This project started out as an intense curiosity and has been a tremendous amount of fun. This is not a professional project. I did this for fun. I have no real idea as to whether it is valuable to anyone. The project is called PRAYCG. It is an exploratory psychophysiology protocol and software stack for testing whether naturalistic narrative stimuli produce measurable EEG/autonomic state trajectories beyond low-level sensory entrainment and beyond analytic task demand.

This is not a clinical tool, not a diagnostic system, and not a claim that OpenBCI can prove consciousness, memory biology, or any metaphysical theory. It is best understood as an open-source methods workbench for stimulus preparation, acquisition logistics, timing/QC, and exploratory analysis.

What the protocol does
The core design is a three-arm naturalistic viewing protocol:

Phase-scrambled sensory control
The source video is scrambled so recognizable story/faces/narrative meaning are damaged while low-level audiovisual structure is partially preserved. This is meant to ask: “Is the response just light, sound, motion, cuts, rhythm, cue timing, or audiovisual energy?”
Target narrative
The intact narrative is watched naturally. This asks: “What happens when the story is allowed to land?”
Contextual Override
The same intact narrative is watched again, but with an analytic task, usually an upper-right number-cue running-sum task. Target and Override use the same cue-embedded video; only the instructions differ. This asks: “What changes when the participant is measuring/extracting rather than receiving?”

Current hardware stack
My current setup uses:
OpenBCI Cyton + Daisy
OpenBCI gelfree/gelless cap workflow
BrainFlow/LSL EEG stream
LabRecorder XDF recording
ALS-PT19 or similar light sensor for physical screen timing
Polar H10 R-R intervals
Vernier respiration belt
optional Faraday/shielded enclosure setup(see the box section on github)
PRAYCG2.0 protocol runner with structured event logs and self-report/confound reports

What the software release includes
The repository contains or will contain:
MediaPrep + StimulusFingerprint
prepares Target, Override, and phase-scrambled Control videos
embeds cue schedules
generates stimulus-side QC and exogenous regressors
supports predeclared anchor files for event-locked analysis

PRAYCG2.0 Protocol Runner
loads Control, Target, Override, cue schedule, and anchor file
logs structured events
includes pre-run display/audio calibration
includes branch-level self-report and confound reports
includes final reflection baseline
supports override task reporting
BrainFlow / ALS / LSL tools
OpenBCI streaming helpers
ALS/PT19 timing-sensor support
LabRecorder workflow documentation
stream checklist and timing-QC workflow
Master Comprehensive Analysis Suite
time-resolved feature extraction
artifact scoring
stimulus-regressor residualization
MRED/A-MRED endpoint-compression modules
MRED-Peak vs MRED-Resolution
DGA / Decoder Gate Availability
NUPI / Narrative Update Polarity Index
visualization tools
offline plain-English report generator
Documentation
full end-to-end SOP from media procurement to analysis
hardware- how to build a faraday cage for eeg studies
formula/module explanation report
Main analysis idea

The current primary analysis direction is called MRED: Meaning Recognition / Encoding Dissociation.
In plain language, MRED asks whether the system can distinguish:
immediate recognition of meaningful material, versus
delayed integration, after-state, or regulatory carryover.

The compressed endpoint layer currently separates:

MRED-Peak: acute anchor-locked recognition + delayed integration
MRED-Resolution: slower reflective/regulatory recovery after the stimulus
A-MRED: anchor-locked MRED candidate that must pass condition, artifact, and QC gates

I am not treating gamma as a “meaning biomarker.” Lower-gamma/high-frequency features are treated as artifact-sensitive candidate work-signal proxies. The pipeline tries to penalize or flag muscle activity, eye/forehead artifact, cue burden, sensory drive, respiration, confounds, and exogenous stimulus timing.

Why I am posting here
I am looking for critique from people who know OpenBCI, EEG acquisition, artifact handling, and LSL better than I do.

