ProjectTwo weekends plus the IRB$0 beyond the amplifierTier 2Needs a lab

Project D: Your own ERPs

P300, the face-sensitive N170, and the error-related negativity, recorded from friends under a lab's protocol with PsychoPy and LSL. Learning to run a clean human experiment is a skill most engineers never acquire.

AssumesLab Streaming Layer and PsychoPy Designing an experiment that can fail SpineDecoding / signal processing / MLClinical / regulatory / human factors

On this pageTwo weekends plus the IRB

Three experiments
Doing it cleanly
Analysis

An event-related potential is the brain’s stereotyped electrical response to an event, invisible in one trial and clear after averaging fifty. Three classic ones, each a fifteen-minute experiment: the P300 to a rare target, the N170 to a face, and the error-related negativity when you press the wrong button. Recording them on other people, under a real protocol, with consent, triggers, and a saved dataset, is the experience that turns an engineer who can build an amplifier into one who can be trusted with participants.

How the ERP emergesSimulated signal

This interactive needs JavaScript. If you are reading a printout, the caption describes what it shows.

Figure 1. The averaging principle again. Every ERP experiment is this figure, with a different stimulus.

This project requires a lab

Recording other people for anything you will analyze, present, or publish requires an IRB-approved protocol with a principal investigator. Your lab almost certainly has a standing EEG protocol you can be added to. You need CITI training complete, a consent script, and a data storage plan with no names in the files. Recording yourself alone needs none of this; recording a friend “informally” and using the data is the thing the IRB exists to prevent. Do it properly.

Three experiments

P300, oddball. A stream of tones or shapes, 80 percent one kind, 20 percent another. The participant counts the rare ones. Two hundred trials, ten minutes. Average rare versus frequent at Pz: a positive peak of 5 to 15 µV at 300 to 400 ms for the rare ones. The oddball is the fruit fly of ERP research.

N170, faces. Grayscale photographs of faces and of houses, matched for size and contrast, each shown for 300 ms, participant presses a key for an occasional target. Sixty of each. Average faces versus houses at P8 (right, back, low): a negative peak around 170 ms, larger for faces by a few microvolts. The most robust cognitive ERP there is, and it says something about a brain region specialized for faces.

ERN, flanker task. Arrows: press left for a left-pointing centre arrow surrounded by distractors. Fast responses, forced errors. On error trials, average time-locked to the button press at FCz: a sharp negative wave at 50 to 100 ms after the press, before the person consciously knows they erred. You need enough errors, so push for speed.

Doing it cleanly

Fixed trial timing with jittered inter-stimulus intervals, so nothing periodic lines up with the heartbeat or alpha. Markers via LSL at the flip. Photodiode check of screen delay once. Blink breaks every forty trials. A fixation cross. Baseline correction from −200 to 0 ms. Rejection rule (say, ±100 µV) fixed before looking. Filter 0.1 to 30 Hz, continuous, zero-phase, before epoching. At least sixteen participants for anything you would present; three for learning.

Analysis

Mean amplitude in a window (for example 300 to 500 ms for P300) is more robust than peak amplitude. Compare conditions with a paired test across participants, not across trials. For the whole-waveform question “where do the conditions differ,” use cluster-based permutation, which the statistics explainer covers.

Get added to your lab’s EEG protocol. Read it. Practice the consent script out loud.
Implement the oddball in PsychoPy with LSL markers. Pilot on yourself. Confirm a P300.
Run three participants. Save data with participant codes, not names, and the consent forms in the locked place the protocol specifies.
Analyze: grand-average waveforms at Pz with both conditions, shaded standard error, and the mean-amplitude comparison.
Repeat for N170 and ERN.
Write a two-page report in the format of a methods and results section. Include the rejection counts.

What goes wrong

The P300 is there but shifted 40 ms late in everyone. Monitor delay. Photodiode. A huge component before stimulus onset. Baseline window includes the previous trial’s response; lengthen the inter-trial interval. N170 is bigger for houses. Check your condition labels, then check the images are matched in contrast and size. Too few error trials for an ERN. Shorten the response deadline until people err about 15 percent of the time.

Stop here if you like. Three ERPs, three participants, one honest report. Everything about human experiments you will ever need to know started here.

ExhibitGrand-average P300, N170, and ERN waveforms from at least three participants recorded under a lab protocol, with a methods-and-results write-up including rejection counts.

Recall

Why must trial timing be jittered?

So that nothing periodic in the recording (heartbeat, alpha, hum) lines up consistently with the stimulus and survives averaging as a fake component.

Recall

Why compare conditions across participants rather than across trials?

Trials within one person are not independent samples of the population; the question is whether the effect holds across people. A paired test across participants answers it.

Recall

Why is mean amplitude in a window preferred to peak amplitude?

Peaks are biased upward by noise (the maximum of noisy data is always high) and depend on the number of trials; the mean over a window is unbiased and stable.