Spike sorting for large, dense electrode arrays

Spike sorting for large, dense electrode arrays Rossant Kadir Goodman Schulman Hunter Saleem Grosmark Belluscio

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Spike sorting for large, dense electrode arrays 2016 Nature Neuroscience 10.1038/nn.4268 Rossant2016-spike-sorting-for-large-dense-electrode-arra Rossant2016-spike-sorting-for-large-dense-electrode-arra.txt, the project's staged copy 1b05873fbebd117c653c88744db7f07eebb75626 not-redistributable neuroscience

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them how to recover the firing times of individual neurons from the raw voltage recorded by a dense electrode array - the step that manufactures the single unit Quotation withheld because this record is not cleared for redistribution. whether a practical system can sort high-channel-count data at error rates comparable to tetrode methods, in human and computer time a working lab will accept Quotation withheld because this record is not cleared for redistribution. a three-stage pipeline (masked detection, masked EM clustering, guided manual curation) reaching about 5% error, plus a measurement of how much of the result depends on which human curates it Quotation withheld because this record is not cleared for redistribution. method Quotation withheld because this record is not cleared for redistribution. spike sorting for high-density electrode arrays Quotation withheld because this record is not cleared for redistribution. system-or-tool-building system-or-tool-building Quotation withheld because this record is not cleared for redistribution. mechanisms noise on many uninformative channels swamps signal on few informative channels, so Gaussian-mixture clustering breaks down as dimensionality grows Quotation withheld because this record is not cleared for redistribution. mechanisms probabilistically disenfranchising low-mask channels removes their contribution to cluster assignment and so prevents threshold-crossing artefacts from generating spurious clusters Quotation withheld because this record is not cleared for redistribution. mechanisms requiring a strong-threshold crossing inside each weak-threshold connected component discards isolated noise islands while merging weakly connected regions of one spike Quotation withheld because this record is not cleared for redistribution. mechanisms cost scales with unmasked channels per spike rather than total channels, which is why the method stays tractable as arrays grow Quotation withheld because this record is not cleared for redistribution. overlapping spikes are separated not by their features, which are nearly identical, but by which channels they appear on - the mask vector Quotation withheld because this record is not cleared for redistribution. the curse of dimensionality: on a large array the noise on uninformative channels outweighs the signal on the few informative ones Quotation withheld because this record is not cleared for redistribution. masked EM disenfranchises low-mask channels probabilistically, so no false cluster is created when an amplitude happens to cross a threshold Quotation withheld because this record is not cleared for redistribution. two thresholds rather than one: the weak one prevents a spike being split by noise, the strong one prevents noise being kept as a spike Quotation withheld because this record is not cleared for redistribution. the human cost of curation is quadratic in clusters unless the software orders the candidate merges, which makes it linear Quotation withheld because this record is not cleared for redistribution. complexity scales with unmasked channels per spike, a quantity that does not grow as arrays get bigger Quotation withheld because this record is not cleared for redistribution. Clustering can still take hours or days on large datasets with a single-core machine. what this limits: the practical claim of usability at scale, not the accuracy results: IT CAN TAKE HOURS OR EVEN DAYS TO COMPLETE ON A STANDARD SINGLE-CORE MACHINE Quotation withheld because this record is not cleared for redistribution. Spikes that overlap in both space and time cannot be resolved by the detection algorithm. what this limits: every spike count and every unit derived from dense-array data: IT CANNOT RESOLVE SPIKES THAT OVERLAP IN BOTH SPACE AND TIME - those spikes are absent from the output Quotation withheld because this record is not cleared for redistribution. Operators' judgements of which clusters count as well-isolated units are subjective, so the pipeline still depends on a non-reproducible human step. what this limits: any claim that a set of well-isolated single units is operator-independent: Quotation withheld Quotation withheld because this record is not cleared for redistribution. spikes that overlap in both space and time cannot be resolved at all what this limits: the same scope as limitations[1] - spatiotemporally overlapping spikes are unresolved Quotation withheld because this record is not cleared for redistribution. ground truth is simulated, and rests on the assumption that the extracellular medium is a linear conductor so waveforms can simply be added what this limits: the accuracy figures only, which are measured against this truth set: WE CREATED A SIMULATED GROUND TRUTH DATASET ... SINCE THE EXTRACELLULAR MEDIUM IS A LINEAR CONDUCTOR - the