nimbus-bci 0.4.2

Nimbus BCI: Bayesian classifiers for brain-computer interfaces

nimbus-bci

Bayesian BCI classifiers with sklearn compatibility, streaming inference, low-cost online updates (partial_fit vs full refit), measured scoring time via predict_batch · latency_ms, active-learning calibration loops, and rich diagnostics.

PyPI Python License

Documentation in this repo: Why Nimbus? (vs sklearn / pyRiemann) · NimbusBench — main conclusions · Latency (NimbusBench + in-SDK scope) · Trust, calibration, and rejection · Active-learning calibration loops · Notebooks index. Hosted docs: docs.nimbusbci.com.

Features

• Four sklearn-compatible classifiers: three static Bayesian decoders — LDA, QDA, Softmax (Polya–Gamma) — plus NimbusSTS for latent-state / non-stationary settings (EKF-style updates, experimental)
• sklearn-compatible API: Works with pipelines, cross-validation, and GridSearchCV
• Streaming inference: Real-time chunk-by-chunk processing; per-call predict_batch → BatchResult.latency_ms
• Online update cost: Conjugate / variational partial_fit on Nimbus heads — measured vs batch refit in NimbusBench (main conclusions, latency scope)
• Active learning: suggest_next_trial (BALD on LDA/QDA/Softmax), should_query streaming gate, and label-free calibration_sufficient stopping — cut cued calibration time without manual heuristics
• Rich diagnostics: Entropy, Mahalanobis distance, calibration metrics (ECE/MCE)
• Online learning: Update models with new data without retraining
• BCI-specific utilities: ITR calculation, temporal aggregation, quality assessment
• MNE-Python integration: Convert between MNE Epochs and Nimbus data formats

Installation

pip install nimbus-bci

To use the optional JAX-based softmax model:

pip install nimbus-bci[softmax]

From source:

git clone https://github.com/nimbusbci/nimbuspysdk.git
cd nimbuspysdk
pip install -e ".[all]"

Quick Start

sklearn-Compatible API (Recommended)

from nimbus_bci import NimbusLDA, NimbusQDA, NimbusSoftmax, NimbusSTS
import numpy as np

# Create and fit classifier
clf = NimbusLDA()
clf.fit(X_train, y_train)

# Predict
predictions = clf.predict(X_test)
probabilities = clf.predict_proba(X_test)

# Online learning
clf.partial_fit(X_new, y_new)

Works with sklearn Pipelines

from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import cross_val_score, GridSearchCV

# Simple pipeline
pipe = make_pipeline(StandardScaler(), NimbusLDA())
pipe.fit(X_train, y_train)

# Cross-validation
scores = cross_val_score(NimbusLDA(), X, y, cv=5)
print(f"Accuracy: {scores.mean():.2%} (+/- {scores.std():.2%})")

# Hyperparameter tuning
param_grid = {'mu_scale': [1.0, 3.0, 5.0], 'class_prior_alpha': [0.5, 1.0]}
grid = GridSearchCV(NimbusLDA(), param_grid, cv=5)
grid.fit(X, y)
print(f"Best params: {grid.best_params_}")

Streaming Inference (Real-Time BCI)

from nimbus_bci import NimbusLDA, StreamingSession
from nimbus_bci.data import BCIMetadata

# Setup
metadata = BCIMetadata(
    sampling_rate=250.0,
    paradigm="motor_imagery",
    feature_type="csp",
    n_features=16,
    n_classes=4,
    chunk_size=125,  # 500ms chunks
    temporal_aggregation="logvar",
)

# Train model
clf = NimbusLDA()
clf.fit(X_train, y_train)

# Create streaming session
session = StreamingSession(clf.model_, metadata)

# Process chunks in real-time
for chunk in eeg_stream:
    result = session.process_chunk(chunk)
    print(f"Chunk prediction: {result.prediction} ({result.confidence:.2%})")

# Finalize trial with aggregation
final = session.finalize_trial(method="weighted_vote")
print(f"Final: class {final.prediction} (entropy: {final.entropy:.2f} bits)")

For NimbusSTS specifically (stateful latent dynamics), use StreamingSessionSTS so the latent state can be propagated and updated with delayed feedback:

from nimbus_bci import NimbusSTS
from nimbus_bci.inference import StreamingSessionSTS
from nimbus_bci.data import BCIMetadata

metadata = BCIMetadata(
    sampling_rate=250.0,
    paradigm="motor_imagery",
    feature_type="csp",
    n_features=16,
    n_classes=2,
    chunk_size=125,
    temporal_aggregation="mean",
)

clf = NimbusSTS().fit(X_train, y_train)
session = StreamingSessionSTS(clf, metadata)

result = session.process_chunk(chunk)  # propagates state by default
session.provide_feedback(label=0)      # when label arrives later

