Pith. sign in

REVIEW 4 major objections 6 minor 14 references

Atrial Fibrillation Prediction Using a Lightweight Temporal Convolutional and Selective State Space Architecture

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2508.19361 v1 pith:43H66B3F submitted 2025-08-26 cs.LG cs.AI

Yongbin Lee , Ki H. Chon This is my paper
classification cs.LGcs.AI
keywords atrialfibrillationpredictionRRintervalMambatemporalconvolutionalnetworkselectivestatespacemodelparoxysmalAFwearablemonitoringdeeplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

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.

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 (4)
  1. [III, V, VI] 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.
  2. [III, Table I] 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).
  3. [III, V] 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.
  4. [IV.C, V.A] 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.
minor comments (6)
  1. [IV.B heading] Typo: 'Performance Evaluaation' should be 'Performance Evaluation.'
  2. [Author affiliations] 'United State of America' should be 'United States of America.'
  3. [V and Table III] 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.
  4. [Table IV] 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.
  5. [Fig. 1] The red and green highlighted windows are not visible in a black-and-white print; consider using distinct textures or labels.
  6. [III] 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).

Circularity Check

1 steps flagged · score 6.0 of 10

Headline 'up to two hours in advance' is inherited from label construction; underlying AF-vs-NSR classification is independent.

  1. self definitional [Abstract; Section III Dataset; Section VI Discussion]
    "Notably, the model can predict AF up to two hours in advance using just 30 minutes of input data... we used 2-hour pre-AF segments (2h before AF onset) from 151 AF subjects... To generate 30-minute input segments, all recordings were divided into non-overlapping 30-minute windows, resulting in a total of 604 AF segments... A key advantage of our approach is its ability to predict AF up to two hours before onset using only 30-minute RRI segments."

    The two-hour lead-time claim is not measured from model outputs; it is inherited from the label construction. Any 30-minute window inside the 2-hour-before-onset interval is labeled as a positive AF segment, so 'predicting AF up to two hours in advance' is equivalent to predicting membership in that predefined window set. The earliest window tested, [-120,-90) minutes, ends 90 minutes before onset, so the maximum possible demonstrated lead time is 90 minutes, not 120. Because all four windows are pooled into aggregate metrics without time-to-event evaluation, the headline 'up to two hours' restates the sampling definition rather than a model capability. The classification of pre-AF vs NSR windows is a genuine held-out prediction, so the circularity is limited to the lead-time claim.

full rationale

The paper's core classification is a genuine subject-wise held-out prediction: models are trained on disjoint subjects and evaluated on unseen subjects, with multiple random splits. No fitted parameter is relabeled as a prediction, and no load-bearing result is imported from the authors' own prior work; the comparison methods are re-implemented and trained on the same data. The one circular element is the lead-time claim. The positive class is defined as any 30-minute window within the 2-hour pre-AF interval, and the model outputs a binary pre-AF probability, not a time-to-onset value. The assertion that the model 'can predict AF up to two hours in advance' therefore reduces to the label construction, and the maximum lead time actually testable with the described windows is 90 minutes, because the earliest window ends at -90 minutes. Since all four windows are pooled in aggregate metrics, no experiment separately demonstrates a 120-minute lead time. This makes the headline partially self-definitional, but it does not invalidate the classification accuracy as such.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on label construction that defines all 2-hour pre-AF windows as positive, on the validity of NSR controls, and on hand-selected hyperparameters. No code, full preprocessing pipeline, or per-lead-time analysis is provided.

free parameters (3)
  • TCN architecture hyperparameters = channels=32, kernel=3, dilation rates 1,2,4, dropout 0.2
    Hand-selected via validation (Section IV-A) and directly affect all reported metrics.
  • Optimizer hyperparameters = lr=1e-4, weight decay=1e-4, batch size=16, patience=10
    Chosen for training (Section IV-C); not fitted to test data, but part of the undocumented tuning space.
  • Input window length and label horizon = 30-minute windows from -120 to 0 minutes
    The label construction defines the positive class; the two-hour lead-time claim is inherited from this choice, not measured independently.
assumptions (4)
  • domain assumption Pre-AF RRI windows are a learnable, homogeneous class across all four 30-minute positions.
    Section III creates non-overlapping 30-minute windows from the 2-hour pre-AF segment and labels them all as AF without lead-time distinction.
  • domain assumption NSR RR subjects are true negatives with no undiagnosed AF during recording.
    Section III uses the NSR RR dataset as negative controls; undiagnosed AF in these subjects would bias the discrimination.
  • domain assumption IRIDIA-AF annotations of AF onset are correct.
    Section III relies on IRIDIA-AF onset annotations; mislabeling would corrupt the positive windows.
  • standard math Standard deep learning training assumptions (cross-entropy loss, AdamW, early stopping) are valid.
    Section IV-C, standard optimization framework.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Atrial Fibrillation Prediction Using a Lightweight Temporal Convolutional and Selective State Space Architecture." pith.science (2026). https://pith.science/paper/43H66B3F