validation assumes linearity Quotation withheld because this record is not cleared for redistribution. the bursting interspike intervals that make real spike trains hard were deliberately not simulated what this limits: the generalisation of those accuracy figures to real bursting cells: THE INTERSPIKE INTERVALS TYPICAL OF BURSTING NEURONS WERE NOT SIMULATED Quotation withheld because this record is not cleared for redistribution. which clusters count as well-isolated single units is not reproducible between operators what this limits: the reproducibility of the unit yield between labs: THE JUDGEMENT OF WHETHER A CLUSTER SHOULD BE CONSIDERED A WELL-ISOLATED SINGLE UNIT was not consistent across operators Quotation withheld because this record is not cleared for redistribution. the waveform changes that force manual correction have resisted mathematical modelling what this limits: any expectation that manual curation can be removed: WAVEFORMS ARE NOT CONSTANT, BUT CHANGE ACCORDING TO FIRING - the errors forcing manual correction resist modelling Quotation withheld because this record is not cleared for redistribution. runtime can reach hours or days on a single core what this limits: the same scope as limitations[0] - runtime, not accuracy Quotation withheld because this record is not cleared for redistribution. template matching, which would solve the overlap problem, is blocked less by the algorithm than by what it would do to the curation step what this limits: the prospect of solving the overlap problem by template matching: SUCH METHODS ARE NOT AT PRESENT WIDELY APPLIED TO IN VIVO RECORDINGS Quotation withheld because this record is not cleared for redistribution. the EM algorithm for mixtures of Gaussians, modified so that uninformative dimensions are disenfranchised rather than weighted Quotation withheld because this record is not cleared for redistribution. the flood fill, used to grow a spike as a connected component in space and time Quotation withheld because this record is not cleared for redistribution. the Fowlkes-Mallows index, used here to score agreement between two human operators rather than between two clusterings Quotation withheld because this record is not cleared for redistribution. the precedent that automatic methods in data-intensive science run under human supervision Quotation withheld because this record is not cleared for redistribution. every added spike displaced by a RANDOM SUB-SAMPLE OFFSET via cubic spline interpolation before being written into the simulated data eliminates: the test being passable without sub-sample alignment - spikes added on the sampling grid would let a sorter that ignores sub-sample timing score as well as one that does not, so the benchmark would not test the thing it claims to · outcome (stated by the tagger, no quotation): held Quotation withheld because this record is not cleared for redistribution. hybrid datasets hybrid datasets Quotation withheld because this record is not cleared for redistribution. ground truth here = SIMULATED - a known cell's waveforms added into another recording - not intracellular recording of the same neuron; the word is used for both in this literature Quotation withheld because this record is not cleared for redistribution. well-isolated single unit = a cluster an operator accepts; the paper shows this judgement is not shared between operators, so the term does not denote a fixed object Quotation withheld because this record is not cleared for redistribution. manual curation = correcting the automatic clustering, not selecting data; it is a required step of the measurement, not a post-hoc filter Quotation withheld because this record is not cleared for redistribution. mask vector = which channels a spike appeared on, rescaled to [0,1]; distinct from the feature vector, which is the PCA summary of its shape Quotation withheld because this record is not cleared for redistribution. theoretical optimum = the Best Ellipsoid Error Rate, a quadratic decision boundary fitted to ground truth and cross-validated - a benchmark, not a bound on any possible algorithm Quotation withheld because this record is not cleared for redistribution. claim_type: method a second reading reaching the same value. Recorded: method — note: a second reading reaching the same value from a different sentence. Not applied - a person decides. Quotation withheld because this record is not cleared for redistribution. method: algorithm design validated on simulated ground truth built by adding a donor cell's waveforms into a second recording, plus a study of 8 human operators curating the same dataset a second reading DIFFERING from the recorded value. Recorded: system-or-tool-building — note: a second reading, verbatim-verified, disagreeing with the value on the record. Not applied - a person decides. Quotation withheld because this record is not cleared for redistribution. phenomenon: spike sorting: the inference from extracellular waveforms to the identity of the neuron that produced them a second reading DIFFERING from the recorded value. Recorded: spike sorting for high-density electrode arrays — note: a second reading, verbatim-verified, disagreeing with the value on the record. Not applied - a person decides. Quotation withheld because this record is not cleared for redistribution. builds-on → KlustaKwik, first version (tetrode-oriented hard EM mixture-of-Gaussians