The areas where I most want feedback are:

OpenBCI Cyton + Daisy acquisition reliability
gelfree/gelless cap best practices
BrainFlow configuration
LSL stream naming and synchronization
LabRecorder setup and XDF recording workflow
ALS/PT19 photodiode/light-sensor timing validation
whether my start-pulse design should be longer or barcode-like
artifact control for gamma-band claims
EOG/EMG channel additions
jaw/forehead/eye artifact separation
respiration and HRV integration
whether the analysis suite is overbuilt and how it should be simplified
how to make the public GitHub repository easier for other OpenBCI users to run

Current limitations
The project is exploratory and self-run so far. Several pilot runs revealed exactly the kinds of issues that need better engineering:
timing-pulse detection needing a better physical ALS holder
BrainFlow/OpenBCI effective sample-rate variability
cue visibility/legibility issues in some stimuli
gamma-band artifact vulnerability
the need for EOG/EMG
the need for cleaner locked-anchor and timing workflows

I am not presenting pilot outputs as confirmatory. The point of the current release is to expose the machinery and invite technical critique before treating anything as strong evidence.

Links
GitHub repository:
https://github.com/hbanks87/praycg-open

Here is a list of the fun ideas/theories that I have come up with as a result of running this protocol. The following are not proven claims. They are the current working hypotheses made visible by the protocol,
software, and pilot failures/successes:

1. Meaning recognition and meaning integration can dissociate. A subject can recognize meaningful geometry without showing clean delayed integration under current proxies.
2. Peak-like meaning and resolution-like meaning appear to have different physiological shapes. Some events strike sharply; others settle slowly into afterglow.
3. Analytic extraction does not simply erase meaning. It can reroute, delay, tax, or alter the route by which
   meaning is accessed.
4. Sensory clarity and emotional impact are separable. Clearer semantic access can occur in one branch while stronger emotional absorption occurs in another.
5. Meaning requires access conditions. Sensory access, prior state-space, personal relevance, autonomic
   availability, low task burden, and low confound load all matter.
6. Baseline2 may be one of the most important windows for integration, but it is cumulative and cannot be attributed to Target alone without stronger design.
7. The phase-scrambled Control remains essential because sensory entrainment and meaning integration are not the same construct.
8. The suite is increasingly a theory of availability: whether a living system is available to decode, receive, and be changed by meaning.

Again, this is not a claim of proof. It is an open-source exploratory pipeline, and I am sharing it because I want the engineering and analysis assumptions challenged.

Thank you for taking the time to read this and have a wonderful day,
Hoyt

• 3D printed my own Mark III headset with the electrode holders and whatnot
• used normal regular jumper cables
• used comfort combs from the store (link)
• screwed on my own nuts and screws

Here's what I see on the OpenBCI GUI when I used my own 8-channel montage - one is with 50+60Hz filter and the other is without:


How do I know if my headset is working properly? I am very perplexed by the straight lines after filtering. Thank you in advance everyone!

I'm Aaron, an independent maker/builder — sole developer of NeuralInducer.org, an open-source, non-commercial project exploring neural entrainment through light, sound, and vibration (Arduino-based devices).

I want to be honest upfront: I don't have EEG data yet. What I have is consistent subjective feedback from myself and several family members who use the devices regularly — reports of relaxation, easier sleep onset, and a "settling" feeling during the alpha protocol. That's it. No clinical claims, no "proof" — just a device that seems to do something, and a builder who wants to know if it actually induces measurable brain activity.

If anyone here has a Muse, OpenBCI, or similar setup and is curious to test it and look at the EEG, I'd love the feedback — good or bad. I can share build files, firmware, and full protocol details. Everything is open-source (CC BY-NC).

If you want to dig deeper, the full project — including the scientific background and research papers behind each protocol — is documented at neuralinducer.org.

If you're in the US and interested in testing the physical device itself rather than replicating it, I currently have two NeuroSync ALPHA units built (firmware v6.x — includes vibration/beep, Modo D) ready to send out — happy to work out the details with whoever's interested.

I'd genuinely appreciate any help with this project — whether it's testing, feedback, or just pointing out where I might be wrong.

