{"id":"8e040cab-4760-42e5-bc0c-512c4441211f","arxiv_id":"2506.19427","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A particle filter combined with a two-pathway AV node model enables beat-to-beat estimation of AV nodal refractory period and conduction delay from surface ECG during AF, with best accuracy for the slow-pathway refractory period.","lead":"Researchers built a computer model plus a particle filter to estimate the atrioventricular node's refractory period and conduction delay from ECG, beat by beat, during atrial fibrillation. The method tracks expected changes during a tilt test, but its accuracy was mainly tested using data generated by the same model it uses for estimation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tilt-test validation is confounded by atrial-rate changes: the ECG-PF synthesizes its atrial input from f-waves, and HUT alters f-wave frequency, so the reported AV-node changes may be an artifact of changed atrial input rather than true AV nodal modulation.","rationale":"The reader's weakest assumption identified model fidelity and the circular simulation validation. My concern is closely related but more specific: even granting the model, the only external evidence (tilt test) is confounded by concurrent changes in the atrial input that the estimator itself generates from f-waves. This does not overturn the paper's value as a methods contribution with transparent limitations and publicly available code, and the reader's CONDITIONAL verdict already captures the need for external validation. I therefore recommend no change to the verdict, but the condition should explicitly require addressing the atrial-rate confound, for example by the simulation test above. The numbers make the concern concrete: the tilt effect sizes (4-22 ms) are smaller than or comparable to the one-minute averaged estimation errors (24-152 ms), so the group-level significance is not by itself evidence that the estimates track true AV nodal modulation. I credit the authors for acknowledging the limitation in Section IV and for making the model code available, which makes the proposed test feasible. No ad hominem is intended; the critique is on the inference from the tilt data.","tokens_in":16983,"tokens_out":7926,"duration_ms":93706,"concrete_test":"Simulate the tilt protocol using the authors' code with true AV nodal properties held constant: generate 21 supine and 21 HUT ECG-like inputs from the AV node model, with f-wave frequency increased during HUT by the amount observed in the tilt study (e.g., from [24]), and otherwise identical true phi trajectories. Run the ECG-PF and smoothing exactly as in Section II-G, compute phase differences, and apply the same one-sided Wilcoxon tests. If a significant decrease in R_SP or D_FP appears under constant true phi, the tilt result is an artifact of atrial-rate-driven AA generation; if no false decrease appears, the confounder is unlikely to explain the reported result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that ECG-based estimates 'agree with expected AV nodal modulation' rests on the tilt-test result (Section III-C). This is the only external evidence; the simulation study is self-consistency because ground truth is generated from the same model (Section IV.A). The tilt result is vulnerable to a specific confound: the ECG-PF does not measure atrial activations; it synthesizes AA series from f-wave characteristics (Section II-B2). Head-up tilt changes autonomic state and the atrial fibrillatory rate, as the authors themselves note for HDT (Discussion, citing [24]). A systematic change in the generated AA series between supine and HUT can therefore shift the estimated AV nodal properties even if the true AV node is unchanged. The magnitude of the reported effects sharpens this concern. Table III shows mean HUT-vs-supine changes of -4.7 ms (R_FP), -21.6 ms (R_SP), -4.1 ms (D_FP), and +2.0 ms (D_SP), while the ECG-PF one-minute averaged errors (Table I, ECG l1_min) are 92.3 ms, 23.6 ms, 89.6 ms, and 152 ms, respectively. The tilt signal is at or below the estimator's noise floor, so the significant Wilcoxon results could arise from a systematic bias induced by the changed f-wave input rather than from true AV nodal modulation. The authors acknowledge that RR interval changes during tilt cannot be attributed solely to AV node properties (Discussion), yet still present the tilt result as supporting the physiological relevance of the estimates. Without controlling for the atrial-rate confound, the external validation is too weak to support the abstract's claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a framework for beat-to-beat estimation of four AV-nodal conduction properties (refractory periods R_FP and R_SP, conduction delays D_FP and D_SP) during atrial fibrillation, combining a two-pathway network model of the AV node (Eqs. 1-3) with a particle filter and forward-filtering backward-sampling smoother. Two variants are presented: EGM-PF, which uses intracardiac EGM-derived atrial activations, and ECG-PF, which synthesizes atrial activation series from f-wave characteristics. The method is evaluated in three steps: against simulated data generated from the same AV-node