{"id":"7a4508c8-69ba-4ce8-b9c1-8fd22c366c22","arxiv_id":"2510.23819","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A dual ECG-PCG wearable with a burst-adaptive NLMS filter reports 30.32 dB (ECG) and 37.01 dB (PCG) SNR improvements in a 14-second real hospital recording.","lead":"This paper describes a portable device that records ECG and heart-sound (PCG) signals together and uses a modified adaptive filter to remove hospital noise in real time. It reports large signal-quality gains (30-37 dB) on a single hospital recording, suggesting a cheaper path to cardiac screening in noisy clinics.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 30–37 dB real-world SNR gains probably come from the fixed band-pass filters, not the adaptive canceller; the hand-defined SNR metric (Eq. 6) rewards band-limiting and no ablation is reported.","rationale":"The reader's verdict is CONDITIONAL, and my concern reinforces that conditionality without moving it. The reader identified the hand-defined SNR as partly rewarding band-limiting, but their 'weakest assumption' was the reference-signal correlation. I argue the SNR metric is actually more load-bearing: even if the ambient microphone reference were perfectly correlated, the reported 30–37 dB improvements would still be invalid as evidence because they are computed with a non-standard metric that a simple fixed filter can inflate. The paper's own description of the ECG pipeline shows no adaptive component, which makes the ECG claim particularly vulnerable to this critique. The article does provide honest simulation results under controlled conditions, which support the algorithm's basic functionality, so a full rejection is not warranted. However, the real-world headline numbers need either a proper ablation or a standard SNR evaluation before they can be accepted. Since the reader already conditioned acceptance on similar evidence, my recommendation is UNCHANGED (CONDITIONAL). The agreement is partial because we both flag the SNR metric, but we differ on the single weakest assumption.","tokens_in":7415,"tokens_out":4688,"duration_ms":49219,"concrete_test":"Re-record or reuse the 14-s hospital signals and compute ΔSNR using Eq. 6 for three conditions: (i) raw signal, (ii) after only the fixed 20–200 Hz band-pass (or the ECG elliptic filter cascade), and (iii) after the full pipeline including BA-NLMS. If ΔSNR(ii) is within a few dB of ΔSNR(iii), the adaptive filter adds little. Additionally, report a standard SNR (e.g., using a clean reference PCG recorded in a quiet room with the same sensor, time-synchronized, or using the PhysioNet 2016 challenge noise-augmentation protocol) to verify the 30–37 dB figures.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest link in the central claim is the real-world SNR evaluation, not the reference-signal assumption. The paper reports ΔSNR = 37.01 dB (PCG) and 30.32 dB (ECG) on a single 14-s recording, computed via Eq. 6: SNR = 10 log10( P_LF / (P_LFN + P_HF) ). The passbands are never specified, and the PCG pipeline ends with a fixed 20–200 Hz band-pass. A fixed filter that removes all energy outside the 'signal band' will by definition increase Eq. 6's SNR, potentially by tens of dB if the ambient noise is broadband. The paper provides no ablation isolating the adaptive BA-NLMS contribution from this fixed filtering. The ECG improvement is even more clearly a fixed-filter artifact: Section II-C describes only a cascade of elliptic filters (0.5 Hz HP, 150 Hz LP, 50 Hz notch), no adaptive component at all, yet the abstract attributes the 30.32 dB to the 'proposed system.' Without a baseline (e.g., ΔSNR from the fixed filters alone) or a standard SNR metric anchored to a known clean reference, the headline numbers are not evidence that the adaptive noise canceller works in real hospitals.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a portable simultaneous ECG-PCG acquisition device with a real-time adaptive noise cancellation pipeline. The PCG channel uses a proposed burst-adaptive NLMS (BA-NLMS) algorithm that temporarily increases the step size when the reference-signal energy exceeds a moving average; the ECG channel uses a cascade of fixed elliptic filters (0.5 Hz high-pass, 150 Hz low-pass, 49.5–50.5 Hz bandstop). The BA-NLMS is evaluated on 3934 simulated noisy PCG recordings (per class) built from the AUTHHSDB dataset and hospital noise, comparing against LMS and NLMS. The complete prototype is then evaluated on a single 14-second hospital recording, reporting ΔSNR of 37.01 dB for PCG and 30.32 dB for ECG using a hand-defined SNR metric (Eq. 6). The abstract additionally states that a complexity analysis confirms the pipeline's suitability for embedded implementation, but no such analysis appears in the manuscript.","tokens_in":7760,"tokens_out":6199,"duration_ms":67366,"significance":"If the real-world SNR improvements were fully supported, this device would be a meaningful low-cost cardiac screening tool for