Quality metrics and screening

Targets SpikeInterface 0.105.0; reviewed 2026-10-01. An apparently clean unit is not proof of a single neuron. Inspect waveforms, refractory behavior, stability, spike counts, artifacts and neighboring units together.

Compute metrics with their dependencies

import spikeinterface.full as si

analyzer = si.create_sorting_analyzer(sorting, recording, sparse=True)
analyzer.compute("random_spikes", max_spikes_per_unit=500, seed=42)
analyzer.compute("waveforms", ms_before=1.0, ms_after=2.0)
analyzer.compute("templates", operators=["average", "std"])
analyzer.compute("noise_levels")
analyzer.compute("spike_amplitudes")
analyzer.compute("amplitude_scalings")
analyzer.compute("quality_metrics",
    metric_names=["firing_rate", "snr", "isi_violation", "presence_ratio", "amplitude_cutoff"],
    metric_params={"isi_violation": {"isi_threshold_ms": 1.5, "min_isi_ms": 0.0},
                   "presence_ratio": {"bin_duration_s": 60.0}},
)
qm = analyzer.get_extension("quality_metrics").get_data()

The default presence bin is 60 s: shorter recordings return NaN. Short smoke fixtures may use shorter bins, but this changes the scientific quantity. An amplitude cutoff can also be NaN when too few spikes support the histogram. Never fill these values with a passing score. The bundled classifiers leave any missing/nonfinite required metric unsorted.

Metric names differ from result columns

Pass the left column to metric_names; query the returned right-hand columns. Inspect si.get_quality_metric_list() and si.get_default_quality_metrics_params() for the installed release. Version 0.105 uses nested metric_params.

analyzer.compute("principal_components", n_components=3, mode="by_channel_local")
analyzer.compute("spike_locations", method="center_of_mass")
analyzer.compute("quality_metrics", metric_names=["mahalanobis", "nearest_neighbor", "drift"],
                 metric_params={"nearest_neighbor": {"n_neighbors": 4}}, n_jobs=1)

NaN or inf from sparse clusters, rank-deficient PCA, too few spikes, or too little recording is a reason to inspect the data, not a threshold to relax automatically. Do not compute every expensive metric by default.

Interpret the measurements

• ISI violation ratio is rate-normalized, not the fraction of all spikes that violate a refractory period and not a percentage contamination estimate. It can exceed 1. The chosen refractory/censor windows depend on physiology and sorter duplicate handling; 1.5 ms is a convention, not a universal neuronal constant.
• Presence ratio is the fraction of bins meeting an activity criterion. It depends on bin size, rate and duration; absence may be real stimulus/state dependence.
• Amplitude cutoff estimates missing spikes under distribution assumptions. Drift, multimodality and censoring can make it misleading. It is not a direct recall measurement.
• SNR is template amplitude divided by estimated channel noise; it does not measure isolation by itself. Use matched physical scaling and preprocessing.
• Drift from sorted spikes can be confounded by activity changes and incorrect assignments. Spatial extent of different neurons is not temporal drift.
• Waveform shape alone does not establish molecular cell type or somatic origin.

Institutional criteria versus local presets

The AllenSDK Visual Coding tutorial documents default filtering at amplitude_cutoff < 0.1, presence_ratio > 0.9, isi_violations < 0.5. SpikeInterface names the last output isi_violations_ratio. Allen data releases and pipelines can use different metrics/parameters; state exactly what was used.

The bundled allen preset adds SNR > 3 and is therefore Allen-inspired. The legacy ibl name is a local tighter screen, not the IBL classifier. IBL's brainbox.metrics.single_units.quick_unit_metrics uses sliding-RP, noise-cutoff and amplitude criteria; its pass indicator and contamination measures are not interchangeable with SpikeInterface's Hill ISI ratio.

All three bundled entry points use the same screening behavior:

Boundaries fail strict inequalities. Finite SNR < 1.5 is flagged noise; excess ISI ratio is flagged mua (strict preset uses > 0.05 for this flag); otherwise failed screens remain unsorted, including low completeness without contamination. These are screening labels to review, not validated diagnoses.

# Run from the skill root so bundled helpers are importable.
import sys
sys.path.insert(0, "scripts")
from compute_metrics import curate_units
labels = curate_units(qm, method="allen")
good_ids = [uid for uid, label in labels.items() if label == "good"]

Keep all metrics and labels, selected IDs, parameter values, package versions, recording duration, and a review rationale. Never discard the uncurated result.

Sources: SpikeInterface API, AllenSDK quality metrics, IBL single-unit metrics.