{"id":"0eea5376-4d35-49c6-aef3-85d8c4cd57a2","arxiv_id":"2508.13451","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A first ground-truth benchmark shows center-of-mass localization, although less accurate than model-based methods in ideal conditions, is the most robust to electrode decay.","lead":"This paper benchmarks three algorithms that estimate neuron locations from high-density electrode recordings, using both simulated and real paired patch-clamp and Neuropixels data. It finds that the simple center-of-mass heuristic is most robust to electrode decay, while more complex models are more accurate in ideal conditions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Robustness ranking relies on a binary random-electrode dropout decay model; under gradual or correlated electrode degradation the COM advantage may not generalize to real long-term recordings.","rationale":"The reader's weakest assumption—that electrode decay is faithfully modeled by randomly replacing a fraction of electrodes with Gaussian noise—is the same load-bearing concern I identify. The abstract and discussion generalize from this simulation to real long-term recordings, so the robustness ranking is only as strong as the decay model. My proposed alternative-decay tests directly probe whether the COM advantage is an artifact of binary random dropout. The reader's CONDITIONAL verdict already appropriately conditions on validation of the decay model and code release, so I do not recommend changing the verdict. Secondary issues such as the selective use of 11 experimental pairs and the GC equation typo are worth fixing but are less central to the main claim than the unvalidated decay mechanism.","tokens_in":8524,"tokens_out":3250,"duration_ms":41100,"concrete_test":"Rerun the Figs. 3-4 benchmark on the same MEArec and SPE-1 ground-truth datasets under at least two alternative decay models: (a) gradual multiplicative attenuation of electrode sensitivity (e.g., log-normally distributed gain factors from 0.1 to 0.9 instead of binary replacement) and (b) spatially correlated dead electrodes (clustered blocks rather than uniform random selection). Compare COM/MT/GC accuracy, RMSE, and drift at equivalent aggregate signal-loss levels (30%, 60%, 90%). If COM retains its advantage, the concern is resolved; if MT or GC improves relative to COM, or COM's error increases sharply, the paper's central claim must be narrowed to complete random electrode failure only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central conclusion that COM is 'more suitable for long-term neural recordings' rests on the electrode-decay experiments in Figs. 3-4, where decay is simulated by replacing a random fraction of electrode signals with Gaussian noise. This is a binary, spatially uncorrelated failure model. Real chronic probes typically show gradual impedance increases, reduced gain, and spatially localized or correlated damage; failures are not instantaneous and may cluster near insertion sites or damaged regions. The comparison is not neutral across algorithms: MT solves an optimization over surviving electrodes, so abrupt binary dropout removes constraints and can trigger overfitting, while COM's estimate shrinks toward the convex hull of the remaining electrodes—under random dropout this hull remains centered, yielding 'stable' but biased estimates. With graded sensitivity loss (e.g., 20–80% attenuation) or clustered failures, MT's loss landscape changes less catastrophically and COM's shrinkage could become a larger source of error. Because no validation of the decay model against real chronic recordings is provided, the claimed superiority of COM under long-term conditions is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a benchmark of three spike source localization algorithms—center of mass (COM), monopolar triangulation (MT), and grid convolution (GC)—using two ground-truth datasets: a simulated MEArec dataset with known neuron positions, and the experimental SPE-1 dataset combining patch-clamp and Neuropixels recordings. The authors evaluate localization accuracy, runtime, and robustness under ideal conditions and under simulated electrode decay, where an increasing fraction of electrode signals is replaced with Gaussian noise. The main finding is that MT and GC are more accurate than COM in ideal simulated conditions, but COM (and to a lesser extent GC) is more robust to electrode decay, leading the authors to conclude that simple heuristics such as COM are more suitable for long-term neural recordings. The paper also provides a public code repository and a framework for future benchmarking.","tokens_in":8726,"tokens_out":3308,"duration_ms":36762,"significance":"If the central