@misc{pith2026250819361,
  author       = {Pith},
  title        = {Pith review of: Atrial Fibrillation Prediction Using a Lightweight Temporal Convolutional and Selective State Space Architecture},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/43H66B3F}},
  note         = {Machine review of arXiv:2508.19361}
}
read the original abstract

Atrial fibrillation (AF) is the most common arrhythmia, increasing the risk of stroke, heart failure, and other cardiovascular complications. While AF detection algorithms perform well in identifying persistent AF, early-stage progression, such as paroxysmal AF (PAF), often goes undetected due to its sudden onset and short duration. However, undetected PAF can progress into sustained AF, increasing the risk of mortality and severe complications. Early prediction of AF offers an opportunity to reduce disease progression through preventive therapies, such as catecholamine-sparing agents or beta-blockers. In this study, we propose a lightweight deep learning model using only RR Intervals (RRIs), combining a Temporal Convolutional Network (TCN) for positional encoding with Mamba, a selective state space model, to enable early prediction of AF through efficient parallel sequence modeling. In subject-wise testing results, our model achieved a sensitivity of 0.908, specificity of 0.933, F1-score of 0.930, AUROC of 0.972, and AUPRC of 0.932. Additionally, our method demonstrates high computational efficiency, with only 73.5 thousand parameters and 38.3 MFLOPs, outperforming traditional Convolutional Neural Network-Recurrent Neural Network (CNN-RNN) approaches in both accuracy and model compactness. Notably, the model can predict AF up to two hours in advance using just 30 minutes of input data, providing enough lead time for preventive interventions.

Figures

Figures reproduced from arXiv: 2508.19361 by the authors.

Figure 1
Figure 1. Comparison of RRIs between (a), (b) pre-AF segments from the IRIDIA-AF dataset and (c) an NSR segment from the NSR RR dataset. (a) RRIs exhibit frequent ectopic beats, especially 30 minutes before AF onset. (b) Ectopic beats increase before the 30-minute mark, but pre-AF is more difficult to detect. (c) An NSR segment with frequent ectopic beats, resembling the pre-AF pattern in (b). TABLE II. NUMBER OF SUBJECTS (SE… view at source ↗
Figure 2
Figure 2. Our proposed AF model architecture TABLE III. PERFORMANCE METRICS OF OUR MODEL COMPARED TO PREVIOUS DEEP LEARNING APPROACHES Model Sens. Spec. F1 AUROC AUPRC Grégoire et al. [13] 0.839 0.867 0.869 0.909 0.788 Lin et al. [9] 0.930 0.892 0.905 0.965 0.907 Hannun et al. [14] 0.813 0.884 0.876 0.900 0.755 Our model 0.908 0.933 0.930 0.972 0.932 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 5 canonical work pages

  1. [1]

    Atrial Fibrillation Associated with Heart Failure, Stroke and Mortality,

    S. Bordignon, M. Chiara Corti, and C. Bilato, “Atrial Fibrillation Associated with Heart Failure, Stroke and Mortality,” J. Atr. Fibrillation, vol. 5, no. 1, p. 467, Jun. 2012, doi: 10.4022/jafib.467

    doi:10.4022/jafib.467 2012

  2. [2]

    Atrial fibrillation detection on reconstructed photoplethysmography signals collected from a smartwatch using a denoising autoencoder,

    F. Mohagheghian et al., “Atrial fibrillation detection on reconstructed photoplethysmography signals collected from a smartwatch using a denoising autoencoder,” Expert Syst. Appl., vol. 237, p. 121611, Mar. 2024, doi: 10.1016/j.eswa.2023.121611

    work page arXiv 2024

  3. [3]

    Cooperative PPG/ECG Wearable System for Atrial Fibrillation Diagnosis,

    Y. Lee, S. Lee, S. Kyu Kim, D. Keon Yon, Y. Nam, and J. Lee, “Cooperative PPG/ECG Wearable System for Atrial Fibrillation Diagnosis,” IEEE Sens. J., vol. 25, no. 4, pp. 7331 –7344, Feb. 2025, doi: 10.1109/JSEN.2025.3526245

    arXiv 2025

  4. [4]

