{"id":"4b026b52-f454-4c3c-9157-ea686eb8e25b","arxiv_id":"2505.01738","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A low-power earbud processor runs a compact neural net that detects heartbeats from single-ear dry electrodes, with heart-rate errors of 5.91 bpm single-ear and 0.49 bpm cross-ear.","lead":"The paper shows a single earbud-style device with dry electrodes can detect heartbeats and compute heart rate and heart rate variability in real time on the wearable itself. It reports good accuracy on four resting volunteers and very low energy use, which could make continuous heart monitoring in everyday earbuds more practical.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set threshold selection likely inflates the reported subject-independent F1 and HR/HRV errors; the correction algorithm may add further optimism.","rationale":"The paper has real strengths: a working on-device deployment, an order-of-magnitude model compression, public code and data, and explicit documentation of energy per inference. These support the engineering contributions. The reader's weakest assumption about Algorithm 1 is legitimate: the correction removes missed or duplicate beats using a running-median integer-multiple assumption, which can bias HRV toward regularity and make post-correction errors look smaller than the model's true accuracy. However, I see a more upstream and equally concrete problem: the detection threshold, a key hyperparameter, appears to be selected on the LOSO test subjects. Section III-G does not describe a held-out validation set; with only four subjects, 'across all subjects' in a LOSO protocol can only mean using the test folds' labels. This inflates F1 and therefore also improves the corrected HR/HRV errors. If the threshold were chosen on training folds, the headline numbers could shift materially. The reader did not list threshold tuning as the weakest assumption, though it appears in the rationale; hence partial agreement. My recommendation is unchanged: the paper should be accepted only after the evaluation protocol is fixed to avoid test-set hyperparameter selection and after the correction algorithm's assumptions are validated on data with genuine beat-to-beat variability. Neither issue requires rejecting the engineering work, but they are load-bearing for the accuracy claims.","tokens_in":10074,"tokens_out":5716,"duration_ms":63287,"concrete_test":"Re-run the LOSO evaluation selecting the detection threshold on the training folds only: for each LOSO fold, choose the threshold that maximizes F1 on the three training subjects' predictions, then apply Algorithm 1 exactly as in Section III-E and recompute F1, HR error, and HRV error for both single-ear and cross-ear configurations. Compare these numbers with Table II and the abstract. If the cross-ear HR error rises above roughly 1 bpm, the single-ear F1 falls below 0.75, or the cross-ear HRV error exceeds about 40 ms, then the reported accuracy depends on test-set threshold tuning and the subject-independent claim is not established by the current protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a subject-independent pipeline achieves F1 of 0.82/0.95 and HR/HRV mean errors of 0.49 bpm/25.82 ms, with only 36.7 uJ/inference. For this to hold, every hyperparameter, especially the detection threshold, must be chosen without using test labels, and the reported errors must reflect the deployed pipeline. Section III-G states that 'the threshold maximizing the F1-score was selected for rolling output inference across all subjects.' The dataset has only four subjects and evaluation is LOSO; no validation split is described. Selecting the threshold on all test subjects is equivalent to tuning a hyperparameter on the test set, which biases F1 and all downstream HR/HRV metrics upward. This directly undermines the 'subject-independent' claim. Additionally, Algorithm 1 (Section III-E) corrects missed or duplicate detections by dividing RR intervals by the rounded ratio to the running median. Because the headline HRV error of 25.82 ms is computed after this correction, the reported accuracy may reflect the correction's assumption that true RR intervals are near integer multiples of the running median, not the model's genuine beat-to-beat fidelity. On a four-subject rest-only cohort, no arrhythmic or motion-corrupted intervals are present to stress this assumption. The two issues compound: test-set tuning can make the raw detections look better than they are, and Algorithm 1 can then hide residual errors. The on-device deployment results (quantization, energy) are credible and well documented, but the evaluation protocol does not yet