model, on five patients with synchronized EGM/ECG from the iafdb, and on 21 AF patients undergoing a tilt protocol. The paper reports large mean absolute errors in the simulation study and a statistically significant decrease in R_FP, R_SP, and D_FP from supine to head-up tilt, concluding that beat-to-beat estimation is feasible and agrees with expected autonomic modulation.","tokens_in":17329,"tokens_out":4671,"duration_ms":53714,"significance":"If the estimated quantities corresponded to true physiological AV-nodal properties, the work would be a useful advance: a non-invasive, beat-to-beat readout of AV-nodal modulation would enable new studies of autonomic control during AF and could inform individualized rate-control therapy. The paper's strengths include publicly available model code, a clinically relevant tilt-test dataset, transparent reporting of computational cost, and an honest limitations paragraph. However, the current evidence does not yet establish the physiological validity of the estimates: the simulation ground truth is generated from the same model that the estimator inverts, and the only external validation signal is smaller than the estimator's reported noise floor and is confounded by changes in atrial input. The significance of the contribution therefore depends on whether the manuscript is revised to either supply independent validation or substantially temper its claims.","major_comments":[{"comment":"The simulation-based accuracy evaluation is circular with respect to physiological validity. The ground truth phi*(k) is generated by running the same network model (Eqs. 1-3) with the same assumptions used in inference, so the reported l1 errors measure the estimator's ability to invert a known model under synthetic noise rather than its accuracy against true human AV-nodal behavior. The authors acknowledge this in Section IV-A ('only been validated using ground truth data generated from the same AV node model'), but the abstract and conclusion nonetheless present these error values as evidence that ECG-based beat-to-beat estimation is feasible. This is load-bearing: the central claim needs either an external reference (e.g., invasive measurements during a protocol with induced autonomic changes) or at least a model-misspecification sensitivity analysis showing that the estimator remains accurate when the simulated AV node differs from the assumed network structure.","section":"Section II-E and IV-A"},{"comment":"The tilt-test result is confounded by atrial-input changes and lies below the estimator's reported noise floor. The ECG-PF does not measure atrial activations; it generates AA series from f-wave characteristics (Section II-B2), and head-up tilt is expected to change the atrial fibrillatory rate, as the authors themselves note for HDT based on reference [24]. A systematic shift in the generated AA input between supine and HUT could therefore produce the observed decreases in estimated phi even if the true AV-nodal properties were unchanged. Moreover, the mean HUT-vs-supine changes in Table III (-4.7 ms for R_FP, -21.6 ms for R_SP, -4.1 ms for D_FP, +2.0 ms for D_SP) are all smaller than the corresponding ECG-PF one-minute averaged errors in Table I (92.3, 23.6, 89.6, and 152 ms, respectively). The significant Wilcoxon results are thus not sufficient to establish that the estimates track AV-nodal modulation rather than systematic input-related bias. The authors should control for this confound, for example by including the f-wave frequency trend as a covariate, by fixing the input AA statistics across phases, or by validating against EGM-based estimates during tilt.","section":"Section III-C and Table III"},{"comment":"The abstract's claim that 'beat-to-beat estimation of AV nodal conduction properties during AF from ECG is feasible' is substantially stronger than the paper's own error analysis supports. Section IV states that variations smaller than 169 ms in R_FP, 178 ms in D_FP, and 178 ms in D_SP cannot be identified, and that only R_SP has a beat-to-beat error (67 ms, 6% relative) that may allow detection of typical beat-to-beat changes. The conclusion similarly narrows the feasible claim to 'capturing beat-to-beat changes in the refractory period of the SP.' The central claim should be rephrased to specify which properties are trackable and at what temporal resolution, and the abstract should not present the overall simulation errors as evidence of uniform feasibility.","section":"Abstract and Section IV"}],"minor_comments":[{"comment":"The caption lists 'RF P' for all four panels (a)-(d); the last three should be RSP, DFP, and DSP, respectively.","section":"Figure 3 caption"},{"comment":"The sentence 'reducing its affect on the output' should read 'reducing its effect on the output.'","section":"Section IV, paragraph on estimation accuracy"},{"comment":"The caption contains a typo: 'following five minues in HDT' should be 'five minutes.'","section":"Figure 6 caption"},{"comment":"The choices of the variance factor 4 in the AA-generation model and sigma_alpha = 4 sigma_f are described as empirical, but no sensitivity analysis is provided; the robustness of the ECG-PF estimates to these tuning parameters is therefore not established.","section":"Section II-B2"},{"comment":"The inconsistent spacing in 'A V node' (sometimes 'AV node') should be made uniform.