{"id":"7a808c3e-04fd-4440-affc-3aa0f5c633e3","arxiv_id":"2504.13233","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Auto-FEDUS, an autoregressive convolutional model, generates realistic Doppler ultrasound signals from fetal ECGs, with morphology and heart rate estimates similar to real data.","lead":"A new deep learning model converts fetal ECG signals into Doppler ultrasound waveforms, creating realistic synthetic fetal heart monitoring data. The model could help overcome the shortage of labeled Doppler ultrasound data for fetal health monitoring in low-resource settings.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FHR validation is circular: input FECG dictates heart rate, so Bland-Altman agreement is trivially expected and does not demonstrate DUS morphological realism.","rationale":"The reader's weakest assumption focuses on generalization from only 5 subjects and 3,923 beats, which is a valid concern. My stress-test identifies a more specific and arguably more decisive flaw: the downstream FHR evaluation is circular because the input FECG encodes the exact heart rate used as the label. This does not change the overall verdict (conditional acceptance with requests for stronger validation), but it adds a concrete reason why the headline claim of 'comparable FHR results' and 'realism' is not supported by that experiment. The metrics in Table 1 (RMSE, MAE, FD, etc.) and qualitative figures remain the primary evidence, but they are averaged over the acknowledged subject- and time-dependent variability of DUS and are reported without statistical significance tests across the five folds. The near-zero CD and SF values in Table 1 also warrant scrutiny. I therefore agree with the reader's conditional verdict but sharpen the rationale: the paper needs either non-circular downstream validation or explicit demonstration that generated waveforms carry morphology beyond timing.","tokens_in":12125,"tokens_out":5984,"duration_ms":54896,"concrete_test":"Generate 3.75 s DUS segments with a baseline that places a single averaged real DUS beat at the FECG R-peak locations (correct timing, no beat-to-beat morphology variation). Run the same FHR estimator and compute Bland-Altman/PICP against the FECG-derived labels. If this timing-only baseline achieves comparable agreement (e.g., <=4.5 bpm, PICP >=96%), the FHR test is not discriminative and the realism claim loses its strongest quantitative support. Also report the same metrics for real DUS segments as a sanity check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest quantitative evidence for realism is the use-case FHR agreement (Bland-Altman 4.5 bpm, PICP 96.7%, Table 3). This evidence is circular. Generated 3.75 s DUS segments are produced by feeding consecutive FECG heartbeats into Auto-FEDUS; the input FECG therefore fully determines the timing (FHR) of the output DUS, and the reference label is derived from the same FECG R-peaks. Any model that outputs a periodic signal at the input heart rate, regardless of waveform fidelity, will show high FHR agreement. The test measures only periodicity preservation, not morphological realism. The paper itself concedes in the Qualitative Evaluation that DUS signals are highly variable and 'even for relatively similar FECG signals, their structures can vary over time,' so the FECG-to-DUS mapping is not deterministic. The quality-model result (all 303 segments 'good') is similarly uninformative without a baseline for real segments. The central claim of clinical utility thus rests on validation that cannot distinguish a realistic DUS waveform from a timing-carrying template.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Auto-FEDUS, an autoregressive convolutional model (based on WaveNet-style dilated causal convolutions) that maps fetal electrocardiogram (FECG) heartbeats to corresponding Doppler ultrasound (DUS) waveforms. The authors evaluate the model with leave-one-subject-out cross-validation on a dataset of five subjects (3,923 beats), comparing against an autoencoder, LSTM-GAN, DCGAN, and WGAN-GP using eight time/frequency domain metrics. They also generate 3.75 s DUS segments by feeding consecutive FECG heartbeats, then assess these segments with an external DUS quality classifier and a fetal heart rate (FHR) estimation model, reporting that all generated segments are classified as good quality and that FHR agreement with labels is comparable to real signals (Bland-Altman 4.5 bpm, PICP 96.7%). The central claim is that Auto-FEDUS produces realistic DUS signals that can augment scarce DUS datasets for fetal monitoring applications.","tokens_in":12373,"tokens_out":4074,"duration_ms":38951,"significance":"If the claims were fully validated, the work would be a useful contribution to a genuinely under-explored area: synthetic DUS generation for fetal monitoring, with a sensible subject-independent evaluation protocol and a direct practical motivation (data scarcity). The paper also provides a broad comparison against several generative baselines, which is a strength. The leave-one-out scheme is appropriate given the small dataset, and the idea of using external quality and FHR models as downstream validators is sound in principle. However, the current evidence does not establish the main claim of morphological realism: the FHR validation is largely circular, the quality-model result lacks a baseline on real signals, and the quantitative metrics show small differences with overlapping standard deviations and no significance testing. These issues are fixable and do not invalidate the approach itself, but they currently undermine the paper's central assertions.","major_comments":[{"comment":"The FHR-based validation is circular. The generated 3.75 s segments are produced by feeding consecutive FECG heartbeats into Auto-FEDUS, and the label FHR is derived from the R-peaks of that same FECG. The input FECG therefore fully determines the timing (and hence the heart rate) of the output DUS; any model that outputs a periodic signal at the input heart rate, regardless of waveform fidelity, will show high FHR agreement. The Bland-Altman limit of 4.5 bpm and PICP of 96.7% in Table 3 only demonstrate periodicity preservation, not morphological realism. The authors should either remove the FHR agreement as evidence for realism, or compare against a trivial baseline that re-synthesizes DUS at the input heart rate and show that Auto-FEDUS achieves significantly better morphology metrics than that baseline.","section":"Use Case Scenario; Appendix B; Figure 4(c); Table 3"},{"comment":"The quality assessment result that all 303 generated segments are classified as 'good' is not interpretable without a baseline. The authors do not report the quality-model's classification distribution for the real DUS segments in their test set. If the quality model is biased toward 'good', or if the real segments have a different distribution (e.g., many are 'poor' or 'interference'), the result would not support the claim that generated signals match real ones. Please report the quality-model breakdown on real segments and, ideally, the score distributions for real vs. generated signals.","section":"Use Case Scenario; first paragraph"},{"comment":"The claim that Auto-FEDUS outperforms the baselines is supported by small metric gaps with overlapping standard deviations, and no significance tests are reported. For example, RMSE is 0.20 ± 0.01 for Auto-FEDUS versus 0.22 ± 0.01 for WGAN-GP, MAE is 0.14 ± 0.01 versus 0.15 ± 0.01, and KLD is 0.05 ± 0.01 versus 0.03 ± 0.02 (the latter actually favoring WGAN-GP). With only five leave-one-out folds, the authors should report per-fold paired comparisons, confidence intervals, or a paired significance test (e.g., Wilcoxon signed-rank) to substantiate the superiority claim.","section":"Quantitative Evaluation and Analysis; Table 1"},{"comment":"The paper concedes that 'DUS signals are highly variable; even for relatively similar FECG signals, their structures can vary over time,' yet Auto-FEDUS is a deterministic mapping trained with MSE loss. Such a model can only produce the conditional mean of the DUS distribution given an FECG beat, which is at odds with the acknowledged variability and will tend to over-smooth high-frequency components. The lower PSD of generated signals in Figure 4(a) is consistent with this concern. The authors should either make the model stochastic (e.g., by modeling a conditional distribution) or explicitly justify why a deterministic conditional-mean mapping is sufficient for the targeted downstream tasks, and evaluate the effect on morphological fidelity.","section":"Qualitative Evaluation; Methodology; Figure 4(a)"}],"minor_comments":[{"comment":"There is a missing space in 'channelAs a result' — it should read 'channel. As a result'.","section":"Dataset paragraph"},{"comment":"The phrase '0.97 milli seconds' should be '0.97 milliseconds'.","section":"Training Details"},{"comment":"The caption should explicitly state that FHRLabel is computed from FECG R-peaks of the input signal, which helps readers understand the circularity concern in the FHR comparison.","section":"Figure 4 caption"},{"comment":"The phrase 'subject-wise training and evaluation' is ambiguous; please clarify whether the 2- and 3-beat ablation used subject-independent folds or within-subject training, since this affects interpretation.","section":"Appendix B, first sentence"},{"comment":"The description of the autoencoder appears to invert the encoder and decoder roles: the decoder is described with max pooling and the encoder with upsampling. Please check and correct the descriptions.","section":"Related Generative Models"},{"comment":"Only two qualitative examples are shown in Figure 3; given the acknowledged variability, including failure cases or a larger random sample would strengthen the visual assessment.","section":"Qualitative Evaluation"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the thing to know: this is the first published FECG-to-DUS synthesis model, and it works reasonably well at the waveform level given the data size. The architecture is a straightforward WaveNet adaptation, so the novelty is the application and the preprocessing pipeline, not the model class.\n\nGood parts: subject-independent leave-one-out evaluation is appropriate; they manually annotated and quality-gated the data, which matters for noisy fetal signals; and they compare against a reasonable set of baselines (autoencoder, LSTM-GAN, DCGAN, WGAN-GP). On most time/frequency metrics Auto-FEDUS is best, though the gaps are small and no significance tests are reported.