support the headline accuracy numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents DeepMF-mini, a lightweight convolutional autoencoder and R-peak classifier derived from the DeepMF architecture, integrated into BioGAP, an ultra-low-power wearable biosignal platform. The authors collect simultaneous arm-ECG and in-ear biopotential data from four healthy subjects in rest conditions and evaluate leave-one-subject-out (LOSO) R-peak detection, ECG reconstruction, and HR/HRV estimation for both single-ear and cross-ear configurations. They report a single-ear F1-score of 0.82, a cross-ear F1-score of 0.95, HR mean errors of 5.91 bpm (single-ear) and 0.49 bpm (cross-ear), HRV mean errors of 91.57 ms and 25.82 ms respectively, and an on-device inference energy of 36.7 uJ/inference with an estimated 36 h battery life. The main claims are that a subject-independent, fully wearable single-ear system can perform real-time HR/HRV monitoring at accuracy comparable to state-of-the-art cross-ear approaches.","tokens_in":10412,"tokens_out":2049,"duration_ms":21125,"significance":"If the reported accuracy holds under a properly unbiased evaluation, this is a useful step toward practical hearable-based cardiovascular monitoring: dry in-ear electrodes, a 10k-parameter model, on-device inference at 36.7 uJ per inference, and a concrete embedded deployment are all meaningful engineering contributions. The manuscript also ships code and data (github.com/pulp-bio/DeepMF-mini), which is a strength for reproducibility. The cross-ear HR error of 0.49 bpm and the sub-40 uJ inference energy are attractive headline numbers. However, the current evidence base is narrow: four male subjects, rest-only recordings, no confidence intervals, and an evaluation protocol that tunes the detection threshold on the test subjects. These issues directly affect the load-bearing 'subject-independent robust extraction' claim and need to be addressed before the quantitative claims can be taken at face value.","major_comments":[{"comment":"The detection threshold is selected on the evaluation set: 'the threshold maximizing the F1-score was selected for rolling output inference across all subjects.' Because the evaluation is LOSO with only four subjects and no separate validation split is described, this is equivalent to tuning a hyperparameter on the test data. This biases the reported F1-score and all downstream HR/HRV errors upward, undermining the subject-independent claim. Please use nested LOSO (or a subject-disjoint validation split) and report metrics for the threshold chosen without access to test labels.","section":"Section III-G"},{"comment":"All reported HRV errors (25.82 ms cross-ear, 91.57 ms single-ear) are computed after applying Algorithm 1, which inserts or removes beats by rounding the ratio of each RR interval to the running median. This correction assumes that missed or duplicate detections are isolated and that the underlying true RR intervals are close to integer multiples of a median interval. On a four-subject, rest-only dataset with no arrhythmic or motion-corrupted intervals, the correction is not stress-tested, and it could suppress genuine beat-to-beat variability if such intervals occur. Please report both raw and corrected HR/HRV errors and validate the correction on data containing artifacts or known ectopic beats.","section":"Section III-E, Algorithm 1"},{"comment":"The subject-independent claim rests on LOSO evaluation over only four healthy male subjects, all recorded at rest, with no confidence intervals or per-subject breakdowns. With n=4, a single subject can dominate the cross-ear HR error of 0.49 bpm, and the absence of variability measures makes it impossible to assess whether the reported accuracy is stable. Please provide per-subject results, confidence intervals, and preferably an independent test cohort or an explicit statistical statement of the precision of the reported errors.","section":"Section III-A and Section IV-C"}],"minor_comments":[{"comment":"The contributions list contains a typo: 'achieves SoA performance for HV and HVR' should presumably read 'HR and HRV'; elsewhere the paper correctly uses 'HRV'.","section":"Section I"},{"comment":"The reported DeepMF precision/recall values are inconsistent between the text (Section II: '95%/92%') and Table II ('91.2% / 94.9%'). Please reconcile these numbers.","section":"Table II"},{"comment":"The relationship between the reported 36.7 uJ per inference, the 400 ms update interval, and the total system power of 7.6 mW is not derived; please state which components contribute to the 7.6 mW and how the 36 h battery estimate is obtained.","section":"Section