program released in 2000) Quotation withheld because this record is not cleared for redistribution. builds-on → Klusters manual spike-sorting environment (ref. 13) Quotation withheld because this record is not cleared for redistribution. builds-on → Best Ellipsoid Error Rate (BEER) measure (ref. 8) Quotation withheld because this record is not cleared for redistribution. extends → semi-automatic tetrode spike sorting, which reached 5% error against intracellular ground truth Quotation withheld because this record is not cleared for redistribution. contradicts → the assumption that tetrode sorting methods carry over to large arrays Quotation withheld because this record is not cleared for redistribution. critiques → the practice of deciding by eye which units are well isolated differs in: THE AXIS IS WHO OR WHAT DECIDES ISOLATION QUALITY. 'experienced operators tended to make similar judgements during the manual curation process, but Quotation withheld ... we therefore recommend that these [quantitative criteria] be used, rather than human judgements.' Note the shape: agreement between operators is granted and is not treated as evidence of correctness - concordance and validity are separated explicitly. Quotation withheld because this record is not cleared for redistribution. distinguishes-from → template-matching sorters validated on in vitro retinal arrays differs in: THE AXIS IS THE NOISE LEVEL OF THE DATA THE METHOD WAS VALIDATED ON. 'TEMPLATE-MATCHING ALGORITHMS HAVE SOLVED THIS PROBLEM IN THE CASE OF IN VITRO RETINAL ARRAY DATA, BUT THESE DATA ARE MUCH LESS NOISY THAN IN VIVO BRAIN RECORDINGS.' The method is granted its success and the success is bounded by the preparation it was demonstrated on - so 'solved' is true of one regime and unproven in the other, and a citation carries only the word. Quotation withheld because this record is not cleared for redistribution. critiques → the existing literature of proposed sorting methods, none of which was adopted Quotation withheld because this record is not cleared for redistribution. contradicts → the strategy of adding time as a clustering feature to track drift Quotation withheld because this record is not cleared for redistribution. RETRACTED (DIRECTION REVERSED) · manual curation —part-of→ spike sorting manual curation spike sorting RETRACTED: DIRECTION REVERSED. The edge says spike sorting is part of manual curation. The paper says the opposite: '…extraction, (2) cluster analysis, and (3) MANUAL CURATION…' Manual curation is a step OF spike sorting. Found while looking for a span naming both endpoints - the search for grounding surfaced a direction error. — strength of the relation: 1 · note: direction is member-to-container: curation is the fourth step of the sorting process, not an optional extra Quotation withheld because this record is not cleared for redistribution. masked EM algorithm —contrasted-with→ classical EM algorithm masked EM algorithm classical EM algorithm strength of the relation: 1 · note: the same data: masked near the BEER optimum, classical above 50% error Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED) · well-isolated —contrasted-with→ multiunit activity well-isolated multiunit activity EVIDENCE CORRECTED: RE-GROUNDED. The old span was about operator consistency. This sentence names both as the two outcomes of one decision - the operator '…WELL-ISOLATED neuron, OR rejects it as MULTIUNIT ACTIVITY…' - so the contrast is a forced choice rather than an observed difference. — strength of the relation: 1 · note: the boundary the operator study found to be operator-dependent Quotation withheld because this record is not cleared for redistribution. bursting —is-a→ nonstationarity bursting nonstationarity strength of the relation: 0.9 Audit 2026-09-28: DOUBTFUL - read against its span and kept, but the span models bursting as a source of waveform variability; is-a 'nonstationarity' is loose. Quotation withheld because this record is not cleared for redistribution. silicon probes —contrasted-with→ tetrode silicon probes tetrode strength of the relation: 1 · note: CRITERION, stated: overlapping spikes are extremely common on the dense probe and not on the tetrode Quotation withheld because this record is not cleared for redistribution. manual curation —part-of→ spike sorting manual curation spike sorting strength of the relation: 1 · note: REPLACES another row of this record, which ran the other way. The convention is from=part, to=whole - 'vertex --part-of--> network'. NO separate quote for the relation: the span's composition cue is 'involves', which is not in CONTAINER_FIRST, and its only other attestation in the corpus (Pickering2008) is fenced. One available occurrence does not earn a cue, so the clause-level check is forgone rather than the rule bent. Quotation withheld because this record is not cleared for redistribution. mask vector —distinct-from→ feature vector mask vector feature vector strength of the relation: 1 · note: RE-GROUNDS another row of this record: the old span named ONE endpoint. Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span compares the feature vectors and mask vectors of two example spikes; distinct-from is loose. Quotation withheld because this record is not cleared for redistribution. t-test t-test Quotation withheld because this record is not cleared for redistribution. two-sample t-test two-sample t-test Quotation withheld because this record is not cleared for redistribution. principal component analysis principal component analysis Quotation withheld because this record is not cleared for redistribution. cross-validation cross-validation Quotation withheld because this record is not cleared for redistribution. standard deviation standard deviation Quotation withheld because this record is not cleared for redistribution. robust estimator (median absolute deviation) robust estimator (median absolute deviation) Quotation withheld because this record is not cleared for redistribution. Bayesian Information Criterion Bayesian Information Criterion Quotation withheld because this record is not cleared for redistribution. maximum likelihood estimation (mixture of Gaussians) maximum likelihood estimation (mixture of Gaussians) Quotation withheld because this record is not cleared for redistribution. false discovery rate / true positive rate false discovery rate / true positive rate Quotation withheld because this record is not cleared for redistribution. Fowlkes-Mallows index (clustering agreement) Fowlkes-Mallows index (clustering agreement) Quotation withheld because this record is not cleared for redistribution. t-test t-test Quotation withheld because this record is not cleared for redistribution. principal component analysis principal component analysis Quotation withheld because this record is not cleared for redistribution. mixture of Gaussians mixture of Gaussians Quotation withheld because this record is not cleared for redistribution. cross-validation cross-validation Quotation withheld because this record is not cleared for redistribution. ROC curve ROC curve Quotation withheld because this record is not cleared for redistribution. Bayesian Information Criterion Bayesian Information Criterion Quotation withheld because this record is not cleared for redistribution. Fowlkes-Mallows index Fowlkes-Mallows index Quotation withheld because this record is not cleared for redistribution. Kullback-Leibler divergence Kullback-Leibler divergence Quotation withheld because this record is not cleared for redistribution. false discovery rate false discovery rate Quotation withheld because this record is not cleared for redistribution. geometric random walk geometric random walk Quotation withheld because this record is not cleared for redistribution. Detection is close to optimal with a strong threshold of 4 SD, a weak threshold of 2 SD and a power weight of 2. Quotation withheld because this record is not cleared for redistribution. The masked EM clustering algorithm reaches roughly 5% false positive and false discovery rates, close to the theoretical optimum. Quotation withheld because this record is not cleared for redistribution. The classical EM algorithm failed on high-count probe data with error rates above 50%, whereas the masked EM algorithm matched the BEER-estimated optimum. Quotation withheld because this record is not cleared for redistribution. Human operators agreed with each other about which clusters to merge but disagreed about which clusters count as well-isolated, so quantitative isolation metrics should replace that judgement. Quotation withheld because this record is not cleared for redistribution. Spacing recording sites further apart than 40 μm substantially degrades sorting performance in rat cortex. Quotation withheld because this record is not cleared for redistribution. The two sides of a staggered shank act as largely independent probes, so the gain from two-sided shanks is more recorded units rather than better isolation of any one unit. Quotation withheld because this record is not cleared for redistribution. Adding time as an extra clustering feature to track amplitude drift made clustering worse rather than better. Quotation withheld because this record is not cleared for redistribution. Manual curation time grows linearly with cluster count at roughly one minute per cluster and is independent of probe size. Quotation withheld because this record is not cleared for redistribution. the masked EM algorithm is statistically indistinguishable from the theoretical optimum, while the classical EM is far from it Quotation withheld because this record is not cleared for redistribution. false positive and false discovery rates are both about 5% Quotation withheld because this record is not cleared for redistribution. experienced operators agree on merges, but disagree on which clusters count as well isolated Quotation withheld because this record is not cleared for redistribution. the authors therefore recommend quantitative isolation metrics in place of the human judgement Quotation withheld because this record is not cleared for redistribution. adding time as a clustering feature to track drift made performance worse, against expectation Quotation withheld because this record is not cleared for redistribution. performance in rat cortex falls substantially once site spacing exceeds the 40um same-side spacing of the test probes Quotation withheld because this record is not cleared for redistribution. the two sides of a staggered shank work independently: the gain is more units, not better isolation Quotation withheld because this record is not cleared for redistribution. detection parameters optimal in cortex were also optimal in CA1 Quotation withheld because this record is not