I’d like to discuss how to handle QC for EEG data, especially when dealing with datasets of varying lengths. Here’s a simplified example of the types of data I’m working with:
• Data A: One long session of EEG data (~1 hour, unlabeled, resting state) (can be as long as 24 hours or more).
• Data B: Multiple short sessions (~5 minutes each, labeled, 20 sessions total).
When performing bad channel/epoch removal, I see three potential approaches:
-
Sub-session-based QA/QC: Split the data into smaller chunks (e.g., 1-minute segments) and perform bad channel/epoch removal at this level.
• Pros: Handles long datasets like Data A by breaking them into manageable chunks.
• Cons: If a bad event only occurs during a portion of the chunk (e.g., the last 10 seconds of a 1-minute segment), the segment might still be kept despite containing bad data. Additionally, the detection algorithm might lack sufficient data within the short segment to make robust decisions. -
Session-based QA/QC: Perform QA/QC on the entire session (e.g., remove a session if it contains bad epochs).
• Pros: Works well for shorter sessions like Data B.
• Cons: For very long sessions like Data A, this approach could result in removing an entire hour of data if a single bad epoch is detected. -
Event-specific QA/QC (does this exist?): Instead of working at the sub-session or session level, implement a pipeline that identifies and removes specific bad epochs, pinpointing the exact time intervals that are problematic.
Because my data is very long (10k hours in total), manual QC is impossible and hence automation seems required. I have looked around a bit, but all current methods seem to either use sub-session or session-level bad channel/epoch detection, which brings questions I mentioned above.
If possible could someone tell to me what the best practices (or convention) for performing QA/QC on very long sessions (~1 hour or more) without discarding entire sessions unnecessarily?**