IV-D and Table I"},{"comment":"The dataset description reports mean age as '31.0 ± 5.1 years' with a decimal point; '31 ± 5 years' would be more natural, and the number of female participants (zero) should be stated explicitly as a limitation.","section":"Section III-A"}],"recommendation":"major_revision","confidential_remarks":"The paper's central feasibility claim is plausible and the engineering is solid, but the evaluation protocol has a test-set threshold selection issue that is load-bearing for the headline numbers. The four-subject, rest-only dataset is thin for a 'subject-independent robust extraction' claim, and the HRV correction algorithm could hide detection errors. These are fixable within the manuscript's scope: re-evaluate with a validation-based threshold, report raw and corrected errors, add per-subject and CI statistics, and temper the generalization claims. I do not see circularity in the model derivation, but the evaluation bias must be addressed before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a credible embedded-systems paper with a real engineering contribution—real-time single-ear R-peak detection and HR/HRV on a low-power wearable—but the abstract accuracy numbers are more optimistic than the evaluation protocol supports. The core deployment story is new and worth publishing; the evaluation needs to be redone before the numbers become citable.\n\nWhat the paper does well: DeepMF-mini reduces the model from 759k to under 10k parameters, a 75x compression that enables on-device inference. Running on BioGAP with 8-bit quantization and 36.7 uJ/inference is a genuine step toward earbud-form-factor cardiac monitoring. The rolling-window postprocessing is sensible. The authors release code and data, and they compare against the original DeepMF on the same recordings. That is reproducible engineering.\n\nThe soft spots are in the evaluation. Section III-G chooses the F1-maximizing threshold 'across all subjects' in LOSO. That is test-set tuning. With four subjects and no validation split, the reported F1 and the downstream HR/HRV errors are biased upward. The subject-independent claim needs threshold selection on training folds only. Second, Algorithm 1 corrects missed or duplicate peaks by rounding RR intervals to integer multiples of the running median. On a rest-only dataset that assumption is convenient, but it can erase genuine beat-to-beat variability in the presence of arrhythmia or motion. The headline HRV error of 25.82 ms is computed after this correction, so it reflects the correction rule as much as the model. The dataset is also small (four young male subjects, rest only, no error bars), and there is a minor internal inconsistency: the intro says 18 h battery life while the abstract and conclusion say 36 h.\n\nWho this is for: researchers working on embedded biosignal processing and hearables. The feasibility demonstration is worth knowing, even if the accuracy numbers need more validation. I would cite it as a feasibility study, not as a benchmark.\n\nRecommendation: send to peer review. The engineering and code/data release make it worth serious referee time, and the review process can fix the threshold-selection protocol, require raw vs corrected errors, and ask for more subjects or at least confidence intervals. With those changes it becomes a solid contribution.","headline":"A credible embedded-systems step toward single-ear HR/HRV monitoring, but the headline accuracy numbers are inflated by test-set threshold tuning and a median-based correction that hides real HRV; the engineering deserves a referee, the evaluation needs rework.","tokens_in":10896,"tokens_out":3350,"would_cite":true,"duration_ms":32286,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single-ear, dry-electrode wearable can reconstruct an ECG, detect R-peaks, and estimate HR/HRV in real time on an ultra-low-power chip, using only 36.7 µJ per inference.","keywords":["ear-ECG","R-peak detection","heart rate variability","wearable ECG","on-device inference","dry electrodes","DeepMF","low-power embedded"],"falsifier":"Record ear-ECG and simultaneous arm-ECG from, say, twenty subjects during rest, walking, and a protocol that induces premature ventricular contractions; compute RMSSD from the corrected peaks and compare with the arm-ECG ground truth. If the correction algorithm removes or inserts beats outside isolated missed/duplicate cases, the single-ear HRV error will climb substantially above the reported 91.57 ms as the arrhythmia or motion burden increases.","tokens_in":9913,"feed_emoji":"🎧","tokens_out":8591,"duration_ms":72804,"temperature":0.7,"pith_summary":"This paper claims that