多通道生物信号间隙填补方案选择及滤波方法技术咨询

Great question—when working with gap-ridden biosignals (like your 16-channel dataset) for PSD calculation via scipy.welch, your gap-filling approach directly dictates the filtering strategy to avoid artifacts. Let’s break down each scenario tailored to your three methods:

1. Direct Concatenation (No Gap Filling, MNE Creates Equal-Interval Structure) #

When you skip gap filling and concatenate signals, MNE enforces the 4ms sampling interval by effectively "compressing" the data. This creates discontinuities (step jumps) at former gap boundaries, which introduce massive high-frequency noise into your PSD.

Recommended Filtering Steps: #

• First, fix discontinuities: Use linear interpolation across concatenation points to smooth step jumps (you can use np.interp or MNE’s mne.preprocessing.interpolate_bads if you mark these points as bad).
• Apply a linear-phase FIR filter: Use MNE’s mne.filter.filter_data to implement:
  • A high-pass filter (e.g., 0.5Hz) to remove baseline shifts caused by concatenation.
  • A low-pass filter (e.g., 50Hz for EEG) to attenuate high-frequency artifacts from the jumps.
• For Welch PSD: Use a Hamming window (reduces spectral leakage) with a window length longer than the largest gap size, plus 50% overlap to improve frequency resolution. Example snippet:

  from scipy.signal import welch
  fs = 250  # 4ms sampling rate = 250Hz
  f, psd = welch(filtered_data, fs=fs, window='hamming', nperseg=256, noverlap=128)
  

2. Gap Filling with np.nan #

np.nan is a clean way to mark missing data, but scipy.welch doesn’t handle NaNs natively, and MNE’s processing needs careful handling to avoid artifacts.

Recommended Filtering Steps: #

• Use MNE’s NaN-aware filtering: MNE’s mne.filter.filter_data automatically skips NaN regions during FIR filtering (it uses edge-effect handling to avoid contaminating valid data). Avoid IIR filters here—they’ll propagate NaN artifacts across the signal.
• Optional: Interpolate NaNs first: If you want continuous data before filtering, replace NaNs with linear or spline interpolation (scipy.interpolate.interp1d or MNE’s mne.preprocessing.interpolate_bads).
• Add notch filtering: For biosignals, remove power-line noise (50/60Hz) with mne.filter.notch_filter to clean up the PSD.

3. Gap Filling with 0s #

Filling gaps with 0s is the riskiest approach—it creates sharp transitions (signal → 0 → signal) that generate broad-spectrum artifacts (from the sinc-shaped frequency response of the rectangular gap fill).

Recommended Filtering Steps: #

• Replace 0-filled regions first: Treat 0s as missing data (mask them with np.nan) and interpolate using linear/spline methods—this eliminates the worst artifacts before filtering.
• Apply strict bandpass filtering: Use a linear-phase FIR bandpass filter (e.g., 0.5–50Hz for EEG) to attenuate high-frequency noise from the 0 transitions. Skip IIR filters here—they can amplify edge artifacts.
• Optimize Welch parameters: Use longer window lengths (e.g., 512 samples) and higher overlap (75%) to average out residual artifacts from any remaining 0-filled segments.

General Best Practices #

• Always visualize your data before/after filtering with MNE’s plotting tools (e.g., mne.plot_raw) to spot artifacts.
• Match filter cutoff frequencies to your signal type (e.g., 1–100Hz for EMG, 0.5–50Hz for EEG).
• For MNE compatibility, convert your data to an mne.io.Raw object after gap handling—this simplifies filtering and preprocessing workflows.

内容的提问来源于stack exchange，提问作者Andrea A.