neurokit2 – Data & Science skill for AI agents

NeuroKit2

Scope and evidence cutoff

Use this skill for method-aware, reproducible biosignal research with NeuroKit2. The

snapshot was checked on 2026-10-01 against:

• stable PyPI 0.2.13, released 2026-03-02;
• Python metadata (>=3.10; classifiers 3.10–3.14) and wheel dependencies;
• GitHub release notes/tags, NEWS.rst, source at tag v0.2.13;
• official API pages/examples (the live site identified itself as

0.2.13.dev214); and

• pinned 0.2.13 runtime signatures and synthetic output schemas.

The current wheel requires pandas<3.0.0. The refreshed core checks use Python 3.13,

NumPy 2.5.3, and pandas 2.3.3; forcing pandas 3 conflicts with upstream metadata.

API fragments below assume caller-provided arrays and imports. Optional MNE source

reconstruction, cvxEDA, EMD, and pyRQA paths are source-reviewed, not runtime-validated

by this refresh. Synthetic checks establish software behavior, not empirical validity.

The live documentation can be ahead of the stable wheel. Prefer the pinned runtime

for reproducible work and name both versions if consulting development docs.

Boundary

NeuroKit2 is a research and educational toolbox. Do not present its output as:

• a diagnosis, treatment recommendation, patient-monitoring decision, or alarm;
• validation, certification, or regulatory evidence for a medical device; or
• proof that a physiological construct is measured validly in a new sensor,

protocol, environment, population, or disease group.

Validate acquisition hardware, electrode/optode placement, units, sampling and clock

accuracy, preprocessing, detector/decomposition method, population, task, and

outcomes for the intended study. Preserve raw data and an auditable exclusion log.

Use deidentified local files only; do not place PHI in prompts, logs, examples, or

bundled fixtures.

Reproducible installation

uv pip install "neurokit2==0.2.13"

For optional features, create a uv project, add only the packages actually required at

reviewed exact versions, and commit/review the resulting uv.lock before

uv sync --locked. NeuroKit2 exposes an upstream full extra, but this skill

intentionally does not install that floating transitive set in an automated workflow.

Optional capabilities can require MNE, cvxopt, Plotly, PyEMD, pyRQA, Pillow, OpenCV,

or file readers. Record the resolved environment with the analysis. Provision any MNE

data/template download as an explicit, checksummed study input. Do not install a moving

development branch for a reproducible study.

Required data contract

Before processing, record:

1. signal identity and sensor/channel configuration;

2. native sampling rate in Hz and physical unit (or explicitly arbitrary_unit);

3. clock, timestamp origin, drift correction, and synchronization evidence;

4. polarity/orientation and acquisition-side filters/gain;

5. missing samples, discontinuities, saturation, flatlines, motion, and annotations;

6. whether event onsets are zero-based sample indices or seconds;

7. planned preprocessing order, methods, parameters, exclusions, and outputs; and

8. participant-level grouping needed to prevent leakage in later statistics.

Never infer units from a column name. Do not silently treat samples as milliseconds,

volts, microsiemens, or arbitrary units.

Core workflow

1. Inspect before transforming

python skills/neurokit2/scripts/inspect_signal.py \
  --input recording.csv --root . --deidentified \
  --columns ECG,RSP,EDA --time-column time_s \
  --units ECG=mV,RSP=a.u.,EDA=uS

The inspector is bounded and emits no row values or paths. Resolve non-monotonic time,

duplicate samples, gaps, non-finite values, flat runs, and sampling-rate disagreement

before filtering.

2. Preserve preprocessing order

Use this default reasoning order, adapting it to the acquisition and cited method:

1. preserve immutable raw signal and annotations;

2. verify time base, units, polarity, clipping, gaps, and artifacts;

3. segment at long gaps; only interpolate short gaps under a declared policy;

4. apply modality-specific cleaning at the native sampling rate;

5. detect peaks/onsets or decompose components;

6. inspect quality outputs and raw overlays;

7. correct peaks only with logged categories and sensitivity checks;

8. derive rates/features;

9. align continuous modalities on a declared common time grid; and

10. map event indices to that grid, epoch, baseline, and analyze.

Do not resample binary markers or peak-index arrays as ordinary continuous signals.