Happy to answer questions here — my written English is decent, please bear with me if I take a bit to reply.

https://neuralinducer.org/

My name is Rahela and am currently conducting doctoral research on how employees perceive wearable Brain-Computer Interface (BCI) technologies in the workplace, and I would love your input. I`d be grateful if you could take 10–15 minutes to complete my anonymous survey. There are no right or wrong answers, your honest perspective is what matters most. All responses are completely anonymous and voluntary. The survey explores employee attitudes toward wearable BCI devices at work, expectations and concerns regarding emerging workplace technologies, and how these innovations may shape the future of work.

Take the Survey here: https://1ka.arnes.si/a/0c999ab2

There are no direct material benefits for participants. However, your contribution will directly support the development of the first empirically validated guidelines for the ethically responsible implementation of BCI technology in organizations. Every response makes a real difference to my research.

Should you have any questions, please do not hesitate to contact me.
Thank you sincerely for your time and valuable contribution to this research.
Rahela

Disclaimer:
This research is conducted solely for academic purposes as part of a self-funded doctoral study. The study is entirely independent and has no affiliation with, sponsorship from, or representation of any current or previous employer. All analyses, interpretations, and conclusions are exclusively those of the researcher.

We are currently looking for OpenBCI users interested in making money selling their neural data and will be conducting a pilot study this summer. If you're interested, please feel free to reach out here or at my email ucabmk5@ucl.ac.uk. If you're not interested, feel free to leave some feedback and let me know why : )

These wrist mounted sleep trackers really aren't very good. EEG is the best.

I think some kind of adhesive electrodes with a light, small data logger would be the best idea. I mean when you are tossing around with your head on a pillow, any kind of low cost headware is probably going to get messed up.

Any hot tips on suitable amplifiers, data loggers, electrodes, analysis software? I'm willing to hack stuff together etc. and indeed really wish these scientists publishing these articles would share their source code. The world is a bit nuts, the road to getting promising technology/techniques out there is ridiculous. There are various products that claim they can do this stuff but they are never ready to roll, and probably never will be. We do it like this or it doesn't get done, basically.

Can anyone point me to hardware or DIY efforts towards this? We are in the process of adapting a Kernel fNIRS and making it portable - and would love to use OpenBCI systems.

Thanks in advance,
Cat

I'm new to this forum, and I'm looking for some advice. I’m a software engineer with a Bachelor’s in Computer Science, and I’m interested in eventually working in Brain-Computer Interfaces (BCI).

My company offers professional growth funding and is willing to cover the cost of individual college courses/certificates (excluding full degrees or startup expenses) or just opportunities in general. I want to use this opportunity wisely and take courses that will actually move me closer to BCI work. I am more interested in the implementation of machine learning (more specifically, deep learning) with the brain and how that can bridge a gap between people and their prosthetics. I've been interested in this since high school and it never really went away.

I’m trying to figure out which subjects matter most at this stage.

Some options I’m considering:

• Machine learning / AI (especially time-series or signal processing)
• Neuroscience fundamentals (neuroanatomy, electrophysiology, cognitive neuroscience)
• Biomedical engineering–related courses

I didn't take a biology course during college, as I already had a science gen. Ed done, and the last time I took a biology-like course was my senior year of high school.

For those working in or near BCI/neurotech:

• What specific courses would you prioritize first?
• Are there any classes you found especially useful (or wish you’d taken earlier)?
• Is it better to focus on math + ML first, or start building neuroscience knowledge right away?

My long-term goal is to work on software in BCI (nothing really specific in mind right now), possibly pursuing graduate school later, but right now, I want to make the best use of employer-funded coursework. Nebraska doesn't have many options, so it would most likely have to be something online.

Thanks — I really appreciate any guidance.

My name is Elon Paul Anozie Nnamdi, and I’m excited to join this community. I’m based in Nigeria, and I’ve recently started a journey into exploring Brain-to-Brain Interfaces (BTBI).

I’ve set up a GitHub repo called btbi-experiments
, where I’m documenting my ideas, early experiments, and long-term vision for this field. The project is called Synapse, and it’s focused on how we might extend Brain-Computer Interfaces (BCI) toward more direct brain-to-brain communication.

My interests include:

Exploring how BCI/EEG signals can be processed and transmitted.

