{"id":"11722023-cd18-438d-95c4-507af4c85aae","arxiv_id":"2508.19361","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A TCN+Mamba model trained on 30-minute RR interval windows predicts imminent atrial fibrillation in held-out subjects with AUROC 0.972 and only 73.5K parameters.","lead":"The paper trains a lightweight TCN+Mamba model on 30-minute RR interval windows to flag the two hours before atrial fibrillation onset, reporting high subject-wise accuracy and low computational cost. A generalist would read it because early AF warning could enable preventative treatment and wearable deployment.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'two hours in advance' claim is unsupported: the earliest 30-min input window ends 90 min before onset, so the maximum possible lead time is 90 minutes, not 120.","rationale":"The reader's weakest assumption focused on whether all four windows carry equal signal, suggesting that aggregate metrics could be dominated by windows closest to onset. That is a legitimate concern, but my analysis reveals a more fundamental and specific problem: even under the most favorable interpretation, the experimental design cannot support a 120-minute lead time because the earliest input window ends 90 minutes before onset. The lead-time claim is therefore not merely an evidential gap—it is an internal mismatch between the stated prediction horizon and the data construction in Section III. This directly undermines the central claim in the abstract (the 'notably' sentence) and the conclusion. I agree with the reader that the paper merits a conditional verdict: the architecture and subject-wise protocol are plausible, but the headline lead-time claim must be corrected and the dataset inconsistencies (e.g., 2,380 vs. 595 NSR segments; 151 vs. 152 AF subjects) resolved. My concern is an additional, sharper reason for the same conditional verdict, not a change in its direction.","tokens_in":6333,"tokens_out":9783,"duration_ms":103550,"concrete_test":"Stratify the test-set metrics by the four 30-minute windows ([-120,-90), [-90,-60), [-60,-30), [-30,0) relative to onset) and report sensitivity/specificity per window. If the earliest window (ending at -90 min) is near chance, the two-hour claim fails. If it is not near chance, the claim should still be downgraded to at most 90 minutes unless a separate evaluation uses a 30-minute input window ending at -120 min before onset. If such data are unavailable, the abstract's 'two hours in advance' must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the model 'can predict AF up to two hours in advance using just 30 minutes of input data.' However, the experimental setup makes this impossible. The authors use 2-hour pre-AF segments (from 2h before onset to onset) and divide them into four non-overlapping 30-minute windows: [-120,-90), [-90,-60), [-60,-30), and [-30,0) minutes relative to onset (Section III). A prediction is made at the end of each input window, so the window ending at -90 minutes provides a lead time of 90 minutes, not 120. To claim a 2-hour lead time, one would need a 30-minute input window ending at -120 minutes (e.g., [-150,-120)), which is never used in training or evaluation. Even if the model performs perfectly on the earliest window tested, the maximum lead time demonstrated is 90 minutes. The abstract, introduction, and conclusion explicitly state 'two hours in advance,' so this is not a minor wording issue—the headline claim is not supported by the described procedure. Additionally, the model is only tested against NSR subjects who never have AF, not against AF patients' recordings far from an AF episode; therefore, high performance may partly reflect identifying AF patients rather than the dynamic approach to onset. This confound further weakens the lead-time interpretation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a lightweight TCN-Mamba-FC architecture that predicts atrial fibrillation (AF) from 30-minute RR-interval segments. Using the IRIDIA-AF and NSR RR datasets with subject-wise 60/20/20 splits repeated over five random seeds, the authors report high mean test performance (sensitivity 0.908, specificity 0.933, F1 0.930, AUROC 0.972, AUPRC 0.932) and low model complexity (73.5K parameters, 38.3 MFLOPs). The central claim is that the model can predict AF up to two hours in advance from a 30-minute input window, enabling preventive intervention.","tokens_in":6706,"tokens_out":2830,"duration_ms":33169,"significance":"If the headline lead-time claim were supported, this would be a clinically valuable and computationally efficient result for wearable AF prediction. The subject-wise evaluation protocol, repeated random splits, re-implemented baseline comparisons, and explicit reporting of parameters and FLOPs are strengths. However, the central 'two hours in advance' claim is not supported by the experimental design as described, and the control population does not isolate dynamic pre-AF changes from subject-level AF propensity. The core architecture and evaluation pipeline are promising, but the lead-time claim must be re-derived and re-validated before the contribution can be assessed.","major_comments":[{"comment":"The