AF-Mamba: Efficient Long-Term Signal Modeling for Early Prediction of Atrial Fibrillation Onset
cs.LG
Quick verdict
TCN+Mamba on public RRIs reports strong subject-wise 1h-pre-AF results (AUROC 0.974) with honest cross-dataset tests, but separate AF/NSR databases leave dataset confounds open.
Summary
AF-Mamba is a hybrid TCN plus Mamba architecture that predicts atrial fibrillation onset one hour ahead from one-hour RR-interval windows. Using five public Holter datasets, strict subject-wise 5-fold cross-validation, and a validation-set threshold targeting 0.90 sensitivity, it reaches AUROC 0.974, sensitivity 0.889, and specificity 0.943 with a compact 34.6K-parameter model (2.75 ms inference). In paired cross-dataset holdout across all six unseen AF-NSR database pairs, AUROC varies widely (0.797-0.991) and threshold-dependent metrics collapse, showing decision thresholds do not transfer across sources. Ectopy stratification and GEE adjustment for five HRV irregularity measures suggest discrimination beyond ectopic burden. The main caveat is that AF and NSR windows come from disjoint databases, so dataset confounding cannot be excluded, and validation remains retrospective.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 4
● low 0
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A hybrid TCN plus Mamba architecture predicts AF onset one hour in advance from one-hour RRI windows, achieving Sens 0.889, Spec 0.943, AUROC 0.974, AUPRC 0.933 under subject-wise 5-fold evaluation (Table III).
- AF-Mamba offers a favorable performance-efficiency trade-off for long sequences: 34.6K parameters, 127.0M FLOPs, and 2.75 ms latency, versus 955.3M-1780.5M FLOPs for Transformer variants on the same 1-h input (Table IV).
- Discrimination is retained when both the AF and NSR source databases are unseen during development, with mean AUROC 0.897±0.070 across all six paired holdouts (Table V).
- Model predictions are not explained by ectopic burden or five conventional RR-irregularity measures: discrimination persists in low-ectopy windows (AUROC 0.842) and the score remains associated with pre-AF status after GEE adjustment (OR 2.20).
- The score reflects a sustained state of elevated pre-AF susceptibility rather than a temporally precise 1-h onset estimate (Fig. 5).
Strengths
Rigorous subject-independent evaluation protocol
Strict subject-wise 5-fold CV with mutually exclusive splits (Sec. III-B, IV-E), non-overlapping 1-h windows, and a decision threshold fixed on validation at Sens 0.90 and applied unchanged to test — avoiding common leakage pitfalls in this literature.
Generalization testing beyond in-sample CV
A paired cross-dataset holdout over all six AF-NSR database combinations (Table V, AUROC 0.797-0.991, mean 0.897±0.070), a careful design given each source contributed only one class.
Confounder-hunting analyses rarely seen in ML-first AF prediction papers
Unusually honest reporting
Weaknesses
Structural dataset confounding is acknowledged but not resolvable within this design
The 1-h framing is weaker than the trajectory evidence
No variance or significance testing for the primary comparison
Ectopy-independence evidence rests on very small strata
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
AF-Mamba — Structured Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is AF-Mamba and how does it predict atrial fibrillation before it starts?
AF-Mamba is a deep learning model that processes one-hour sequences of RR intervals (beat-to-beat timings extracted from ECG Holter recordings) to flag windows recorded one hour before an atrial fibrillation onset. It combines temporal convolutional networks for local patterns with a Mamba state-space layer for long-range context, and is very lightweight: about 35 thousand parameters and roughly 2.75 ms per inference on a GPU.
How accurate is AF-Mamba?
In strict subject-independent five-fold testing it reached AUROC 0.974, sensitivity 0.889 and specificity 0.943. When tested on database pairs entirely unseen during training, AUROC stayed between 0.797 and 0.991, but threshold-dependent metrics varied greatly, so the alerting threshold would need recalibration for a new source.
Is the model just detecting extra heartbeats (PACs/PVCs)?
The authors provide evidence it is not: the model still discriminated pre-AF windows with very low ectopic burden (AUROC 0.842 in their low-ectopy subgroup), and its score remained associated with pre-AF status after statistically adjusting for five heart-rate-variability irregularity measures (adjusted odds ratio 2.20).
Can I reproduce the results and is the code available?
Yes, per the paper: the code is on GitHub (yongbin98/AF_Mamba) and all five datasets are public PhysioNet/PhysioBank resources. Preprocessing, architecture, and hyperparameter search space are documented, so a re-run is feasible, though class-imbalance handling during training is not detailed in the main text.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 are present in the source PDF but only as images; their numeric content is unavailable to this review, so figure-dependent judgments (e.g., calibration curve shape) rely on the in-text descriptions.
- Supplementary Tables S1-S7 are referenced by the main text but were not part of the extracted material; all cited numbers here come from the main text and Tables I-VII.
- Table I appears to contain a typographical error: the MIT-BIH NSR sex ratio is printed as 5/13, reversed relative to the underlying public dataset documentation; this affects no reported result.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.