    Atrial Fibrillation Detection from Wrist Photoplethysmography Signals Using Smartwatches,

    S. K. Bashar et al. , “Atrial Fibrillation Detection from Wrist Photoplethysmography Signals Using Smartwatches,” Sci. Rep., vol. 9, no. 1, p. 15054, Oct. 2019, doi: 10.1038/s41598-019-49092-2

    doi:10.1038/s41598-019-49092-2 2019

  5. [5]

    Atrial Fibrillation in the ICU,

    N. A. Bosch, J. Cimini, and A. J. Walkey, “Atrial Fibrillation in the ICU,” Chest, vol. 154, no. 6, pp. 1424 –1434, Dec. 2018, doi: 10.1016/j.chest.2018.03.040

    work page doi:10.1016/j.chest.2018.03.040 2018

  6. [6]

    Artificial intelligence for atrial fibrillation detection, prediction, and treatment: A systematic review of the last decade (2013–2023),

    M. Salvi et al., “Artificial intelligence for atrial fibrillation detection, prediction, and treatment: A systematic review of the last decade (2013–2023),” WIREs Data Min. Knowl. Discov. , vol. 14, no. 3, p. e1530, 2024, doi: 10.1002/widm.1530

    work page doi:10.1002/widm.1530 2013

  7. [7]

    Temporal Convolutional Networks for Action Segmentation and Detection,

    C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager, “Temporal Convolutional Networks for Action Segmentation and Detection,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI: IEEE, Jul. 2017. doi: 10.1109/cvpr.2017.113

    doi:10.1109/cvpr.2017.113 2017

  8. [8]

    Mamba: Linear -Time Sequence Modeling with Selective State Spaces,

    A. Gu and T. Dao, “Mamba: Linear -Time Sequence Modeling with Selective State Spaces,” May 31, 2024, arXiv: arXiv:2312.00752. doi: 10.48550/arXiv.2312.00752

    arXiv doi:10.48550/arxiv.2312.00752 2024

Show all 14 references
  1. [9]

    Artificial -intelligence-based risk prediction and mechanism discovery for atrial fibrillation using heart beat -to-beat intervals,

    F. Lin et al. , “Artificial -intelligence-based risk prediction and mechanism discovery for atrial fibrillation using heart beat -to-beat intervals,” Med, vol. 5, no. 5, pp. 414 -431.e5, May 2024, doi: 10.1016/j.medj.2024.02.006

    2024 doi

  2. [10]

    PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals,

    A. L. Goldberger et al., “PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals,” Circulation, vol. 101, no. 23, pp. E215 -220, Jun. 2000, doi: 10.1161/01.cir.101.23.e215

    doi

  3. [11]

    IRIDIA-AF, a large paroxysmal atrial fibrillation long -term electrocardiogram monitoring database | Scientific Data

    “IRIDIA-AF, a large paroxysmal atrial fibrillation long -term electrocardiogram monitoring database | Scientific Data.” Accessed: Jul. 10, 2025. [Online]. Available: https://www.nature.com/articles/s41597-023-02621-1

    2025

  4. [12]

    ECGMamba: Towards Efficient ECG Classification with BiSSM,

    Y. Qiang, X. Dong, X. Liu, Y. Yang, Y. Fang, and J. Dou, “ECGMamba: Towards Efficient ECG Classification with BiSSM,” Jun. 14, 2024, arXiv: arXiv:2406.10098. doi: 10.48550/arXiv.2406.10098

    2024 doi arXiv

  5. [13]

    Role of the autonomic nervous system and premature atrial contractions in short - term paroxysmal atrial fibrillation forecasting: Insights from machine learning models,

    J.-M. Grégoire, C. Gilon, S. Carlier, and H. Bersini, “Role of the autonomic nervous system and premature atrial contractions in short - term paroxysmal atrial fibrillation forecasting: Insights from machine learning models,” Arch. Cardiovasc. Dis., vol. 115, no. 6, pp. 377–38...

    2022 doi

  6. [14]

    Cardiologist -level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network,

    A. Y. Hannun et al. , “Cardiologist -level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network,” Nat. Med. , vol. 25, no. 1, pp. 65 –69, Jan. 2019, doi: 10.1038/s41591-018-0268-3. TABLE IV. MODEL COMPLEXITY COMPARISON IN TERMS O...

    2019 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.