headline claim that the model predicts AF 'up to two hours in advance' is not supported by the described procedure. The 2-hour pre-AF segment is divided into four non-overlapping 30-minute windows: [-120,-90), [-90,-60), [-60,-30), and [-30,0) min relative to onset. A prediction from the earliest window ends at -90 min, so the maximum achievable lead time is 90 minutes, not 120. Moreover, metrics are pooled across all four windows, so even the 90-minute capability is not demonstrated unless results are reported per lead-time bin. Please report sensitivity, specificity, F1, AUROC, and AUPRC separately for each of the four windows, and revise the abstract, introduction, and conclusion to state the maximum lead time actually supported by the data.","section":"III, V, VI"},{"comment":"There is an inconsistency in the number of AF subjects: Table I lists 152 IRIDIA-AF subjects, while the Section III text says '151 AF subjects.' Please reconcile this and clarify how the 604 AF segments arise from the subject/episode counts, including whether multiple episodes per subject are used and whether any subject appears in both AF and NSR groups (the latter seems impossible given the datasets, but the segment-level counts need clarification).","section":"III, Table I"},{"comment":"The control group consists only of NSR subjects with no documented AF episodes. The AF class is composed exclusively of pre-AF segments. Consequently, the model may be learning subject-level AF propensity or differences between AF and NSR populations rather than the dynamic progression toward AF onset. To support a lead-time interpretation, include as negative controls RRI segments from AF patients recorded far from any AF episode (e.g., more than 2 hours before onset, or after episode termination) and show that the model does not simply classify 'this is an AF patient.' Without this control, the reported discrimination between pre-AF and NSR does not establish early prediction of imminent AF.","section":"III, V"},{"comment":"The baseline comparisons are described as 're-implemented... based on the architecture described in the original publications,' but the original models often used different input representations, demographic features, or optimization schemes. The manuscript states that demographic features were excluded from Lin et al. but does not state whether the same 30-minute RRI windows, same preprocessing, and same class-balancing were used for all baselines. Please specify the exact common protocol and any architecture-specific adaptations, because the reported performance gap could otherwise reflect implementation choices rather than architectural superiority.","section":"IV.C, V.A"}],"minor_comments":[{"comment":"Typo: 'Performance Evaluaation' should be 'Performance Evaluation.'","section":"IV.B heading"},{"comment":"'United State of America' should be 'United States of America.'","section":"Author affiliations"},{"comment":"The reported metrics are means over five splits, but no standard deviations or per-split ranges are given. Reporting variability would help assess stability across the random subject-wise splits.","section":"V and Table III"},{"comment":"The FLOPs calculation is not described. Please specify the input size and the counting method (e.g., multiply-accumulates vs. floating-point operations) so that the comparison is reproducible.","section":"Table IV"},{"comment":"The red and green highlighted windows are not visible in a black-and-white print; consider using distinct textures or labels.","section":"Fig. 1"},{"comment":"The phrase 'all recordings from 54 NSR subjects' is ambiguous: the NSR RR dataset contains 24-hour recordings, and later the text says all recordings were divided into 30-minute windows. Please clarify how many NSR windows are used per subject and how the 2,380 NSR segments are obtained (54 × 24 h × 2 windows/h = 2,592, so some exclusion or overlapping scheme must be in effect).","section":"III"}],"recommendation":"major_revision","confidential_remarks":"The paper's core architecture and evaluation protocol are reasonable, and the reported computational efficiency is credible. The main issue is the two-hour lead-time claim, which is not merely a wording problem but is contradicted by the window construction. I believe the authors can address it by re-analyzing per lead-time bin, correcting the headline, and adding far-from-event AF controls. If those changes are made, the paper would be a solid candidate for acceptance. I would also encourage the authors to make code and trained model available to support reproducibility, but I do not make that a condition for revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the model work is plausible and the baseline comparison is fairly done, but the 'two hours in advance' claim is not supported by the experiment. The four input windows run from -120 to -90, -90 to -60, -60 to -30, and -30 to 0 minutes relative to onset. Since the prediction is made at the end of each window, the earliest decision point is 90 minutes before onset, not 120. To support a two-hour lead time you'd need a window ending at -120 (e.g., -150 to -120), which is never used. The metrics are pooled across all four windows, so we also don't know whether the -90-minute window contributes or whether the model only works in the hour right before onset. That is the paper's load-bearing claim, and it's not backed.