\n\nSoft spots: the use-case validation is weaker than the paper suggests. The Bland-Altman FHR agreement is partly circular because the generated DUS inherits its timing from the input FECG, and the label is derived from the same FECG peaks. A model that emits a periodic template at the input rate would pass that test without capturing real DUS morphology. The quality-model result is also uninformative without a real-segment baseline on the same classifier. The dataset is five subjects; 3,923 heartbeats from five subjects means the generalizability claim rests on very little. No code or data is released, which makes it hard to judge the reproducibility of the preprocessing and annotation steps.\n\nProportion: these are real issues but not fatal. The paper's central contribution—that a causal dilated convolutional model can map FECG to plausible DUS waveforms—survives. The circularity inflates the realism claim but doesn't undermine the basic feasibility result.\n\nWho it's for: people working on fetal monitoring signal generation, data augmentation, or cross-modal physiological translation. As a workshop paper it's a reasonable contribution.\n\nRecommendation: deserves peer review, but the authors should add significance tests, compare the quality classifier on real versus generated segments, and ideally release the model and annotations. The FHR use case should be reframed as a periodicity-preservation check, not a realism proof.","headline":"A useful first FECG-to-DUS synthesis result, but the realism claims rest on validation that is partly circular and a very small dataset.","tokens_in":12852,"tokens_out":1913,"would_cite":false,"duration_ms":18913,"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":"Fetal ECG alone can drive realistic Doppler ultrasound waveform generation, down to heart-rate agreement within 4.5 bpm.","keywords":["fetal monitoring","Doppler ultrasound","fetal electrocardiogram","autoregressive generative model","dilated causal convolutions","cross-modal signal synthesis","heart rate estimation","data augmentation"],"falsifier":"Generate a new simultaneous FECG-DUS dataset from a larger, more diverse cohort (e.g., 30 subjects spanning gestational weeks 20-40 and a broad fetal-heart-rate range), run the leave-one-subject-out pipeline, and check whether the FHR estimated from generated DUS stays within the 4.5 bpm Bland-Altman limit and the morphology metrics hold. If errors widen or a quality model begins flagging synthesized segments as non-good, the generalizability of the mapping is refuted.","tokens_in":11977,"feed_emoji":"🫀","tokens_out":5643,"duration_ms":48635,"temperature":0.7,"pith_summary":"This paper introduces Auto-FEDUS, an autoregressive generative model that maps a fetal electrocardiogram (FECG) heartbeat into the corresponding one-dimensional Doppler ultrasound (DUS) heartbeat. The goal is to overcome a bottleneck in fetal health monitoring: DUS datasets are scarce and imbalanced, while FECG data are comparatively abundant. Auto-FEDUS is built on dilated causal convolutions, works directly on raw waveforms, and extrapolates from input sampled at 250 Hz to output at 2 kHz. In leave-one-subject-out tests on 3,923 beats from five subjects, it produces signals that resemble real DUS in both time and frequency, and that a quality model labels as good while a heart-rate estimator agrees with the FECG-based reference to within 4.5 beats per minute.","feed_headline":"AI generates fetal Doppler ultrasound from ECG alone","feed_subtitle":"A WaveNet-style model turns fetal heart electrical signals into realistic ultrasound waveforms for fetal monitoring.","key_machinery":"The load-bearing mechanism is a WaveNet-style stack of five residual blocks of dilated causal convolutions, with dilation rates 1, 2, 4, 8, and 16, a kernel size of 20, and gated activations (tanh times sigmoid) feeding residual and skip connections. This setup expands the receptive field without adding layers, letting a single causal pass see both the immediate waveform shape and long-range temporal context. That combination is what the paper credits with capturing the short- and long-range dependencies needed to extrapolate a low-frequency electrical signal into a high-frequency mechanical waveform, rather than merely reproducing mean morphology.","core_discovery":"The central claim is that electrical FECG morphology carries enough information about mechanical cardiac activity that a well-designed autoregressive network can synthesize faithful DUS waveforms from FECG alone. On the paper's own terms, Auto-FEDUS is the first generative model for FECG-to-DUS cross-modal extrapolation, and it outperforms alternative generative architectures (autoencoder, LSTM GAN, DCGAN, WGAN-GP) on seven of eight time- and frequency-domain similarity metrics. The generated DUS segments are classified as good quality by a dedicated signal-quality assessment model, and the fetal heart rate estimated from them agrees with the real FECG-derived heart rate almost as well as the real DUS does: Bland-Altman limit 4.5 bpm, RMSE 2.3 vs 2.2 bpm, and PICP 96.7% vs 98.1%.","pith_inferences":["If the cross-modal mapping is as faithful as reported, FECG-only recordings from low-resource settings could be converted into synthetic DUS, effectively