noisy, resource-constrained settings. The simulation study is a genuine strength: it uses a public dataset, a large number of recordings, several power-independent metrics (NMSE, ΔSNR, CC, NMAE), and direct comparisons against LMS and NLMS. The BA-NLMS shows consistent but modest gains over NLMS in simulation (about 0.5 dB in the full noisy set and about 1.9 dB on burst-only recordings). However, the headline real-world numbers are not currently supported: the frequency bands in the SNR metric are never specified, there is no ablation isolating the adaptive contribution from fixed band-pass filtering, the ECG improvement comes entirely from fixed filters rather than adaptive cancellation, and the validation rests on a single 14-second trace. The missing complexity analysis also undercuts the embedded-suitability claim. With a proper ablation, more recordings, and a defined evaluation protocol, this work could become a useful contribution.","major_comments":[{"comment":"The real-world validation rests on a single 14-s recording and an undefined SNR metric. Eq. (6) defines SNR = 10 log10( P_LF / (P_LFN + P_HF) ), but the frequency ranges LF, LFN, and HF are never specified. Since the PCG pipeline ends with a 20–200 Hz band-pass filter (Section III-B), a fixed filter that removes energy outside the assumed 'signal band' will increase this ratio by construction. The reported 37.01 dB improvement may therefore largely reflect band-limiting rather than the BA-NLMS algorithm. Please specify the frequency bands, report the ΔSNR achieved by the fixed filters alone, and evaluate on more than one recording. Ideally, use a metric anchored to a known clean reference or a blind quality measure.","section":"Section III-B, Eq. (6)"},{"comment":"The ECG improvement of 30.32 dB is not attributable to the proposed adaptive noise-cancellation pipeline. Section II-C describes only a cascade of fixed elliptic IIR filters (0.5 Hz high-pass, 150 Hz low-pass, 49.5–50.5 Hz bandstop); there is no adaptive component in the ECG path. The abstract and conclusion attribute this improvement to the proposed system's noise-cancellation capabilities. Please provide an ablation showing the contribution of each stage, or clearly state that the ECG ΔSNR is a fixed-filter result and not the output of the adaptive algorithm.","section":"Section II-C"},{"comment":"The abstract states that a complexity analysis confirms the pipeline's suitability for embedded implementation, but no such analysis appears in Sections II–IV. In fact, Section IV lists 'implementing the denoising algorithms on the microcontroller for fully embedded operation' as future work, and Section II-A says data are streamed to a computer for real-time ANC. Thus the claims of an integrated, embedded, real-time system are not currently supported. Either add the complexity analysis and clarify where processing is performed, or temper the claims accordingly.","section":"Abstract / Section IV"},{"comment":"The central ANC assumption is unverified on the prototype. Section II-B states that the reference signal r(n) is uncorrelated with heart sounds but closely correlated with the noise reaching the stethoscope microphone. No measurement of that correlation is presented. In the synthetic evaluation, the reference is generated from the same noise source through a moving-average filter (Section III-A), so it is highly correlated by construction and does not test the real-world assumption. Please report the measured coherence/correlation between the ambient microphone array and the noise component in the stethoscope channel, and evaluate the algorithm under a decorrelated reference.","section":"Section II-B"},{"comment":"The BA-NLMS parameters (μ0 = 0.05, α = 0.99, η = 5, β = 6, ε = 0.001, filter length 10) are stated without a fitting procedure or sensitivity analysis. The burst-detection mechanism depends critically on η and β, and no evidence is provided that these values generalize beyond the selected AUTHHSDB subset. Please include a sensitivity study or justify the values from a principled criterion; otherwise the robustness of the simulated and real-world gains is unclear.","section":"Section III-A"}],"minor_comments":[{"comment":"The sentence 'This process generated 3934 noisy recordings, each for normal and abnormal' is ambiguous. The numbers suggest 7 clean normal × 562 noise = 3934 per class; please state explicitly whether the total is 3934 or 7868.","section":"Section III-A"},{"comment":"The input SNR values reported for the real-world recordings (e.g., -9.08 dB for ECG and -2.08 dB for PCG) are not defined. State whether they use the same Eq. (6) and, if so, give the exact frequency bands used.","section":"Equation (6), Fig. 4"},{"comment":"Notation is inconsistent: Eq. (2) uses ||r(n)||² while Eq. (5) uses r^T(n)·r(n) for the same quantity. Unify the notation.","section":"Section II-B, Eqs. (2) and (5)"},{"comment":"The abstract in the submission says 'the proposed system achieved signal-to-noise ratio improvements,' whereas the abstract in the full text says 'the proposed algorithms achieved.' Please use consistent wording, especially because the ECG path is not adaptive.","section":"Abstract"},{"comment":"The time-domain plots and spectrograms lack axis labels/units (time in seconds, amplitude in what units) and the spectrogram color scale is not described. Please add these details.