claim holds, this is a useful contribution to a nascent field: it provides the first ground-truth comparison of these widely used localization algorithms using both a biophysical simulator and a paired patch-clamp/Neuropixels dataset. The open code and use of public datasets are strengths that support reproducibility. However, the headline conclusion about long-term suitability depends heavily on the realism of the electrode-decay model and on the representativeness of the selected experimental subset, both of which are currently not sufficiently validated. The paper is therefore potentially significant, but the strength of the claim exceeds what the evidence currently supports.","major_comments":[{"comment":"The electrode-decay model is implemented by replacing the signals of a randomly selected fraction of electrodes with Gaussian noise. This is a binary, abrupt, spatially uncorrelated failure model. Real chronic electrode degradation in vivo is typically gradual (e.g., rising impedance, reduced gain) and can be spatially clustered or correlated. The central abstract claim that COM is 'more suitable for long-term neural recordings' rests on this simulation. No validation against real chronic recordings is provided, and the choice of model is not neutral: binary dropout removes constraints from MT's optimization, potentially explaining MT's overfitting, while COM's shrinkage toward the convex hull of the remaining electrodes remains centered under random dropout. The authors should either validate the decay model against empirical chronic-recording data or substantially soften the long-term","section":"Benchmarking results against decay (Figs. 3–4)"},{"comment":"The paper states that the SPE-1 dataset contains 43 paired recordings but that only 11 pairs were selected 'with the strongest signal quality and confidence of patch-clamp and MEA pairing.' No objective or pre-specified criteria are given for this selection. Since the experimental results (Fig. 2D and Fig. 4) are based entirely on these 11 pairs, the selection could bias the comparison, especially if pairing confidence correlates with signal-to-noise ratio or neuron depth. The manuscript should report the total number of available pairs, the criteria used, and ideally show that the main conclusions are robust to including more pairs or a more inclusive subset.","section":"Methods, Experimental Dataset"},{"comment":"There is a direct internal contradiction in the statistical reporting. The Fig. 2 caption states that 'all metrics exhibited significant differences (p < 0.05),' while the text states that 'No significant differences were found in template accuracy across algorithms (one-way ANOVA p = 0.0409; all pairwise comparisons p >0.05).' If the ANOVA is significant but pairwise comparisons are not, the caption should not claim that all metrics differed significantly. This inconsistency undermines the reader's ability to trust the statistical summaries and must be corrected.","section":"Results, 'Benchmarking results in absence of electrode decay' vs. Fig. 2 caption"},{"comment":"Even under the chosen binary dropout model, the decay-robustness results are reported only as bar/error-bar plots without any statistical tests. The text states that 'COM and GC are more robust against electrode decay than MT,' but no ANOVA, paired comparison, or confidence interval is provided across decay levels. Given that the central long-term-recording claim is based on these data, the authors should provide quantitative statistical support, e.g., repeated-measures comparisons across decay levels, rather than relying on visual trends.","section":"Figs. 3–4 and decay robustness"}],"minor_comments":[{"comment":"The probe layout is described as 'spaced 30nm apart.' This should presumably be 30 µm (micrometers), not 30 nm. Please correct.","section":"Materials and methods, Simulated Dataset"},{"comment":"The loss function is written as Li = Σ_j (ptp_ij − V_ij), with the text stating that the location is estimated 'usually via least squares.' The least-squares objective should involve the square of the residual, i.e., Σ_j (ptp_ij − V_ij)^2. Please clarify whether the implementation uses squared residuals or absolute residuals, as this affects the algorithm's behavior.","section":"Eq. for Monopolar Triangulation"},{"comment":"The equation τ_kj(t) = e^{(p_k − p_i)^2/(2σ^2)} w_k(t) appears to contain a typo: the exponent should likely involve the distance between the theoretical template position p_k and the electrode position p_j, not a spike position p_i. Please define all symbols and correct the equation.","section":"Eq. for Grid Convolution"},{"comment":"The abstract describes 'simple heuristics' as demonstrating superior