zpybci 0.4.0

High-performance signal processing for BCI research

zpybci -- Zerostone Python for BCI

High-performance, real-time signal processing for brain-computer interface research. Powered by Rust with zero-copy NumPy integration.

Installation

pip install zpybci

Wheels are available for Linux (x86_64, aarch64), macOS (Intel + Apple Silicon), and Windows (x86_64). Python 3.8+.

Quick Start

import numpy as np
import zpybci as zbci

# Bandpass filter for alpha band (8-12 Hz)
bpf = zbci.IirFilter.butterworth_bandpass(sample_rate=256.0, low_cutoff=8.0, high_cutoff=12.0)
signal = np.random.randn(1000).astype(np.float32)
filtered = bpf.process(signal)

# ICA for artifact removal
ica = zbci.Ica(channels=16, contrast="logcosh")
ica.fit(eeg_data, max_iter=200)
cleaned = ica.remove_components(eeg_data, exclude=[0, 2])  # remove blink/muscle artifacts

# Load an EDF file -- no MNE dependency needed
rec = zbci.read_edf("recording.edf")
ch1 = rec.get_channel("Fp1")
all_data = rec.get_all_channels()  # (n_channels, n_samples) numpy array

Validated Results

Tested on the BCI Competition IV 2a motor imagery benchmark (9 subjects, 4-class, session-to-session transfer):

Pipeline	Mean Accuracy	Published Baseline
CSP+LDA	40.5%	~40-50%
TS+LDA	64.4%	~60-68%
MDM	59.0%	~55-62%
xDAWN+MDM	57.4%	~55-60%

TS+LDA exceeds the original competition winner (FBCSP, ~63%). All pipelines built entirely with zpybci primitives.

Features

Filters

• IIR -- Butterworth lowpass, highpass, bandpass (order 2/4/6/8, proper pole placement matching scipy)
• FIR -- arbitrary-length finite impulse response
• AC coupling -- DC removal for streaming data
• Median -- nonlinear smoothing
• Adaptive -- LMS and NLMS for noise cancellation
• Notch -- narrowband rejection (e.g., 50/60 Hz line noise)

Spatial Filters

• CAR -- common average reference
• Surface Laplacian -- current source density approximation
• Channel Router -- flexible channel remapping
• xDAWN -- supervised spatial filters maximizing ERP signal-to-noise ratio

Spectral Analysis

• FFT -- fast Fourier transform (magnitude/phase)
• STFT -- short-time Fourier transform
• Multi-band power -- concurrent power in multiple frequency bands
• Welch PSD -- power spectral density estimation
• CWT -- continuous wavelet transform (Morlet)

Independent Component Analysis

• FastICA -- symmetric parallel extraction with LogCosh, Exp, and Cube contrast functions
• Artifact removal -- remove blink/muscle/cardiac components and reconstruct clean signal
• Channel counts -- 4, 8, 16, 32, or 64 channels

Kalman Filter

• State estimation -- predict/update cycle for real-time decoder smoothing
• Joseph form -- numerically stable covariance update (preserves positive definiteness)
• Flexible dimensions -- 8 state/observation combinations from (2,1) to (8,8)

Linear Discriminant Analysis

• Fisher's LDA -- binary classification with shrinkage regularization
• Calibrated probabilities -- predict_proba() via sigmoid scaling
• Feature dimensions -- 2, 4, 6, 8, 12, 16, 32, or 64

Spike Sorting

• Waveform extraction -- extract spike waveforms around detected events
• Waveform PCA -- dimensionality reduction for spike clustering
• Template matching -- classify spikes by Euclidean distance or normalized cross-correlation
• MAD noise estimation -- robust noise floor estimation for threshold setting
• Batch detection -- negative-threshold crossing with refractory period
• Convenience pipeline -- spike_sort() for end-to-end spike sorting with inline k-means

EDF/EDF+ File Reader

• read_edf() -- load EDF/EDF+ files without MNE or pyedflib
• Channel access -- by index or label name
• Physical units -- automatic digital-to-physical conversion
• Mixed sample rates -- zero-padded multi-channel extraction

Detection

• Threshold -- fixed-threshold event detection
• Adaptive threshold -- self-adjusting threshold based on signal statistics
• Zero-crossing -- rate estimation

Artifact Handling

• Amplitude-based -- flag samples exceeding a threshold
• Z-score -- flag statistically outlying segments

Analysis

• Envelope follower -- instantaneous amplitude via rectification + smoothing
• Windowed RMS -- streaming root-mean-square
• Hilbert transform -- analytic signal, instantaneous phase/frequency

Statistics

• Online mean/variance -- Welford's algorithm, no buffer required
• Online covariance -- streaming covariance matrix
• Connectivity -- coherence and phase locking value (PLV)

BCI Paradigms

• Motor imagery -- CSP with online adaptation, sklearn-compatible transformer
• SSVEP -- CCA-based frequency detection, reference signal generation
• P300/ERP -- epoch averaging, xDAWN spatial filters

Riemannian Geometry

• Tangent space -- SPD manifold projection for classification
• MDM classifier -- minimum distance to mean on SPD manifold
• Frechet mean -- geometric mean of SPD matrices
• Riemannian distance -- affine-invariant distance metric
• Recentering -- domain adaptation via Riemannian transport

Advanced

• OASIS deconvolution -- calcium transient inference from fluorescence traces

Utilities

• Pipeline -- declarative stage chaining with a single process() call
• Resampling -- integer decimation and interpolation
• Streaming percentile -- approximate quantiles on unbounded streams
• Clock sync -- offset estimation, linear drift correction, sample clock alignment
• Cross-correlation -- full, valid, and circular modes
• Window functions -- Hann, Hamming, Blackman, flat-top, Kaiser

Version

0.4.0

License

GPL-3.0