repurposing the abundant FECG modality as a source of paired training data.","The paper's own ablation suggests the model degrades beyond three consecutive beats, so segment-length synthesis in practice may need error correction or conditioning on longer context.","The five-subject dataset limits the plausible range of FECG-DUS couplings; testing on varied gestational ages and pathologies (e.g., growth restriction, congenital heart defects) would reveal whether the learned mapping is universal or cohort-specific."],"forward_implications":["Synthesized DUS signals can augment scarce, imbalanced DUS datasets for training machine-learning models that assess fetal health.","Auto-FEDUS generates clinically usable 3.75-second DUS segments by sequentially feeding heartbeat FECGs, and a quality model labels all such generated segments as good.","Because FHR estimates from generated DUS agree with FECG-derived references to within 4.5 bpm, the synthetic signals preserve the heart-rate information needed for screening.","The autoregressive approach handles mismatched sampling frequencies, so FECG recordings made with inexpensive hardware can be leveraged for DUS-based monitoring pipelines."],"supporting_citations":[{"why":"Supplies the WaveNet autoregressive dilated-causal-convolution architecture that Auto-FEDUS adapts for signal-to-signal mapping.","marker":"Oord et al. 2016"},{"why":"Provides the simultaneous FECG-DUS dataset and the preprocessing conventions (sampling rates, filters, lag correction) used throughout.","marker":"Valderrama et al. 2019"},{"why":"FECGSYN toolbox used to extract fetal ECG from seven-channel abdominal recordings and to auto-detect R-peaks.","marker":"Andreotti et al. 2016"},{"why":"Defines the WGAN-GP baseline and the adversarial loss the comparison GANs are built on.","marker":"Arjovsky, Chintala, and Bottou 2017"},{"why":"CardioGAN, the prior PPG-to-ECG adversarial framework whose time-frequency discriminator and Frechet distance calculation are referenced for evaluation.","marker":"Sarkar and Etemad 2021"},{"why":"The signal-quality model used to classify generated DUS segments as good quality.","marker":"Motie-Shirazi et al. 2023"},{"why":"The AutoFHR model used to estimate fetal heart rate from generated and real DUS for the Bland-Altman and PICP comparison.","marker":"Rafiei et al. 2024"}],"fun_headline_variants":["AI model synthesizes fetal Doppler ultrasound from ECG","Fetal ECG translated to realistic ultrasound by AI","Autoregressive net generates fetal ultrasound from ECG","From fetal ECG to Doppler ultrasound: AI does it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's core assumption is that the electrical morphology of a fetal heartbeat determines the Doppler ultrasound morphology closely enough that a mapping learned from 3,923 beats of five subjects will generalize to new subjects; if the FECG-to-DUS relationship is subject-specific or highly variable, the generated signals can look plausible without being clinically faithful.","fun_headline_variants_meta":{"raw":{"variants":["AI model synthesizes fetal Doppler ultrasound from ECG","Fetal ECG translated to realistic ultrasound by AI","Autoregressive net generates fetal ultrasound from ECG","From fetal ECG to Doppler ultrasound: AI does it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00012,"raw_usage":{"total_tokens":1107,"prompt_tokens":977,"completion_tokens":130,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":70}},"tokens_in":593,"tokens_out":130,"duration_ms":2165,"temperature":1.0,"reasoning_tokens":70,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:15:51.292566+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a new simultaneous FECG-DUS dataset from a larger, more diverse cohort (e.g., 30 subjects spanning gestational weeks 20-40 and a broad fetal-heart-rate range), run the leave-one-subject-out pipeline, and check whether the FHR estimated from generated DUS stays within the 4.5 bpm Bland-Altman limit and the morphology metrics hold. If errors widen or a quality model begins flagging synthesized segments as non-good, the generalizability of the mapping is refuted.","supporting_citations":[{"cited_title":"E.; Stroux, L.; Katebi, N.; Paljug, E.; Hall-Clifford, R.; Rohloff, P.; Marzbanrad, F.; and Clifford, G","cited_arxiv_id":null,"evidence_quote":"Provides the simultaneous FECG-DUS dataset and the preprocessing conventions (sampling rates, filters, lag correction) used throughout."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FECGSYN toolbox used to extract fetal ECG from seven-channel abdominal recordings and to auto-detect R-peaks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the WGAN-GP baseline and the adversarial loss the comparison GANs are built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CardioGAN, the prior PPG-to-ECG adversarial framework whose time-frequency discriminator and Frechet distance calculation are referenced for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The signal-quality model used to classify generated DUS segments as good quality."},{"cited_title":"D.; and Katebi, N","cited_arxiv_id":null,"evidence_quote":"The AutoFHR model used to estimate fetal heart rate from generated and real DUS for the Bland-Altman and PICP comparison."}],"review_version":1}