robustness, but the results identify COM and, in some cases, GC as robust. GC is not a simple heuristic; it is a grid-based template method. Please align the language with the actual algorithms tested.","section":"Abstract and Discussion"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is plausible but currently overreaches the evidence. The binary random-dropout decay model, the selective use of 11/43 experimental pairs, and the inconsistent statistical reporting all need to be addressed before publication. The code repository availability is a positive feature, but I would recommend verifying that the repository contains the exact analysis scripts and version used. The scope is appropriate for q-bio.NC."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before the field re-discovers it: it is the first ground-truth comparison of COM, MT, and GC spike localization under simulated electrode decay, using both a biophysical simulation and the patch-clamp/Neuropixels SPE-1 dataset. The core finding—that simple COM is more robust than MT when electrodes drop out—is plausible and worth taking seriously. The decay protocol is explicit, the datasets are independent, and the framework is reusable. That is a real contribution.\n\nThe soft spots are real but not fatal. The biggest is the decay model itself: replacing a random fraction of electrodes with Gaussian noise is a binary, spatially uncorrelated failure mode. Real chronic probes show gradual impedance increases, gain loss, and clustered damage. The stress-test note is right that MT, which optimizes over surviving electrodes, may be unfairly penalized by abrupt dropout, while COM's shrinkage toward the center of the convex hull looks stable under random dropout but could be misled under graded or correlated decay. The authors do not validate the model against any real chronic recordings, so the extension to \"long-term recordings\" is not yet established. I would ask for either a graded/clustered decay version or a clear caveat.\n\nThe manuscript also has avoidable internal problems. The Fig. 2 caption says all metrics showed significant differences (p < 0.05), while the text correctly reports that template accuracy showed no significant pairwise differences (ANOVA p = 0.0409, all pairwise p > 0.05). That contradiction will confuse readers. The selective use of 11 of 43 SPE-1 pairs needs a stated selection criterion, not just \"strongest signal quality.\" The GC equation uses (pk−pi) where pi is the unknown spike location—surely a typo for an electrode position—and should be fixed. The code repository is promised but not yet verifiable.\n\nWho is this for? Someone choosing a localization method for chronic recordings will get a clear, caveated ranking. Someone building the next algorithm will get a benchmark they can extend. The paper deserves a serious referee, not a desk reject, but the referee should push on the decay model and the statistical reporting.\n\nMy recommendation: send it out, with the expectation of major revisions. The central benchmark is solid enough to justify referee time, even if the robustness claim needs to be reined in.","headline":"Useful first benchmark of spike localization under electrode decay, but the headline robustness claim rests on a narrow decay model and the manuscript has a few internal cracks.","tokens_in":9167,"tokens_out":1412,"would_cite":true,"duration_ms":17314,"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 ground-truth benchmark of three neuron-localization algorithms finds that the simple center-of-mass heuristic, though less accurate in ideal conditions, stays most accurate as electrodes decay.","keywords":["spike source localization","high-density probes","center of mass","monopolar triangulation","grid convolution","electrode decay","long-term neural recordings","ground-truth benchmarking"],"falsifier":"Record from a chronically implanted high-density probe for months while monitoring electrode impedance, and compare localization accuracy against patch-clamp ground truth for neurons near electrodes that degrade gradually (rising impedance) rather than failing outright. If monopolar triangulation matches or beats center of mass under that natural decay pattern, the paper's central claim about long-term suitability fails; if COM still wins, the claim is confirmed outside the random-death simulation.","tokens_in":8410,"feed_emoji":"🧠","tokens_out":7763,"duration_ms":78966,"temperature":0.7,"pith_summary":"This paper sets out to settle, with ground-truth data, which of three spike source localization algorithms should be trusted when electrodes degrade over months of recording. The authors benchmark center-of-mass (COM), monopolar triangulation (MT), and grid convolution (GC) on a biophysically realistic simulated dataset and on paired patch-clamp and Neuropixels recordings. In ideal conditions the more physically elaborate methods MT and GC locate neurons more accurately, but once a fraction of electrodes is replaced by Gaussian noise, the simple COM heuristic (and, to a lesser degree, GC) stays accurate while MT overfits and produces large errors, especially for individual spikes. The conclusion, stated for long-term neural interfaces, is that COM's robustness makes it the most suitable method for chronic recordings despite its lower ceiling in ideal conditions. A benchmark framework of this kind matters because localization accuracy feeds directly into spike sorting and probe-drift monitoring.","feed_headline":"Simple center-of-mass rule wins when electrodes decay","feed_subtitle":"A ground-truth benchmark of three neuron-location algorithms shows the fastest, simplest one holds up best over long recordings.","key_machinery":"The argument is carried by three named localization algorithms compared under a common degradation protocol. COM estimates a spike's position as the peak-to-peak amplitude-weighted average of electrode positions, which is fast and simple but confined to the convex hull of the array. MT assumes a monopolar point source with voltage decaying as inverse distance and fits the location by least squares. GC builds a dense grid of theoretical templates with exponential signal decay and weights each grid position by the convolution between measured and theoretical waveforms. The degradation protocol replaces the recorded signal at a randomly chosen fraction of electrodes with Gaussian noise, and per","core_discovery":"The central discovery is an accuracy–robustness tradeoff among the three algorithms, measured against biological ground truth rather than theoretical appeal. On a simulated dataset with known neuron positions, grid convolution and monopolar triangulation beat center of mass on accuracy and RMSE in clean conditions, consistent with their more realistic physical models. On the noisy experimental dataset, however, center of mass and grid convolution outperform monopolar triangulation—and under simulated electrode decay, center of mass retains its accuracy while monopolar triangulation degrades sharply, producing overfit estimates for spike events. The paper reads this as evidence that the simpl","pith_inferences":["If real electrode aging is gradual impedance rise or correlated multi-electrode failure rather than random single-electrode death, the COM-versus-MT ranking is not guaranteed; the paper's decay model is the assumption most worth testing.","A hybrid strategy—use MT or GC while the array is healthy and switch to COM as electrode loss crosses a threshold—could combine ideal-condition accuracy with long-term stability.","The same ground-truth benchmark could be extended to probe drift and tissue micromotion, since the paper names drift monitoring as a downstream use but does not test localization under active drift.","The near-linear geometry of Neuropixels may amplify COM's advantage; on a square 2D array with denser coverage the accuracy gap in clean conditions could matter more in practice."],"forward_implications":["For chronic implants, where electrodes fail progressively, the simple COM and the grid-based GC should be preferred over MT for tracking neurons across days or months.","MT's overfitting under electrode loss is mainly a single-spike problem, so template-level localization is safer than spike-level localization when using MT on degraded arrays.","COM's speed, combined with its stability, makes it practical for real-time spike sorting on embedded or low-power devices.","Drift estimates computed from COM or GC locations will be less contaminated by electrode-loss artifacts than those from MT.","Future localization algorithms should be evaluated not only on clean accuracy but on the same decay-robustness axis, since real arrays always age."],"supporting_citations":[{"why":"supplies the monopolar triangulation approach and its use for three-dimensional spike localization on Neuropixels recordings","marker":"[8]"},{"why":"provides the unified spike-sorting framework in which COM, MT, and GC are implemented and compared","marker":"[9]"},{"why":"supplies the biophysically realistic simulator that generates the simulated ground-truth recordings","marker":"[20]"},{"why":"provides the paired patch-clamp and Neuropixels ground-truth dataset used for experimental benchmarking","marker":"[21]"},{"why":"sets the simulation and benchmarking protocol the paper follows for ground-truth recordings","marker":"[5]"},{"why":"introduces the inverse-distance monopolar