Investigating possible steps toward multi-user neural communication.

Connecting with others working on signal processing, hardware design, and neuroscience research.

I’m here to learn, share, and collaborate. I’d really appreciate any guidance on good beginner-friendly experiments, open resources, or ways I could contribute meaningfully as I develop this project.

Looking forward to engaging with you all!

Best regards,
Elon Paul
(GitHub: NAZBROTECH
)

I'm looking at two studies as a reference:

First, this study, which used fMRI to capture DMN/PCC deactivation in meditation versus a focused task.
https://pmc.ncbi.nlm.nih.gov/articles/PMC4529365/

Second, this study, which used LORETA to capture DMN deactivation.
https://drfredtravis.com/Papers/final paper DMN and effortless.pdf

Basically I'm trying to replicate the findings of study 1, but as fMRI is prohibitive, I'm looking to use the LORETA approach of study 2.

Does this sort of thing seem plausible to put together with an openbci-based solution?

Thanks very much for any insight anyone can offer.

import time
import numpy as np
import brainflow
from brainflow.board_shim import BoardShim, BrainFlowInputParams, LogLevels, BoardIds
from brainflow.data_filter import DataFilter, FilterTypes, AggOperations
BoardShim.enable_dev_board_logger()
params = BrainFlowInputParams()
params.serial_port = '/dev/cu.usbmodem11'
board = BoardShim(BoardIds.GANGLION_BOARD.value, params) # added ganglion board id here
srate = board.get_sampling_rate(BoardIds.GANGLION_BOARD.value)
board.prepare_session()
board.start_stream()
eeg_chan = BoardShim.get_eeg_channels(BoardIds.GANGLION_BOARD.value)
aux_chan = BoardShim.get_accel_channels(BoardIds.GANGLION_BOARD.value)
print('EEG channels')
print(eeg_chan)
print('Accelerometer channels')
print(aux_chan)
SCALE_FACTOR_EEG = (4500000)/24/(2**23-1) #uV/count
SCALE_FACTOR_AUX = 0.002 / (2**4) 

while True:
    data = board.get_current_board_data(1) # this gets data continiously
    # don't send empty data
    if len(data[0]) < 1 : continue
    time.sleep(1)
    scaled_eeg_data = data[eeg_chan]*SCALE_FACTOR_EEG
    scaled_aux_data = data[aux_chan]*SCALE_FACTOR_AUX
    print(scaled_eeg_data)
    print(scaled_aux_data)

• a ganglion board with 4 channels around Oz
• An Arduino stimulus flickering LED bulbs at 12Hz

I have verified that I am getting data at 200Hz by dividing the number of samples by the amount of time. The sample rate is always near-perfect.

I have good connections with the electrodes. The impedance is always below 20 kOhms.

I have verified that my data analysis pipeline is good by running it on publicly available SSVEP data. Using the publicly available data, I easily see a peak in the power spectrum for the stimulated frequency. So I'm pretty sure my analysis steps are good. They are:
1. Highpass filter above 5Hz
2. Bandstop 59-61Hz (I am in the US)
3. Get power spectrums with DFT, FFT, Bartlett's Method and Welch's Method

I record my data using BrainFlow for Java. I have verified that my data is coming in the same way it does in OpenBCI GUI.

I analyze my data both in java and then export it and and analyze it in MATLAB as well, and see the same results.

I tried putting a makeshift faraday shield with aluminum foil around the OpenBCI board, but this was silly because it just prevented the bluetooth connection from working correctly. So I had to do my recordings with no EMF shielding, and I'm unsure if this could be the source of the issue or not.

My raw data looks really noisy. Even after filtering it, it never looks nearly as smooth as the data I've seen from other devices. The publicly available data I mentioned above is sampled at 256 Hz and looks much less noisy.

Attached is a picture of some of my data as well as a zip file with all of my data and code. In the data, there are markers in milliseconds. The SSVEP stimulus starts at the second marker and ends at the third.

There are a couple of specific concerns I have about the data.

