how spectral entropy is computed (windowing, normalization, frequency range).","section":null}],"recommendation":"reject","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed comments. Below we respond point-by-point to the major concerns. We acknowledge the manuscript is concise and agree that revisions are needed to strengthen the presentation and validation.","responses":[{"response":"We agree the abstract is high-level and does not contain these specifics. The full manuscript describes the ECG arrhythmia classification task, the dataset, the post-hoc XAI methods applied, and reports spectral entropy values across techniques. To address the concern directly, we will expand the abstract to summarize the dataset, methods, and key quantitative comparisons.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the claim that spectral entropy 'demonstrate[s] its usefulness' is unsupported; the text supplies no methods, dataset description, XAI techniques tested, quantitative results, baselines, or statistical comparisons."},{"response":"The current experiments apply spectral entropy to explanations produced by different XAI methods on identical ECG inputs and model outputs, with the observed differences attributed to the distinct heuristics. We acknowledge that explicit controls (raw outputs without XAI, synthetic signals, or ground-truth noise metrics) are not included. We will add a controlled validation subsection using synthetic perturbations to isolate XAI-specific effects.","revision_made":"partial","referee_comment":"The central assumption that spectral entropy differences across XAI methods isolate heuristic-induced noise from the underlying ECG frequency content or model attributions is not tested. No controlled validation (e.g., entropy on raw model outputs, synthetic signals with known perturbations, or comparison to a ground-truth noise metric) is described."}],"tokens_in":1131,"tokens_out":352,"duration_ms":23858,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core idea is to treat spectral entropy as a practical score for how much extra noise post-hoc XAI methods add when explaining ECG arrhythmia models. That framing is straightforward and addresses a real annoyance in medical XAI work.\n\nWhat the paper does is name the problem cleanly and suggest one existing signal-processing tool as a possible fix. The domain choice (ECG) is reasonable because explanations there need to be trusted.\n\nThe soft spot is the missing link between the proposal and any evidence. The abstract claims a demonstration but shows none—no dataset size, no list of XAI methods tested, no baseline entropy on raw model outputs or clean signals, and no check that the entropy differences track XAI heuristics rather than the underlying ECG spectrum or model decision boundaries. Without those controls the metric could simply reflect signal complexity, which undercuts the central claim.\n\nA reader already working on XAI evaluation metrics for time-series medical data might pick up the suggestion and try it themselves. For anyone else the piece is too thin to act on.\n\nI would not send this to peer review until the authors add the actual experiments and a direct test of whether entropy separates the claimed noise source.","headline":"The paper proposes spectral entropy to measure XAI-introduced noise on ECG explanations but supplies no methods, results, or validation that it actually isolates that noise.","tokens_in":2128,"tokens_out":317,"would_cite":false,"duration_ms":14166,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Spectral entropy quantifies noise added by explainability techniques to ECG arrhythmia classifications.","keywords":["spectral entropy","explainable AI","ECG","arrhythmia classification","signal noise","post-hoc XAI"],"falsifier":"Observing that spectral entropy values remain unchanged across XAI techniques or fail to increase when known noise is artificially added to explanations would challenge the claim.","tokens_in":2468,"feed_emoji":"","tokens_out":479,"duration_ms":14190,"temperature":0.7,"pith_summary":"The paper proposes using spectral entropy to measure the noise that post-hoc explainability techniques introduce into the outputs of deep learning models. It demonstrates this approach on an ECG dataset for classifying arrhythmias using various XAI methods. A sympathetic reader would care because healthcare models require trustworthy explanations, and distinguishing model-derived signal from heuristic-induced noise improves reliability. If the proposal holds, it offers a concrete way to evaluate and compare the quality of different explainability tools on signal data.","feed_headline":"Spectral entropy measures XAI noise in ECG explanations","feed_subtitle":"New metric separates model signal from heuristic artifacts in arrhythmia classifiers.","key_machinery":"Spectral entropy applied to XAI-generated explanations to quantify introduced signal noise.","core_discovery":"Spectral entropy serves as a measure of noise in XAI output, shown to be useful when applied to explanations from different post hoc explainability techniques in an ECG arrhythmia classification task.","pith_inferences":["Similar entropy-based measures might apply to other medical imaging or time-series domains.","Integration into XAI toolkits could help practitioners select less noisy explainers.","This raises the question of whether spectral entropy correlates with human-interpretable explanation quality."],"forward_implications":["XAI methods can be ranked by the amount of noise they add, as measured by spectral entropy on ECG data.","Explanations with lower spectral entropy are closer to the core model signal.","The method allows assessment of XAI usefulness in medical signal classification tasks."],"fun_headline_variants":["Spectral entropy quantifies XAI noise on ECG","Entropy tracks XAI noise in ECG arrhythmia classifiers","Measuring XAI signal noise via spectral entropy in ECG","Spectral entropy exposes noise in XAI ECG explanations","XAI noise quantified by spectral entropy in ECG analysis"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Spectral entropy can isolate and quantify the noise introduced by XAI heuristics separately from the underlying model signal.","fun_headline_variants_meta":{"raw":{"variants":["Spectral entropy quantifies XAI noise on ECG","Entropy tracks XAI noise in ECG arrhythmia classifiers","Measuring XAI signal noise via spectral entropy in ECG","Spectral entropy exposes noise in XAI ECG explanations","XAI noise quantified by spectral entropy in ECG analysis"]},"model":"grok-4.3","cost_usd":0.005512,"raw_usage":{"total_tokens":2556,"prompt_tokens":487,"num_sources_used":0,"completion_tokens":73,"cost_in_usd_ticks":55124500,"prompt_tokens_details":{"text_tokens":487,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1996,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":487,"tokens_out":73,"duration_ms":21958,"temperature":1.0,"reasoning_tokens":1996,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T00:58:49.130814+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Observing that spectral entropy values remain unchanged across XAI techniques or fail to increase when known noise is artificially added to explanations would challenge the claim.","supporting_citations":[],"review_version":1}