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a real methods advance—beat-to-beat estimation of four AV-nodal properties from surface ECG using a network model, particle filter, and smoothing—and the authors are unusually candid about its limits. The simulation study is self-consistency (the same model generates and validates the ground truth), which they openly acknowledge in Section IV.A. The only external evidence is the tilt test, and the stress-test note about the f-wave confound lands: the ECG-PF synthesizes its AA series from f-wave characteristics, and HUT changes atrial fibrillatory rate, so the estimated AV-node changes could be an artifact of changed atrial input. The reported tilt effects (a few ms to about 20 ms) are also smaller than the one-minute averaged errors in Table I (up to 152 ms), so the significant Wilcoxon results are less convincing than the abstract claims.\n\nCredit where due: the model and filtering framework are clearly described, the code is on GitHub, and the EGM-vs-ECG comparison on the iafdb data is a useful sanity check. The authors also explicitly state that beat-to-beat variations in R_FP and D_FP cannot be identified given the errors—more honest than most papers in this space. The novelty over their prior 10-minute-resolution work is genuine.\n\nThe soft spots are real but not disqualifying for a methods paper. Besides the circular simulation validation and the f-wave confound, the patient numbers are small (5 for the EGM comparison, 21 for tilt), and the fast pathway estimates are noisy. The authors partially acknowledge the confound in the Discussion but do not control for it, so the external validation is weaker than the abstract's language suggests.\n\nRecommendation: send to peer review. The method is novel, the limitations are transparent, and a good referee will push for either a direct control of the f-wave confound or a more conservative interpretation of the tilt result. It deserves referee time, not a desk reject.","headline":"Genuine beat-to-beat AV-node estimation method with honest limitations, but the tilt-test validation is confounded by f-wave changes and the effect sizes sit below the estimator's noise floor; worth reviewing seriously.","tokens_in":17887,"tokens_out":2266,"would_cite":true,"duration_ms":23298,"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":"The paper claims that the AV node's fast- and slow-pathway refractory periods and conduction delays can be estimated from surface ECG one heartbeat at a time during atrial fibrillation, with the slow-pathway refractory period recovered…","keywords":["atrioventricular node","atrial fibrillation","particle filter","smoothing algorithm","beat-to-beat estimation","refractory period","conduction delay","autonomic nervous system"],"falsifier":"Record simultaneous surface ECG and direct His-bundle or AV-node electrograms in patients during AF while performing a tilt or drug-induced sympathetic challenge; if the ECG-based beat-to-beat posterior distributions fail to cover the invasively measured refractory-period and conduction-delay changes in roughly 95% of beats, or fail to show the expected shortening under sympathetic stimulation, the claimed feasibility and uncertainty calibration would be refuted.","tokens_in":16781,"feed_emoji":"🫀","tokens_out":12789,"duration_ms":129977,"temperature":0.7,"pith_summary":"This paper tries to establish that the electrical gatekeeping properties of the atrioventricular node—how long each of its two pathways must rest before conducting again, and how slowly it conducts—can be estimated from a surface ECG one heartbeat at a time while a patient is in atrial fibrillation. That would matter because in AF the usual sinus-rhythm markers of autonomic nervous system activity are unavailable, and the AV node is the main place where autonomic tone still acts to set the ventricular rate. The method couples a two-pathway network model of the AV node with a particle filter that proposes candidate parameter sets and a smoothing algorithm that draws posterior trajectories, producing beat-to-beat distributions of $\\phi = [R_{FP}, R_{SP}, D_{FP}, D_{SP}]$. On model-generated ground truth the slow-pathway refractory period is recovered with a mean absolute error near 67 ms, the other three properties are noisier, estimates from ECG and synchronized intracardiac recordings differ by less than 5% on average, and head-up tilt produces decreases in the estimated properties in the direction expected from sympathetic activation.","feed_headline":"Beat-to-beat AV node readout during AF is feasible","feed_subtitle":"Particle filter plus AV node model tracks slow-pathway refractoriness to ~67 ms and detects expected autonomic changes.","key_machinery":"The central object is the two-pathway AV node network model: each pathway has ten nodes plus a coupling node, incoming impulses are blocked if the preceding diastolic interval $\\tilde t_i(n)$ is negative, and after each conducted impulse node $i$ updates its refractory period $R_i(n) = R_{\\min} + \\Delta R(1 - e^{-\\tilde t_i(n)/\\tau_R})$ and its conduction delay $D_i(n) = D_{\\min} + \\Delta D e^{-\\tilde t_i(n)/\\tau_D}$. This model is embedded in a particle filter, a sampling-based Bayesian filter whose particles are the twelve model parameters $\\theta$; the filter simulates each candidate parameter's ventricular activation times, weights them against the observed RR series, resamples, and propagates with Gaussian noise. Impulses are propagated through the 21-node network with a modified Dijkstra algorithm, an event-based shortest-path routine. A forward filtering backward sampling smoother then draws $M = 20{,}000$ trajectories from the posterior, yielding beat-to-beat distributions of $\\phi$. For ECG-only data, the atrial activation series are not known, so 40,000 parameter particles are copied 25 times and each copy is evaluated with a different Gaussian-random-walk atrial activation series derived from the f-wave frequency, which is the mechanism that lets the non-invasive version run.","core_discovery":"The paper claims that AV nodal conduction properties, previously not assessable during atrial fibrillation, can be followed from beat to beat from non-invasive recordings. The authors posit that the AV node behaves as two converging pathways whose ten nodes each recover with an exponential time course set by the preceding diastolic interval, and that a particle filter over the model's twelve parameters, followed by forward filtering backward sampling, can turn an observed RR interval series into full posterior distributions for $\\phi$. They report that the most probable estimates track simulated truth best for the slow-pathway refractory period (mean absolute error $67\\pm10$ ms, about 6% of its simulated range), are largest for the slow-pathway conduction delay ($178\\pm28$ ms, about 16%), and that one-minute averages cut the fast-pathway refractory-period error to about 92 ms. Comparing synchronized EGM and ECG estimates in five patients gave average agreement within 5% for all four properties, and in 21 tilt-test patients head-up tilt significantly decreased $R_{FP}$, $R_{SP}$, and $D_{FP}$, matching the expected effect of sympathetic activation. The conclusion is that beat-to-beat ECG-based estimation is feasible with different, quantifiable uncertainty levels per property.","pith_inferences":["If these results transfer to real physiology, beat-to-beat AV node monitoring offers an autonomic nervous system readout during AF, replacing heart rate variability where the sinus node signal is absent; one natural application is titrating rate-control drugs against each patient's AV node response.","The strong asymmetry in identifiability—slow-pathway refractoriness reliably, fast-pathway refractoriness and both delays noisily—suggests that future clinical studies should choose endpoints accordingly, using $R_{SP}$ for beat-to-beat analyses and one-minute averages for the other properties.","The largest source of uncertainty in the ECG version is the unobserved atrial activation timing; a model that jointly infers the atrial activation series and $\\phi$ instead of sampling Gaussian random walks might substantially shrink the $D_{FP}$ and $D_{SP}$ errors.","Because the simulated validation is generated from the same model used for estimation, independent invasive validation, such as direct His-bundle measurements during AF or a pharmacological autonomic challenge, would be the decisive next test of whether the estimated quantities correspond to real AV node physiology."],"forward_implications":["The slow-pathway refractory period can be tracked beat-to-beat from surface ECG alone: a mean absolute error of about 67 ms against simulated truth means changes of roughly that size or larger are resolvable at individual beats.","Fast-pathway refractory period and the two conduction delays are not resolvable beat-to-beat from ECG; one-minute or phase averages are needed, with average errors of about 92 ms for $R_{FP}$ and 90 to 152 ms for the delays.","Autonomic modulation of the AV node during AF is observable non-invasively at the group level: the tilt protocol shows the expected sympathetic shortening of $R_{FP}$, $R_{SP}$, and $D_{FP}$ from supine to head-up tilt.","Because ECG-based estimates agree with EGM-based estimates within 5% on average, much of the information available from an intracardiac catheter near the AV node is preserved in the surface ECG.","The reported 95% credibility regions cover the simulated ground truth in 93–99.8% of beats, so the uncertainty bounds produced by the method are conservative."],"supporting_citations":[{"why":"Supplies the two-pathway, ten-node-per-pathway network model with exponential recovery dynamics and the open-source model code that both particle filters use as their simulator.","marker":"[15]"},{"why":"Supplies the forward filtering backward sampling algorithm that turns the particle filter's weighted samples into the $M = 20{,}000$ posterior trajectories used for beat-to-beat estimates.","marker":"[22]"},{"why":"Supplies the synchronized EGM/ECG atrial fibrillation recordings used to compare invasive and non-invasive estimates and to build realistic simulated data.","marker":"[23]"},{"why":"Supplies