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The manuscript describes a plausible hardware design and a reasonably thorough simulation study, but the real-world validation is not yet adequate for the headline claims. The absence of a complexity analysis, despite being promised in the abstract, is a notable omission. I would be willing to review a revised version that includes a proper ablation, specified evaluation bands, and more real-world data."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a mixed bag. The device is real, the simulated evaluation is decent, and the burst-adaptive NLMS does improve over plain NLMS on burst noise. But the headline claim — 30.32 dB and 37.01 dB SNR improvement in real hospital recordings — does not survive scrutiny. Read it for the hardware and the simulation, not for the field validation.\n\nWhat is actually new: a portable simultaneous ECG-PCG acquisition prototype, streamed via BLE, with a real-time ANC pipeline. The BA-NLMS variant is a threshold on instantaneous versus average reference energy, switching the step size. That is a minor twist on existing variable-step-size NLMS — the paper doesn't cite that literature — but the end-to-end integration is a legitimate engineering contribution. The simulated comparison on AUTHHSDB (3934 recordings) is thorough and honestly reported: BA-NLMS beats NLMS, especially on burst-type noise.\n\nWhere it falls apart: the real-world validation rests on a single 14-second recording and a hand-defined SNR metric (Eq. 6) with unspecified frequency bands. The PCG pipeline ends with a fixed 20–200 Hz band-pass, and the ECG pipeline is just fixed elliptic filters — 0.5 Hz HP, 150 Hz LP, 50 Hz notch — with no adaptive component at all. Yet the abstract credits the \"proposed system\" with the ECG 30.32 dB improvement. That number is almost certainly a fixed-filter artifact. There is no ablation separating what BA-NLMS contributes versus the band-pass. For the PCG, the simulated results suggest BA-NLMS alone might give ~20 dB under ideal reference conditions, so the real-world 37 dB likely includes a large fixed-filter component.\n\nThe reference-signal assumption is also shaky: the ambient mics sit on the device body, but no measurement shows their correlation with the noise reaching the stethoscope mic. The simulation uses a moving-average-filtered version of the same noise as the reference, which is far more favorable than reality.\n\nAnd a smaller overclaim: the abstract says \"real-time adaptive noise cancellation pipeline integrated into a device,\" but the ANC runs on a computer. The conclusion lists \"implementing the denoising algorithms on the microcontroller\" as future work.\n\nBottom line: the simulation is solid, the hardware is plausible, the field validation is not. This deserves peer review, but the authors need to add multi-subject hospital data, a standard SNR definition, an ablation of fixed versus adaptive filtering, and a measurement of the reference correlation. Also they should stop attributing the ECG gain to the adaptive algorithm.\n\nThe reader's stress-test is right. I'd bring this to a reading group as a case study in evaluation overclaim, but I wouldn't cite it yet.\n\nRecommendation: send to peer review, expect heavy revision.","headline":"Solid prototype and honest simulation, but the headline real-world SNR numbers are likely a fixed-filter artifact and the ANC contribution is unproven in the field.","tokens_in":8275,"tokens_out":3855,"would_cite":false,"duration_ms":37407,"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":"A portable ECG–PCG device with burst-adaptive noise cancellation lifts real-world hospital signal-to-noise ratios by 30.32 dB (ECG) and 37.01 dB (PCG).","keywords":["adaptive noise cancellation","phonocardiogram","electrocardiogram","burst noise","normalized least mean square","real-time denoising","portable cardiac screening","hospital noise"],"falsifier":"Compute the coherence between the ambient reference and the noise component in the stethoscope microphone during hospital recordings; if it is low across 20–200 Hz, the reported 37 dB gain would not generalize to other hardware placements.","tokens_in":7349,"feed_emoji":"🩺","tokens_out":4241,"duration_ms":43833,"temperature":0.7,"pith_summary":"This paper argues that a low-cost, portable device can simultaneously record ECG and heart-sound (PCG) signals and clean them in real time even in loud hospital wards. The key move is a burst-adaptive normalized least-mean-square filter that speeds up adaptation whenever the ambient-noise reference