source model that MT assumes for voltage decay","marker":"[22]"},{"why":"extends the monopolar source localization model to cortical tetrode recordings, supporting MT's physical basis","marker":"[23]"},{"why":"supports the claim that MT and GC rest on more physically realistic models than COM","marker":"[26]"},{"why":"provides a modeling study of extracellular action potential waveforms underlying the physical assumptions of the elaborate algorithms","marker":"[27]"}],"fun_headline_variants":["Simple neuron locator beats complex models in long runs","Center of mass survives electrode decay best","Neuron localization: simplest algorithm wins in real conditions","Ground-truth test: simple rule trumps complex physics","Electrode decay? Simpler spike location algorithm wins"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The robustness ranking depends on the assumption that electrode decay is faithfully represented by replacing the signal at a randomly selected fraction of electrodes with Gaussian noise; if real electrode degradation follows a different pattern, the observed advantage of the simple heuristic could be an artifact of that simulation.","fun_headline_variants_meta":{"raw":{"variants":["Simple neuron locator beats complex models in long runs","Center of mass survives electrode decay best","Neuron localization: simplest algorithm wins in real conditions","Ground-truth test: simple rule trumps complex physics","Electrode decay? Simpler spike location algorithm wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000233,"raw_usage":{"total_tokens":1307,"prompt_tokens":701,"completion_tokens":606,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":532}},"tokens_in":445,"tokens_out":606,"duration_ms":6251,"temperature":1.0,"reasoning_tokens":532,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T19:01:41.244811+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record from a chronically implanted high-density probe for months while monitoring electrode impedance, and compare localization accuracy against patch-clamp ground truth for neurons near electrodes that degrade gradually (rising impedance) rather than failing outright. If monopolar triangulation matches or beats center of mass under that natural decay pattern, the paper's central claim about long-term suitability fails; if COM still wins, the claim is confirmed outside the random-death simulation.","supporting_citations":[{"cited_title":"Three-dimensional spike localization and improved motion correction for neuropixels recordings","cited_arxiv_id":null,"evidence_quote":"supplies the monopolar triangulation approach and its use for three-dimensional spike localization on Neuropixels recordings"},{"cited_title":"SpikeInterface, a unified framework for spike sorting","cited_arxiv_id":null,"evidence_quote":"provides the unified spike-sorting framework in which COM, MT, and GC are implemented and compared"},{"cited_title":"MEArec: A fast and customizable testbench simulator for ground-truth extracellular spiking activity","cited_arxiv_id":null,"evidence_quote":"supplies the biophysically realistic simulator that generates the simulated ground-truth recordings"},{"cited_title":"Recording from the same neuron with high-density CMOS probes and patch-clamp: a ground-truth dataset and an experiment in collaboration","cited_arxiv_id":null,"evidence_quote":"provides the paired patch-clamp and Neuropixels ground-truth dataset used for experimental benchmarking"},{"cited_title":"A modular implementation to handle and benchmark drift correction for high-density extracellular recordings","cited_arxiv_id":null,"evidence_quote":"sets the simulation and benchmarking protocol the paper follows for ground-truth recordings"},{"cited_title":"Spike source localization with tetrodes","cited_arxiv_id":null,"evidence_quote":"introduces the inverse-distance monopolar source model that MT assumes for voltage decay"},{"cited_title":"Three-dimensional localization of neurons in cortical tetrode recordings","cited_arxiv_id":null,"evidence_quote":"extends the monopolar source localization model to cortical tetrode recordings, supporting MT's physical basis"},{"cited_title":"The 3D position estimation of neurons in the hippocampus based on the multi-site multi-unit recordings with silicon tetrodes","cited_arxiv_id":null,"evidence_quote":"supports the claim that MT and GC rest on more physically realistic models than COM"},{"cited_title":"On the origin of the extracellular action potential waveform: A modeling study","cited_arxiv_id":null,"evidence_quote":"provides a modeling study of extracellular action potential waveforms underlying the physical assumptions of the elaborate algorithms"}],"review_version":1}