\n\nCredit where due: the subject-wise split is the right protocol, they re-implemented three earlier models under identical conditions, ran five random splits, and report a genuinely light architecture (73.5K parameters, 38.3 MFLOPs). The TCN-Mamba combination for RRI-based AF prediction appears to be new. That is a legitimate engineering contribution.\n\nOther soft spots: the negative control is NSR subjects with no AF, not AF patients' recordings far from an episode, so the model may be discriminating patient state rather than pre-AF dynamics. There are dataset accounting inconsistencies (151 vs 152 AF subjects; the 2,380 NSR segment total doesn't match the split segment counts). No standard deviations are reported, and no code/data release. These are fixable.\n\nThe fix is also straightforward: report per-window performance and either claim the supported lead time (probably 90 minutes) or extend the dataset to include windows that end further in advance. If the earliest window does work, that's still clinically valuable.\n\nOverall, this is a solid engineering study with an overstated abstract. The evaluation is not as strong as the claims, but the core model is plausible and the comparison is honest. It deserves a serious referee; I'd send it to peer review conditional on major revision focusing on the lead-time analysis and the control group.","headline":"The model is a plausible lightweight AF predictor, but the 'up to two hours in advance' claim is not supported by the experimental design; the earliest tested window ends 90 minutes before onset.","tokens_in":7101,"tokens_out":4201,"would_cite":false,"duration_ms":44621,"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":"This paper proposes a 73.5K-parameter TCN-Mamba network that predicts atrial fibrillation up to two hours before onset from 30 minutes of heartbeat intervals, reaching AUROC 0.972.","keywords":["atrial fibrillation prediction","RR interval","Mamba","temporal convolutional network","selective state space model","paroxysmal AF","wearable monitoring","deep learning"],"falsifier":"Take the held-out test windows and compute AUROC separately for each 30-minute interval (120–90, 90–60, 60–30, and 30–0 minutes before onset); if the earliest windows perform near chance while only the window closest to onset performs well, the two-hour lead-time claim fails, and if all four intervals perform similarly, the claim holds.","tokens_in":6283,"feed_emoji":"🫀","tokens_out":9738,"duration_ms":91515,"temperature":0.7,"pith_summary":"This paper tries to show that atrial fibrillation (AF) can be predicted, not just detected, using a small deep network that reads only the time between successive heartbeats (RR intervals). The authors combine a temporal convolutional network (TCN) with Mamba, a selective state space model, and train it on 30-minute intervals cut from the two hours before recorded AF episodes. In subject-wise testing across five random splits, the model reaches a sensitivity of 0.908, specificity of 0.933, F1 of 0.930, AUROC of 0.972, and AUPRC of 0.932, while using only 73.5 thousand parameters and 38.3 million FLOPs. The practical payoff the authors emphasize is that a wearable could warn of AF onset two hours in advance, leaving time for preventive treatment. That result matters because paroxysmal AF is often missed until it becomes persistent, and current detection-focused algorithms only flag AF once it has already started.","feed_headline":"30-minute heartbeat data predicts atrial fibrillation two hours early","feed_subtitle":"A 73.5K-parameter TCN-Mamba model reaches AUROC 0.972 and beats CNN-RNN baselines—making two-hour AF warnings fit on wearables.","key_machinery":"The load-bearing object is the TCN-Mamba-FC stack. The TCN is a stack of three causal dilated convolutional blocks (kernel size 3, dilation rates 1, 2, 4) with residual connections; its role is to encode gradual changes—such as the rising ectopic beats the authors show in pre-AF recordings—without letting future samples leak into the prediction. Mamba, a selective state space model, then models temporal dependencies over the encoded sequence in linear time and in parallel, avoiding the sequential bottleneck of RNNs and the quadratic cost of attention; a convolution-based feed-forward network inside the Mamba block captures local dynamics. Global average and max pooling are concatenated and p","core_discovery":"The central claim is that the pre-AF state is distinguishable from normal sinus rhythm from RR intervals alone, and that a TCN-Mamba classifier can make that distinction early enough to act. The model labels every non-overlapping 30-minute window drawn from the two hours before AF onset as positive, and windows from outpatient normal-sinus recordings as negative; subjects are split so that the same person never appears in both training and test sets. On held-out subjects the model reports AUROC 0.972 and AUPRC 0.932, with sensitivity 