a fully wearable, single-ear device with dry in-ear electrodes can reconstruct an ECG-like signal, detect R-peaks, and compute heart rate (HR) and heart-rate variability (HRV) in real time on an ultra-low-power processor, without any per-user calibration. That matters because it removes two obstacles to continuous at-home cardiovascular monitoring: wet electrodes and wiring between the ears. In the paper's tests, the on-device model used 36.7 µJ per inference and achieved a mean HR error of 0.49 bpm and a mean HRV error of 25.82 ms in the cross-ear configuration, with a single-ear F1-score of 0.82 for R-peak detection.","feed_headline":"A dry earbud does real-time heart-rate and HRV tracking on-device","feed_subtitle":"A 10k-parameter model detects R-peaks at 0.95 F1 cross-ear and 0.82 single-ear, on a 36-hour battery.","key_machinery":"The load-bearing mechanism is the DeepMF-mini network plus its postprocessing pipeline. DeepMF-mini is a 1D convolutional autoencoder (roughly 10k parameters) whose encoder is frozen after training and paired with a classifier made of transpose convolutions; it learns a latent-space representation shared between the noisy ear biopotential and the clean arm-ECG, and the classifier outputs R-peak probabilities. The rolling-window refinement evaluates each 2-second window multiple times with a 0.4 s shift, discards the least-reliable last segment, and takes element-wise maxima over the overlap, so peaks near window edges get a second chance. Finally, Algorithm 1 corrects the predicted peak sequence by adjusting to the running median of RR intervals, which is what makes the reported HRV numbers possible from noisy single-ear signals.","core_discovery":"The paper's central claim is that a compact convolutional network called DeepMF-mini, a roughly 75× smaller variant of the DeepMF architecture, can be trained once on ear biopotentials and then detect R-peaks in real time for new users without recalibration. The model is a two-stage design: an encoder-decoder first maps the ear signal to a latent representation that reconstructs the arm-measured Lead I ECG; the decoder is then replaced by a peak classifier trained on the same latent space. A rolling-window postprocessor merges overlapping predictions to catch peaks near window boundaries, and a correction algorithm that uses the running median of RR intervals removes duplicates and inserts missed beats. With all processing on the wearable, the paper reports an R-peak F1-score of 0.95 for cross-ear and 0.82 for single-ear input, HR errors of 0.49 bpm and 5.91 bpm, and HRV errors of 25.82 ms and 91.57 ms, respectively, at 36.7 µJ per inference.","pith_inferences":["The paper's HRV numbers depend on Algorithm 1, which assumes missed or duplicate beats are isolated and true intervals are near whole-number multiples of the running median; if real arrhythmias or motion artifacts produce genuine short-long-short patterns, this correction would bias HRV toward regularity, so the 25.82 ms error may not hold on ambulatory or arrhythmic data.","The single-ear input's main weakness is the attenuated, less consistent T-wave and post-QRS signal; adding a second sensing modality already present in earbuds (e.g., PPG or an accelerometer for motion rejection) could plausibly close most of the 91.57 ms HRV gap without a major energy cost.","The same latent-space pipeline could be retargeted to estimate other cardiac intervals, such as QT or PR, if the ECG reconstruction quality improves; the paper already shows the reconstruction approximates the arm ECG waveform.","The 36.7 µJ per inference figure suggests an always-on heartbeat detector leaves ample power headroom for audio streaming on the same earbud, which is the practical path to making cardiac monitoring a background feature."],"forward_implications":["Consumer earbuds using this design could track HR and HRV continuously for about 36 hours on a 75 mAh battery, since the whole system draws about 7.6 mW.","Cross-ear R-peak detection (F1 = 0.95) is achieved with 75× fewer parameters than the reference DeepMF model, freeing memory and compute for audio processing on the same chip.","The single-ear configuration reduces hardware complexity (no wiring across the head) and still reaches HR errors around 5.91 bpm, suitable for trend monitoring rather than clinical-grade beat annotation.","The quantized 10 kB model runs in 1.47 ms per inference at the energy-efficient operating point, so it occupies only a small fraction of the duty cycle.","Because the model is trained with leave-one-subject-out, a new user would not need a personal calibration session before the device starts reporting