Signal Pipeline Architecture

This document describes the internal data flow of the Synaptiq decode engine, from raw EEG electrode voltages to motor-imagery class labels. Understanding this architecture helps you tune parameters for your specific electrode configuration and clinical context.

Architecture overview

Input format

The SDK accepts EEG data as a multi-channel time series. The canonical shape is [n_channels, n_samples] as a float32 array, in microvolts (μV). Channel ordering must match the montage passed at session initialization.

• Epoch length: 2–6 seconds (default: 4.0s = 2000 samples at 500Hz)
• Epoch overlap: configurable 0–50% (default: 25%)
• Required channels: C3, Cz, C4 at minimum; full perimotor montage preferred (FC3, FCz, FC4, C3, Cz, C4, CP3, CPz, CP4)
• Amplitude range: ±500μV (beyond this triggers artifact flag)

Preprocessing parameters

Preprocessing is configurable via session.set_preprocess_config(). Defaults are optimized for 16-channel gel-cap recordings at 500Hz.

session.set_preprocess_config({
    "bandpass_low_hz": 8.0,    # mu band low cutoff
    "bandpass_high_hz": 30.0,   # beta band high cutoff
    "filter_order": 4,           # Butterworth filter order
    "artifact_threshold_uv": 150,  # peak-to-peak artifact reject
    "eog_channels": ["Fp1", "Fp2"], # set None to disable EOG rejection
    "csp_n_filters": 2,           # CSP filter pairs per frequency band
    "reference": "car"            # 'car' | 'linked-mastoid' | 'none'
})

Feature extraction

After spatial filtering by the CSP bank, the SDK computes the sample covariance matrix of each epoch's filtered data. For a 4-second epoch at 500Hz, this yields a 16×16 SPD matrix per filter pair. Euclidean alignment (EA) is then applied: the session's mean covariance is computed and each epoch's matrix is divided by the square root of the mean — re-centering the distribution to the identity matrix neighborhood.

Classifier architecture

The minimum distance to Riemannian mean (MDM) classifier:

1. During calibration, compute the Riemannian mean of each class's covariance matrices.
2. At decode time, compute the Riemannian distance from the test epoch's covariance to each class mean.
3. Assign the epoch to the nearest class. Confidence is derived from the normalized distance ratio.

Command output format

Each decode event produces a Command object:

{
  "label": "left_hand",   # class label string
  "class_id": 0,          # zero-indexed class integer
  "confidence": 0.91,    # 0.0–1.0, use >0.7 for device output
  "latency_ms": 72.3,    # EEG sample → command dispatch latency
  "epoch_onset_ts": 1741852928.44,  # Unix timestamp
  "artifact_flagged": false  # true if epoch had artifact (still decoded)
}