jumps in energy, addressing the non-stationary bursts that defeat standard adaptive filters. On real hospital recordings the authors report SNR improvements of 30.32 dB for ECG and 37.01 dB for PCG. If correct, the system makes reliable cardiac screening feasible in resource-constrained settings without expensive commercial digital stethoscopes.","feed_headline":"Real-time filter lifts hospital heart sounds by 37 dB","feed_subtitle":"On a portable device that records ECG and sound together, the algorithm also raises ECG signal quality by 30 dB.","key_machinery":"The central mechanism is the burst-adaptive NLMS update: an energy tracker maintains a moving average of reference power; whenever instantaneous reference power exceeds a threshold times that average, the normalized step size is multiplied by a scaling factor, so the filter reacts quickly to noise bursts, then resumes small-step operation. The reference comes from a quad-microphone array facing outward on the stethoscope body, intended to capture ambient noise uncorrelated with heart sounds.","core_discovery":"The authors set out to show that a single portable device can acquire ECG and PCG simultaneously and clean both in real time with modest compute. The specific discovery is that a normalized LMS filter, modified to boost its step size during detected energy bursts, adapts to non-stationary hospital noise faster than standard LMS/NLMS while staying stable in quiet intervals. On hospital recordings made with the prototype, the PCG pipeline improved SNR by 37.01 dB and the ECG pipeline by 30.32 dB; on a synthetic noisy PCG dataset, the burst-adaptive filter beat both LMS and NLMS on all four metrics for normal and abnormal recordings.","pith_inferences":["The paper does not report the coherence between the ambient reference and the noise reaching the chest microphone; measuring that coherence would show whether the 37 dB gain transfers to other stethoscope geometries and body positions.","The ECG gain comes from fixed elliptic filters rather than adaptive cancellation, so adding an adaptive ECG stage or motion-artifact rejection is a natural extension for ambulatory use.","A testable consequence of the burst-adaptive idea is that the improvement over standard NLMS should grow with the proportion of time spent in burst events; a dataset with controlled burst density could verify that dose–response.","Because the device uses dry electrodes and outward-facing mics, performance during physical motion is an open question; a movement-artifact protocol would map that boundary."],"forward_implications":["Real-time cleaning of both modalities can move onto an embedded microcontroller, so synchronized ECG–PCG screening no longer depends on offline processing or bulky equipment.","A 37.01 dB PCG gain means murmur features that a human or algorithm needs can survive in a busy ward, not just in a quiet exam room.","A 30.32 dB ECG gain preserves the timing landmarks needed to place S1 and S2 within the cardiac cycle when the two signals are read together.","On burst-heavy noise, the kind produced by alarms, trolleys, and conversations, the burst-adaptive filter outperforms conventional LMS and NLMS, indicating the approach targets the noise profile that actually limits auscultation."],"fun_headline_variants":["Burst-adaptive filter lifts heart sounds by 37 dB in real time","Portable ECG-PCG device cleans heart signals by 37 dB live","Real-time burst-adaptive noise cancel boosts heart sound SNR 37 dB","Hospital noise tamed: portable device lifts heart sounds 37 dB","Burst-adaptive noise cancellation cleans ECG-PCG in real time by 30-37 dB"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The result rests on the assumption that the outward-facing ambient microphones provide a noise reference closely correlated with the noise at the stethoscope and free of heart sounds; the paper does not measure that correlation.","fun_headline_variants_meta":{"raw":{"variants":["Burst-adaptive filter lifts heart sounds by 37 dB in real time","Portable ECG-PCG device cleans heart signals by 37 dB live","Real-time burst-adaptive noise cancel boosts heart sound SNR 37 dB","Hospital noise tamed: portable device lifts heart sounds 37 dB","Burst-adaptive noise cancellation cleans ECG-PCG in real time by 30-37 dB"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000634,"raw_usage":{"total_tokens":2769,"prompt_tokens":758,"completion_tokens":2011,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":1924}},"tokens_in":502,"tokens_out":2011,"duration_ms":13919,"temperature":1.0,"reasoning_tokens":1924,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T07:49:34.916371+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the coherence between the ambient reference and the noise component in the stethoscope microphone during hospital recordings; if it is low across 20–200 Hz, the reported 37 dB gain would not generalize to other hardware placements.","supporting_citations":[],"review_version":1}