cleared for redistribution. the classical EM failure is not a parameter choice - rerunning across penalty values could not fix it Quotation withheld because this record is not cleared for redistribution. spike sorting spike sorting Quotation withheld because this record is not cleared for redistribution. hybrid spikes hybrid spikes Quotation withheld because this record is not cleared for redistribution. masked EM algorithm masked EM algorithm Quotation withheld because this record is not cleared for redistribution. mask vector mask vector Quotation withheld because this record is not cleared for redistribution. Wizard Wizard Quotation withheld because this record is not cleared for redistribution. spike sorting spike sorting Quotation withheld because this record is not cleared for redistribution. masked EM algorithm masked EM algorithm Quotation withheld because this record is not cleared for redistribution. mask vector mask vector Quotation withheld because this record is not cleared for redistribution. Best Ellipsoid Error Rate Best Ellipsoid Error Rate Quotation withheld because this record is not cleared for redistribution. hybrid datasets hybrid datasets Quotation withheld because this record is not cleared for redistribution. Fowlkes-Mallows index Fowlkes-Mallows index Quotation withheld because this record is not cleared for redistribution. curse of dimensionality curse of dimensionality Quotation withheld because this record is not cleared for redistribution. fraction of hybrid spikes detected fraction of hybrid spikes detected Quotation withheld because this record is not cleared for redistribution. spike timing jitter spike timing jitter Quotation withheld because this record is not cleared for redistribution. mask accuracy mask accuracy Quotation withheld because this record is not cleared for redistribution. site thinning site thinning Quotation withheld because this record is not cleared for redistribution. weak threshold weak threshold Quotation withheld because this record is not cleared for redistribution. strong threshold strong threshold Quotation withheld because this record is not cleared for redistribution. site spacing site spacing Quotation withheld because this record is not cleared for redistribution. penalty parameter penalty parameter Quotation withheld because this record is not cleared for redistribution. 5% — error rate the whole pipeline achieves on validation data across species and brain regions Quotation withheld because this record is not cleared for redistribution. 95% — proportion of ground-truth (hybrid) spikes correctly detected at the chosen parameters, with 0.5-sample timing jitter Quotation withheld because this record is not cleared for redistribution. p = 0.8 — t-test comparing masked EM total error to theoretical optimum (no difference), versus p = 0.005 for masked vs classical EM Quotation withheld because this record is not cleared for redistribution. p=0.75 — two-sample t-test showing a cell that stopped firing halfway through the recording did not change total sorting error Quotation withheld because this record is not cleared for redistribution. below 80% — average correct spike-detection rate once alternating recording sites were removed (site thinning) Quotation withheld because this record is not cleared for redistribution. over 320 scientists in 50 neurophysiology labs — size of the user base whose feedback shaped the software Quotation withheld because this record is not cleared for redistribution. over 10,000% — speed increase of the reimplemented KlustaKwik clustering program Quotation withheld because this record is not cleared for redistribution. 5% — false positive and false discovery rate of the masked EM algorithm on hybrid data Quotation withheld because this record is not cleared for redistribution. over 50% — error rate of the classical EM algorithm on the same data Quotation withheld because this record is not cleared for redistribution. 4 SD and 2 SD — optimal strong and weak detection thresholds, with power weight 2 Quotation withheld because this record is not cleared for redistribution. 95% — spikes correctly detected at those settings, with 0.5 samples of timing jitter Quotation withheld because this record is not cleared for redistribution. below 80% — correct detection after site thinning - the cost of halving electrode density Quotation withheld because this record is not cleared for redistribution. 40 — micron same-side site spacing beyond which performance in rat cortex falls off Quotation withheld because this record is not cleared for redistribution. about 1 minute per cluster — operator time for curation, linear in cluster count and independent of probe size Quotation withheld because this record is not cleared for redistribution. 8 — human operators tested, 5 experienced and 3 novices Quotation withheld because this record is not cleared for redistribution. 6 of 10 — hybrid cells detectable on only one side of a staggered probe Quotation withheld because this record is not cleared for redistribution. over 320 scientists in 50 neurophysiology labs — the user base whose feedback shaped the software Quotation withheld because this record is not cleared for redistribution. 10-20 — simultaneously recorded cells a tetrode can reveal, against one or two for a sharp electrode Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1038/nn.4268.