drop-bads

name	drop_bads
description	Remove channels marked as bad (flat, noisy, or manually flagged)
layer	L3
group	channel
metadata	{"tags":["operator","channels","bad","reject","quality"],"modalities":["eeg","seeg","ecog","meg"],"step_string":"drop_bads","analysis_goal_allowed":["classification","source_localization","feature_extraction","clinical_screening","exploratory","generic","connectivity","phase_amplitude_coupling","online_inference"],"analysis_goal_forbidden":[]}

Drop Bad Channels

Function

Removes channels from the data array that are marked as bad in meta.bad_channels. Reduces channel count. Irreversible — use interpolate_bads if you want to preserve channel count.

Parameter Format

drop_bads — No parameters. Operates on pre-identified bad channels.

How Bad Channels Are Identified

Bad channels are marked during quality_check or inspect_data and stored in data_dict["meta"]["bad_channels"]. Criteria:

• Flat channels (near-zero variance)
• Excessively noisy channels (variance > 5x median)
• Channels with high artifact ratio
• Manually flagged by researcher

When to Use

• Quality check identified clearly broken channels
• Channels are physically disconnected or have extreme impedance
• You want to reduce dimensionality before spatial filtering (CSP, ICA)
• Channel count is high enough that dropping 1-3 channels is acceptable

When NOT to Use

• No bad channels identified (no-op — returns data unchanged)
• Need to maintain consistent channel count across subjects (use interpolate_bads instead)
• Downstream analysis requires specific channel locations (e.g., source localization)

Constraints

• If no meta.bad_channels list exists, this step does nothing
• Permanently removes channels — cannot be undone without reloading
• Updates data.shape[0] and channels list accordingly

Ordering

• Apply BEFORE: car (bad channels corrupt the average reference)
• Apply BEFORE: ica (bad channels waste ICA components)
• Apply AFTER: quality check has been run to identify bad channels
• Alternative: use interpolate_bads instead if channel preservation needed

Reference Code

Using MNE (on existing Raw object)

raw.drop_channels(['Fp1', 'Fp2'])

Standalone (numpy, from meta.bad_channels)

import numpy as np

bad_channels = ['Fp1', 'Fp2']  # from meta.bad_channels
keep_idx = [i for i, ch in enumerate(channels) if ch not in bad_channels]
data = data[keep_idx]
channels = [channels[i] for i in keep_idx]

Key API

• raw.drop_channels(bad_list) — MNE method
• Operates on data_dict["meta"]["bad_channels"] in pipeline context

drop_bads:auto (Phase 1+)

When the param is the literal string auto, the operator detects bad channels automatically:

Criterion	Threshold	Notes
flat	std < 1e-7	sensor disconnected / amplifier rail
high variance	std > 3 × median(non-zero σ)	usually muscle / electrical artifact
spike	`max(	x

Detected names are written to meta['bad_channels'] and dropped from data + channels arrays. No interpolation is performed — downstream goals that need every channel filled in must explicitly add interpolate_bads after this step.

The trailing dropped_channels list in meta is what qc.py and reasoning.md surface to the user.

Goal-conditional auto-injection (codegen/generator.py:_enforce_clean_output):

analysis_goal	drop_bads:auto auto-inserted?
classification	yes
feature_extraction	yes
clinical_screening	yes
generic	yes
source_localization	no
exploratory	no

For opted-out goals, manually add drop_bads:auto or interpolate_bads if you need them.