Active Learning (Calibration Loop)

Cut cued-calibration time by labeling only the trials the model is genuinely uncertain about, and stop automatically when the posterior settles:

from nimbus_bci import NimbusLDA
from nimbus_bci.active_learning import (
    suggest_next_trial,
    calibration_sufficient,
)

clf = NimbusLDA().fit(X_seed, y_seed)   # small initial cued batch
prev = clf.get_model()

for _ in range(max_rounds):
    # Rank the unlabeled pool by BALD informativeness, label the top 4.
    ranked = suggest_next_trial(
        clf, X_pool, strategy="bald", n=4, num_posterior_samples=64,
    )
    X_new, y_new = collect_labels_for(ranked.indices)   # cue + record
    clf.partial_fit(X_new, y_new)

    # Label-free stopping: when predict_proba over the pool stops moving,
    # more cues will not change predictions much.
    status = calibration_sufficient(
        clf, X_pool,
        criterion="posterior_stability",
        previous=prev, threshold=0.02,
    )
    if status.is_sufficient:
        break
    prev = clf.get_model()

Strategies (entropy, margin, least_confidence, bald) and stopping criteria (posterior_stability, expected_info_gain) are all model-agnostic. STS gets posterior_stability for free; BALD-based features on STS are deferred to v1.1. Full recipe in docs/active_learning.md.

Batch Inference with Diagnostics

from nimbus_bci import predict_batch
from nimbus_bci.data import BCIData, BCIMetadata

# Create BCI data container
metadata = BCIMetadata(
    sampling_rate=250.0,
    paradigm="motor_imagery",
    feature_type="csp",
    n_features=16,
    n_classes=4,
)
data = BCIData(features, metadata, labels)

# Run batch inference with full diagnostics
result = predict_batch(model, data)

print(f"Mean entropy: {result.mean_entropy:.2f} bits")
print(f"Balance: {result.balance:.2%}")
if result.calibration is not None:
    print(f"ECE: {result.calibration.ece:.3f}")
print(f"Latency: {result.latency_ms:.1f}ms")

latency_ms is the SDK’s measured wall time for this predict_batch call (not your full acquisition→feature pipeline). Online update cost vs batch refit is in NimbusBench (main conclusions, Latency & performance).

MNE-Python Integration

import mne
from nimbus_bci import NimbusLDA
from nimbus_bci.compat import from_mne_epochs, extract_csp_features

# Load and preprocess with MNE
raw = mne.io.read_raw_gdf("motor_imagery.gdf")
events = mne.find_events(raw)
epochs = mne.Epochs(raw, events, tmin=0, tmax=4, baseline=None, preload=True)
epochs.filter(8, 30)  # Mu + Beta bands

# Extract CSP features
csp_features, csp = extract_csp_features(epochs, n_components=8)

# Train Nimbus classifier
clf = NimbusLDA()
clf.fit(csp_features, epochs.events[:, 2])

Available Classifiers

Classifier	Description	Best For
NimbusLDA	Bayesian LDA with shared covariance	Fast, when classes have similar shapes
NimbusQDA	Bayesian QDA with class-specific covariances	Complex class distributions
NimbusSoftmax	Bayesian logistic regression (Polya-Gamma VI)	Non-Gaussian decision boundaries
NimbusSTS	Structural time series classifier (latent state + EKF-style inference)	Non-stationary settings, drifting class boundaries (experimental)

Choosing the Right Classifier

Quick Decision Guide

Is your data stationary (distributions don't change over time)?

• Yes → Use static models (LDA/QDA/Softmax)
• No → Use NimbusSTS for temporal adaptation

For stationary data:

• Classes have similar covariance? → NimbusLDA (fastest)
• Classes have different shapes? → NimbusQDA
• Non-Gaussian boundaries? → NimbusSoftmax

For non-stationary data:

• Gradual drift (fatigue, electrode shift)? → NimbusSTS
• Multi-day sessions with state transfer? → NimbusSTS
• Delayed feedback paradigms? → NimbusSTS