Map their timestamps to the target grid. Filtering and interpolation can create edge

artifacts and false precision; retain masks for padded, missing, and rejected regions.

3. Treat schemas as runtime observations

Return columns depend on NeuroKit2 version, function, method, signal availability, and

analysis mode. Never claim that one column list is universal.

signals, info = nk.ecg_process(ecg, sampling_rate=250)
observed_schema = {
    "columns": list(signals.columns),
    "info_keys": sorted(info),
}

Persist the observed schema with package version, method parameters, sampling rate, and

quality/exclusion summary. Reference files list verified default schemas for 0.2.13,

not guarantees for every method.

Current patterns

ECG, corrected peaks, and duration-aware HRV

In stable 0.2.13, ecg_process() performs cleaning, R-peak detection with

correct_artifacts=True, rate, default averageQRS quality, DWT delineation, and phase.

signals, info = nk.ecg_process(ecg, sampling_rate=250, method="neurokit")
time_hrv = nk.hrv_time(info, sampling_rate=250)

Inspect ECG_R_Peaks_Uncorrected and ECG_fixpeaks_*; a corrected series is not

automatically a valid NN series. For frequency/nonlinear HRV, enforce metric-specific

duration and beat-count requirements. Five minutes is the conventional short-term

reference; ULF is a long-recording measure, and VLF interpretation from short records

is unsafe. Do not interpret LF/HF as a direct sympathovagal balance. PPG pulse-rate

variability is not interchangeable with ECG HRV.

Use the bounded pipeline:

python skills/neurokit2/scripts/ecg_hrv_pipeline.py \
  --synthetic --sampling-rate 250 --duration 300 \
  --domains time,frequency,nonlinear

The helper gates frequency analysis using the first-to-last corrected-peak span,

separately from recording duration. It uses Welch, 4 Hz interval interpolation, and

normalize=False; absolute band powers are in ms². Its numerical gates do not

establish valid NN intervals or adequate data for every returned metric. ULF remains

unsupported by this short-recording workflow; VLF needs separate justification.

EDA with explicit decomposition

The stable default eda_process(method="neurokit") uses high-pass tonic/phasic

decomposition, not cvxEDA. Choose and report decomposition explicitly:

clean = nk.eda_clean(eda, sampling_rate=100, method="neurokit")
components = nk.eda_phasic(clean, sampling_rate=100, method="highpass")
markers, info = nk.eda_peaks(
    components["EDA_Phasic"],
    sampling_rate=100,
    method="neurokit",
    amplitude_min=0.1,
)

For neurokit/kim2004, amplitude_min is relative to the largest detected response;

it is not an absolute microsiemens threshold. Other supported detectors ignore that

argument, which the helper records explicitly. cvxEDA needs optional cvxopt.

python skills/neurokit2/scripts/eda_pipeline.py \
  --synthetic --sampling-rate 100 --duration 60 \
  --phasic-method highpass --peak-method neurokit

Events, epochs, and baseline

events_find() reports zero-based sample onsets; duration/spacing arguments are in

samples. epochs_create() takes epoch limits in seconds.

events = nk.events_find(trigger, threshold=0.5, duration_min=2)
epochs = nk.epochs_create(
    signals,
    events,
    sampling_rate=100,
    epochs_start=-0.2,
    epochs_end=0.8,
    baseline_correction=False,
)

Plan sample-exact windows first:

python skills/neurokit2/scripts/plan_epochs.py \
  --events 1000,2500,4000 --event-unit samples \
  --sampling-rate 100 --recording-samples 5000 \
  --epoch-start -0.2 --epoch-end 0.8 \
  --baseline-start -0.2 --baseline-end 0

In 0.2.13 the epoch slice is end-exclusive, but the generated floating time index

includes epochs_end. Built-in baseline correction subtracts the epoch mean from its

start through t=0; use manual correction for a narrower prespecified baseline.

Boundary epochs are padded and can contain NaN. Decide drop/pad/error before analysis.

RSA and multimodal processing

bio_process() assumes all inputs already share one sampling rate and alignment. It

does not resample, synchronize, estimate drift, or create nested modality dictionaries;

its info output is flat. Unequal lengths are concatenated by index and can introduce

NaN. RSA is added only when synchronized ECG and RSP are present.

Validate a stri

… (full content is in the download package)