0.908 and specificity 0.933, outperforming re-implemented CNN-BiGRU, CNN-BiLSTM, and deep residual CNN baselines on most metrics while using far fewer parameter","pith_inferences":["The two-hour lead time is reported as aggregate metrics over all four 30-minute windows in the pre-AF interval; the paper does not break down performance by time-to-onset. A window-resolved analysis (AUROC for the 120–90, 90–60, 60–30, and 30–0 minute windows) would show how much of the signal is truly early versus a near-onset warning. This is my inference, not a result in the paper.","If early windows carry little signal, the method might be better described as early detection of an imminent episode than as a two-hour prediction; a clinician would need the per-window horizon before trusting the lead-time claim.","The same TCN-Mamba backbone could apply to other paroxysmal events with a defined onset, such as seizure or hypoglycemia prediction, because the architecture is not AF-specific; the paper does not test this.","The visually striking rise in ectopic beats suggests that adding hand-crafted features like premature atrial contraction counts or heart-rate variability could sharpen early-window performance; the paper does not evaluate this."],"forward_implications":["If the two-hour lead time is real, wearable monitors could alert patients before AF onset rather than after, creating a window for preventive therapy such as beta-blockers.","Because only RR intervals are needed, the input can be derived from single-lead ECG or photoplethysmography, avoiding multi-sensor wearables.","The parameter and FLOP counts are low enough that the model can run continuously on-device, making real-time outpatient monitoring feasible.","Subject-wise splitting across five random partitions suggests the reported accuracy is not an artifact of a single favorable train/test split.","RRI-only prediction could be integrated into existing arrhythmia detectors as an early-warning pre-filter without adding heavy computation."],"supporting_citations":[{"why":"Supplies the temporal convolutional network (TCN) design that serves as the positional encoder.","marker":"[7]"},{"why":"Supplies the Mamba selective state space model used for parallel linear-time sequence modeling.","marker":"[8]"},{"why":"Supplies the CNN-BiLSTM baseline that the model is compared against, and the NSR RR dataset description.","marker":"[9]"},{"why":"Supplies the public normal sinus rhythm RR recordings used as negative examples.","marker":"[10]"},{"why":"Supplies the outpatient AF recordings with annotated episodes used to build the two-hour pre-AF positive windows.","marker":"[11]"},{"why":"Motivates the convolution-based feed-forward network inside the Mamba block for temporal dynamics.","marker":"[12]"},{"why":"Supplies the CNN-BiGRU baseline re-implemented and compared under the same training protocol.","marker":"[13]"},{"why":"Supplies the deep residual CNN arrhythmia-classification baseline used as a complexity and accuracy comparison.","marker":"[14]"}],"fun_headline_variants":["Two-hour AF head start from 30 minutes of heartbeats","AF predicted 2 hours ahead by 73K-parameter model","RR intervals alone give 2-hour AF prediction edge","Mamba-TCN model forecasts AF two hours before onset"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The paper assumes every 30-minute segment inside the two hours before an AF episode is equally a positive example, so the reported two-hour lead time could be inflated if only the segments closest to onset actually carry the signal.","fun_headline_variants_meta":{"raw":{"variants":["Two-hour AF head start from 30 minutes of heartbeats","AF predicted 2 hours ahead by 73K-parameter model","RR intervals alone give 2-hour AF prediction edge","Mamba-TCN model forecasts AF two hours before onset"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000301,"raw_usage":{"total_tokens":1622,"prompt_tokens":840,"completion_tokens":782,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":712}},"tokens_in":584,"tokens_out":782,"duration_ms":8959,"temperature":1.0,"reasoning_tokens":712,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:47:24.788038+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the held-out test windows and compute AUROC separately for each 30-minute interval (120–90, 90–60, 60–30, and 30–0 minutes before onset); if the earliest windows perform near chance while only the window closest to onset performs well, the two-hour lead-time claim fails, and if all four intervals perform similarly, the claim holds.","supporting_citations":[{"cited_title":"Artificial -intelligence-based risk prediction and mechanism discovery for atrial fibrillation using heart beat -to-beat intervals,","cited_arxiv_id":null,"evidence_quote":"Supplies the CNN-BiLSTM baseline that the model is compared against, and the NSR RR dataset description."},{"cited_title":"IRIDIA-AF, a large paroxysmal atrial fibrillation long -term electrocardiogram monitoring database | Scientific Data","cited_arxiv_id":null,"evidence_quote":"Supplies the outpatient AF recordings with annotated episodes used to build the two-hour pre-AF positive windows."}],"review_version":1}