HR/HRV."],"supporting_citations":[{"why":"Supplies the DeepMF architecture that DeepMF-mini compresses; the encoder–decoder and latent-space classifier design are inherited from this work.","marker":"[8]"},{"why":"Provides the dry in-ear electrode technology used for the wearable measurements; its soft rubber electrodes reduce impedance without gel.","marker":"[4]"},{"why":"Introduces the BioGAP platform that acquires signals and runs the model; its ultra-low-power GAP9 processor enables on-device inference.","marker":"[14]"},{"why":"The only previous single-ear feasibility study; the paper builds on it by adding real-time R-peak detection and HR/HRV.","marker":"[10]"},{"why":"Supplies the prior state-of-the-art HR error (4.52 bpm) that this work's 0.49 bpm is compared against.","marker":"[13]"},{"why":"Prior ear-ECG HRV study with lower R-peak accuracy (68% precision); the paper's 0.95 F1 is compared with this baseline.","marker":"[12]"}],"fun_headline_variants":["Single-ear earbud reconstructs ECG and tracks HR/HRV in real time","On-device ECG: 36-hour battery, 36.7 µJ per inference","A dry earbud does real-time HR/HRV with R-peak detection","Subject-independent ECG from a single ear, no wet electrodes","Earbud ECG: 0.49 bpm HR error, 25.8 ms HRV error, on-device"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The R-peak correction step assumes that missed or duplicate detections are isolated events, so that deleting or inserting a beat according to the running median of RR intervals does not erase genuine beat-to-beat variability; if real arrhythmic or motion-induced intervals occur, the correction would bias the HRV measurement toward regularity and the reported accuracy would be optimistic.","fun_headline_variants_meta":{"raw":{"variants":["Single-ear earbud reconstructs ECG and tracks HR/HRV in real time","On-device ECG: 36-hour battery, 36.7 µJ per inference","A dry earbud does real-time HR/HRV with R-peak detection","Subject-independent ECG from a single ear, no wet electrodes","Earbud ECG: 0.49 bpm HR error, 25.8 ms HRV error, on-device"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000285,"raw_usage":{"total_tokens":1749,"prompt_tokens":1084,"completion_tokens":665,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":700,"completion_tokens_details":{"reasoning_tokens":550}},"tokens_in":700,"tokens_out":665,"duration_ms":6910,"temperature":1.0,"reasoning_tokens":550,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:11:41.228652+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record ear-ECG and simultaneous arm-ECG from, say, twenty subjects during rest, walking, and a protocol that induces premature ventricular contractions; compute RMSSD from the corrected peaks and compare with the arm-ECG ground truth. If the correction algorithm removes or inserts beats outside isolated missed/duplicate cases, the single-ear HRV error will climb substantially above the reported 91.57 ms as the arrhythmia or motion burden increases.","supporting_citations":[{"cited_title":"The deep-match framework: R-peak detection in ear-ecg,","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepMF architecture that DeepMF-mini compresses; the encoder–decoder and latent-space classifier design are inherited from this work."},{"cited_title":"SoftPulse Products: Innovative Solutions for Wearable Sensors,","cited_arxiv_id":null,"evidence_quote":"Provides the dry in-ear electrode technology used for the wearable measurements; its soft rubber electrodes reduce impedance without gel."},{"cited_title":"Biogap: a 10-core fp-capable ultra-low power iot processor, with medical-grade afe and ble connectivity for wearable biosignal processing,","cited_arxiv_id":null,"evidence_quote":"Introduces the BioGAP platform that acquires signals and runs the model; its ultra-low-power GAP9 processor enables on-device inference."},{"cited_title":"Hearables: feasibility of recording cardiac rhythms from single in-ear locations,","cited_arxiv_id":null,"evidence_quote":"The only previous single-ear feasibility study; the paper builds on it by adding real-time R-peak detection and HR/HRV."},{"cited_title":"In-ear ECG Signal Enhancement with Denoising Convolutional Autoencoders","cited_arxiv_id":"2409.05891","evidence_quote":"Supplies the prior state-of-the-art HR error (4.52 bpm) that this work's 0.49 bpm is compared against."},{"cited_title":"Hearables: Heart rate variability from ear electrocar- diogram and ear photoplethysmogram (ear-ecg and ear-ppg),","cited_arxiv_id":null,"evidence_quote":"Prior ear-ECG HRV study with lower R-peak accuracy (68% precision); the paper's 0.95 F1 is compared with this baseline."}],"review_version":1}