Detailed Comparison

Scenario	Recommended Model	Why?
Stable offline datasets	NimbusLDA	Fastest, closed-form solution
P300 spelling (stable)	NimbusLDA or NimbusQDA	Event-related, stationary
SSVEP	NimbusLDA	Highly stationary frequency response
Motor Imagery (short sessions)	NimbusLDA or NimbusQDA	Stationary within session
Motor Imagery (long sessions, fatigue)	NimbusSTS	Tracks drift due to fatigue
Multi-day experiments	NimbusSTS	State transfer across sessions
Electrode repositioning	NimbusSTS	Adapts to impedance changes
Closed-loop with delayed feedback	NimbusSTS	Explicit state propagation
Asynchronous BCI (idle vs active)	NimbusSTS	Models engagement state
Neurofeedback training	NimbusSTS	Tracks learning-induced changes
Long calibration sessions, want to cut label cost	Any head + suggest_next_trial(strategy="bald")	Pool-based BALD on the conjugate posterior; LDA/QDA/Softmax in v1
Don't know when to stop calibrating	Any head + calibration_sufficient	Label-free posterior_stability works for STS too

NimbusSTS Example (Temporal Adaptation)

from nimbus_bci import NimbusSTS

# Train on calibration data
clf = NimbusSTS(transition_cov=0.05, num_steps=50)
clf.fit(X_calibration, y_calibration)

# Online session with delayed feedback
for x_trial, y_feedback in online_trials:
    # 1. Propagate state forward (no label needed)
    clf.propagate_state()
    
    # 2. Make prediction
    prediction = clf.predict(x_trial)
    
    # ... user performs action, feedback arrives later ...
    
    # 3. Update with true label
    clf.partial_fit(x_trial, y_feedback)

# Multi-day state transfer
z_day1, P_day1 = clf.get_latent_state()

# Day 2: Initialize with Day 1 state (increased uncertainty)
clf_day2 = NimbusSTS()
clf_day2.fit(X_day2_calib, y_day2_calib)
clf_day2.set_latent_state(z_day1 * 0.5, P_day1 * 2.0)

Label Conventions (Important)

Nimbus supports common EEG/BCI labeling patterns:

• BCIData labels: can be any non-negative integer codes (e.g., MNE event IDs like 769/770), as long as the number of unique labels does not exceed BCIMetadata.n_classes.
• sklearn estimators (NimbusLDA, NimbusQDA, NimbusSoftmax, NimbusSTS):
  • fit() learns classes_ from your provided labels.
  • predict() returns labels in the original label space (elements of classes_).
• Model-snapshot inference (NimbusModel + predict_batch / StreamingSession):
  • predictions are returned in the model’s label_base convention (label_base is stored in model.params).
  • use nimbus_bci.data.labels_to_zero_indexed(...) for metrics/aggregation that require 0-indexed labels.

NimbusSTS Sequence Semantics (Important)

NimbusSTS has a latent state. For correctness and sklearn compatibility:

• NimbusSTS.predict_proba(X) treats rows as conditionally independent by default.
• For time-ordered evaluation, propagate explicitly:
  • call clf.propagate_state() between trials/chunks, or
  • use the functional API nimbus_sts_predict_proba(model, X, evolve_state=True) when X rows are ordered in time.

Metrics & Diagnostics

from nimbus_bci import (
    compute_entropy,            # Prediction uncertainty
    compute_calibration_metrics,  # ECE, MCE
    calculate_itr,              # Information Transfer Rate
    assess_trial_quality,       # Quality checks
)

# Entropy (uncertainty)
entropy = compute_entropy(posterior)  # bits

# Calibration
calib = compute_calibration_metrics(predictions, confidences, labels)
print(f"ECE: {calib.ece:.3f}, MCE: {calib.mce:.3f}")

# ITR
itr = calculate_itr(accuracy=0.85, n_classes=4, trial_duration=4.0)
print(f"ITR: {itr:.1f} bits/min")

Normalization

Critical for cross-session BCI performance:

from nimbus_bci import estimate_normalization_params, apply_normalization

# Estimate from training data
params = estimate_normalization_params(X_train, method="zscore")

# Apply to all data
X_train_norm = apply_normalization(X_train, params)
X_test_norm = apply_normalization(X_test, params)  # Same params!

Benchmarks

Reproducible MOABB runs live in the separately installable [nimbusbench/](nimbusbench/README.md) package (install, CLI, checked-in CSVs, benchmark_summary.md).

Read in order: Main conclusions (what to take away) → Pinned headline table (numbers + CSV paths) → Limitations. Statistics: benchmark_preregistration.md. What each model track may claim: benchmark_claims.md.

At a glance (IV-2b / Lee2019 unless noted)

• S3: partial_fit matches Nimbus batch refit in eval accuracy; mean head-only update time ~8–10× lower than sklearn batch refit on the same stream (see table + latency for E2E scope).
• S4: Report effective ITR together with accept rate and accuracy on accepted (LOFO / non-oracle).
• S5: Lee2019 — positive mean lift for partial_fit vs static on accuracy and effective ITR; Physionet is supporting; report heterogeneity where summaries include it.
• S1 / S2: Preregistered small-label and tail metrics only—not a universal “beats sklearn” claim.