the tilt-test ECG recordings from 21 patients used to evaluate whether the estimates respond to autonomic changes as expected.","marker":"[24]"},{"why":"Provides the earlier 10-minute-resolution estimation framework on the same AV node model; its errors are the baseline against which the beat-to-beat results are compared.","marker":"[17]"},{"why":"Provides the modified Dijkstra algorithm used to propagate impulses through the 21-node network, the computational engine that evaluates each candidate parameter vector.","marker":"[30]"},{"why":"Documents that autonomic nervous system activity modulates AV nodal conduction at beat-to-beat resolution, which is the physiological signal the method aims to capture.","marker":"[18]"},{"why":"Provides the f-wave signal-quality and frequency model used to generate synthetic atrial activation series for the ECG-only particle filter.","marker":"[28]"}],"fun_headline_variants":["ECG beat-to-beat AV node tracking during AF works","Particle filter reads AV node beat-by-beat in AF","AV node conduction followed beat-to-beat from ECG during AF","Slow-pathway refractoriness tracked to 67 ms in AF"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole validation rests on the assumption that a two-pathway network with ten nodes per pathway and exponential recovery dynamics is an adequate description of the real human AV node under atrial fibrillation; if that model is not faithful, the simulated ground truth used to measure accuracy is not the truth the method needs to estimate.","fun_headline_variants_meta":{"raw":{"variants":["ECG beat-to-beat AV node tracking during AF works","Particle filter reads AV node beat-by-beat in AF","AV node conduction followed beat-to-beat from ECG during AF","Slow-pathway refractoriness tracked to 67 ms in AF"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1358,"prompt_tokens":1131,"completion_tokens":227,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":747,"completion_tokens_details":{"reasoning_tokens":167}},"tokens_in":747,"tokens_out":227,"duration_ms":2903,"temperature":1.0,"reasoning_tokens":167,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:06:42.133704+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record simultaneous surface ECG and direct His-bundle or AV-node electrograms in patients during AF while performing a tilt or drug-induced sympathetic challenge; if the ECG-based beat-to-beat posterior distributions fail to cover the invasively measured refractory-period and conduction-delay changes in roughly 95% of beats, or fail to show the expected shortening under sympathetic stimulation, the claimed feasibility and uncertainty calibration would be refuted.","supporting_citations":[{"cited_title":"Non-invasive characterization of human A V-nodal conduction delay and refractory period during atrial fibrillation,","cited_arxiv_id":null,"evidence_quote":"Supplies the two-pathway, ten-node-per-pathway network model with exponential recovery dynamics and the open-source model code that both particle filters use as their simulator."},{"cited_title":"Chopin, O","cited_arxiv_id":null,"evidence_quote":"Supplies the forward filtering backward sampling algorithm that turns the particle filter's weighted samples into the $M = 20{,}000$ posterior trajectories used for beat-to-beat estimates."},{"cited_title":"Physiobank, physiotoolkit, and physionet: Components of a new research resource for complex physiologic signals,","cited_arxiv_id":null,"evidence_quote":"Supplies the synchronized EGM/ECG atrial fibrillation recordings used to compare invasive and non-invasive estimates and to build realistic simulated data."},{"cited_title":"Autonomic influence on atrial fibrillatory process: head-up and head-down tilting,","cited_arxiv_id":null,"evidence_quote":"Supplies the tilt-test ECG recordings from 21 patients used to evaluate whether the estimates respond to autonomic changes as expected."},{"cited_title":"Model-based estimation of av-nodal refractory period and conduction delay trends from ECG,","cited_arxiv_id":null,"evidence_quote":"Provides the earlier 10-minute-resolution estimation framework on the same AV node model; its errors are the baseline against which the beat-to-beat results are compared."},{"cited_title":"Characterisation of human A V-nodal properties using a network model,","cited_arxiv_id":null,"evidence_quote":"Provides the modified Dijkstra algorithm used to propagate impulses through the 21-node network, the computational engine that evaluates each candidate parameter vector."},{"cited_title":"Rate-related and autonomic effects on atrioventricular conduction assessed through beat-to-beat pr interval and cycle length variability,","cited_arxiv_id":null,"evidence_quote":"Documents that autonomic nervous system activity modulates AV nodal conduction at beat-to-beat resolution, which is the physiological signal the method aims to capture."},{"cited_title":"Model-based assessment of f-wave signal quality in patients with atrial fibrillation,","cited_arxiv_id":null,"evidence_quote":"Provides the f-wave signal-quality and frequency model used to generate synthetic atrial activation series for the ECG-only particle filter."}],"review_version":1}