Regenerate numbers with python -m nimbusbench summarize --input … next to the CSVs; do not copy stale figures from prose.

Quick demo (no MOABB download): notebooks/s3_update_latency_head_vs_sklearn.ipynb - same-accuracy partial_fit vs Nimbus refit + ~8–10× head update vs sklearn refit from the checked-in S3 CSV.

Project Structure

nimbus_bci/
├── models/              # Classifiers
│   ├── nimbus_lda/     # LDA (shared covariance)
│   ├── nimbus_qda/     # QDA (class-specific covariances)
│   └── nimbus_softmax/ # Softmax (Polya-Gamma)
├── data/               # Data contracts (BCIData, BCIMetadata)
├── inference/          # Batch and streaming inference
├── metrics/            # Diagnostics, calibration, ITR
├── utils/              # Normalization, aggregation
└── compat/             # sklearn/MNE compatibility

Functional API (Backward Compatible)

The original functional API is still available:

from nimbus_bci import (
    nimbus_lda_fit, nimbus_lda_predict, nimbus_lda_update,
    nimbus_qda_fit, nimbus_qda_predict,
    nimbus_softmax_fit, nimbus_softmax_predict,
    nimbus_save, nimbus_load,
)

# Fit model
model = nimbus_lda_fit(X, y, n_classes=4, label_base=0, ...)

# Predict
probs = nimbus_lda_predict_proba(model, X_test)

# Update (online learning)
model = nimbus_lda_update(model, X_new, y_new)

# Save/load
nimbus_save(model, "model.npz")
model = nimbus_load("model.npz")

# Legacy trusted artifacts that contain object-serialized params
# require explicit opt-in:
legacy_model = nimbus_load("legacy-model.npz", trusted=True)

Testing

pip install -e ".[dev]"
pytest -v

Requirements

Core (installed with pip install nimbus-bci):

• Python ≥ 3.11
• NumPy ≥ 1.26
• scikit-learn ≥ 1.4
• SciPy ≥ 1.17.1

Optional extras:

• JAX ≥ 0.4.25 — required for NimbusSoftmax and the softmax functional API (pip install nimbus-bci[softmax])
• MNE ≥ 1.6 — EEG integration (pip install nimbus-bci[mne])
• matplotlib ≥ 3.8 — visualization (pip install nimbus-bci[viz])
• all installs all optional extras above (pip install nimbus-bci[all])

License

This Software is proprietary — Nimbus BCI Inc. retains all rights — and is licensed under two tracks:

1. No-cost Non-Commercial License (self-executing, no registration required). Anyone may install, use, modify, and redistribute the Software (and their own derivative works) for non-commercial research, scholarship, teaching, or personal learning — including at universities, government labs, and non-profit research organizations. Modifications and non-commercial distribution (e.g., forks, public reproducibility archives) are permitted, provided derivative works keep this LICENSE.txt and link back to nimbusbci.com. A non-commercial patent grant is included. Citation is strongly requested (it's how a small team sustains no-cost academic access) but is not a legal condition — you don't lose your license by forgetting a BibTeX entry. See CITATION.cff.

2. Paid Commercial License. Any use inside a for-profit company, in a paid product or service, for sponsored/fee-for-service research, or for internal business operations requires a separate written commercial license. Tiers below.

Tier	Use Case
Startup	Companies < $1M revenue
Commercial	Full production rights
Enterprise	Unlimited deployments + SLA
OEM/Embedded	Medical devices, FDA support, white-label redistribution

Full terms: LICENSE.txt. If you are unsure which track applies to your use case, contact [e-mail].

How to cite

If you use the Nimbus BCI Python SDK in your research, please cite it (see CITATION.cff):

@software{nimbus_bci_pysdk,
  author       = {{Nimbus BCI Inc.}},
  title        = {Nimbus BCI Python SDK: Bayesian classifiers for brain-computer interfaces},
  year         = {2026},
  version      = {0.4.2},
  url          = {https://nimbusbci.com},
  note         = {Replace the version above with the version you installed.}
}

Request a commercial license

1. Email [e-mail] with your use case
2. Receive a license agreement and API key (if applicable)
3. Install and start building

Website: https://nimbusbci.com

---

© 2024-2026 Nimbus BCI Inc. — The AI Engine for Brain-Computer Interfaces.

License: Proprietary with a no-cost Non-Commercial License for academic and non-commercial use (citation requested, not required). See License below and LICENSE.txt.