• The raw data has constant jumps. It looks "rectangular" because it will hover around one amplitude for some milliseconds then suddenly jump to another, constantly.
• The filtered data has random spikes, probably reflecting those jumps from the raw data
• The power spectrums always look pretty much the same. In DFTs and FFTs, there are spikes at semi-regular frequency intervals in the power spectrum. In Bartlett's and Welch's methods, there is never any signals above 10 Hz apart from occasionally a bump at 20Hz.

Steps Taken:
* An 8-channel EEG was recorded at a sampling rate of 250 Hz. The electrode positions used were Fp1, Fz, C3, Cz, C4, Pz, O1, and O2. Stimulation was applied with an electrode on the forearm, delivering impulses at a fixed interval of 1 second.
* Blinks were identified as artifacts using the Fp1 electrode and removed through interpolation based on a threshold (mean + 1.75 standard deviations).
* The signal was bandpass-filtered from 1-30 Hz (as my research indicated that PREPs occur within this frequency range) and notch-filtered at 50 Hz (to remove powerline noise).
* I used the Fz, C3, C4, and Pz electrodes to enhance Cz using the Laplacian method. I have attached images showing my Cz channel ('Cz ohne verstärkung' is the raw one and filtered + enhanced is the plot at the top).

Challenges:
* Synchronization of Cap and Stimuli: A major issue I encountered was the lack of synchronization between the electrode cap and the stimuli. This leads to inconsistencies and makes the reliable identification of PREPs difficult.
* Identification and Extraction of PREPs: I have tried to manually identify the PREPs, but it was not possible. There is still too much noise present with a bandpass filter of 1-30 Hz, and nothing significant can be seen at lower frequencies. I also attempted using a wavelet approach, but the identified peaks are not always clear and consistent. Generally, I haven't fully understood the concept of wavelets, and I might be applying them incorrectly.

I hope someone in the forum can help me overcome these challenges and find effective methods for reliably detecting and extracting PREPs in my data.

Thank you in advance for your support!!
Best regards, Ilo

Here are two studies that suggest that more delay results in poorer neurofeedback training. They both had a method to test total delay, but for the life of me, I did not grasp it from their explanations.
https://pubmed.ncbi.nlm.nih.gov/33166941/
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8913511/

Does anyone have experience with other suppliers or know of third-party vendors where I could source the necessary hardware? Any help or suggestions would be greatly appreciated!

Could someone provide a tutorial, because the exported csv data needs to be analyzed using Matlab, and it is easier to analyze the signal after it is smoothed.

The EOG stays with its midline elevated toward 2000 micro volts.

These are the filters I have

I have a ganglion with EOG set up on channel one, with one electrode on each outside corner of the eye, being referenced to each other. I have one analog ground on my mastoid.

I'd expect each sample to have a unique index. What's going on?

Sample from .txt file:
%OpenBCI Raw EXG Data %Number of channels = 4 %Sample Rate = 200 Hz %Board = OpenBCI_GUI$BoardGanglionBLE Sample Index, EXG Channel 0, EXG Channel 1, EXG Channel 2, EXG Channel 3, Accel Channel 0, Accel Channel 1, Accel Channel 2, Other, Other, Other, Other, Other, Timestamp, Other, Timestamp (Formatted) 2.0, 0.0, -10.411881476640701, -83.82611760683358, -76.86990368925035, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158299904E9, 0.0, 2024-08-27 22:35:58.299 2.0, 0.0, 20.08700266201049, 39.82993452809751, 39.0557752372697, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158299904E9, 0.0, 2024-08-27 22:35:58.299 3.0, 0.0, 1.4660407826304436, -57.12697182316333, -39.49708343204111, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158326373E9, 0.0, 2024-08-27 22:35:58.326 3.0, 0.0, 20.0832627620548, -41.09028081316501, -27.80615617055446, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158326373E9, 0.0, 2024-08-27 22:35:58.326 4.0, 0.0, 22.49923813343048, 36.32938816957176, 37.95998455025256, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158330895E9, 0.0, 2024-08-27 22:35:58.330 4.0, 0.0, 8.549411298707128, -23.621208120137453, -15.935713711194694, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158330895E9, 0.0, 2024-08-27 22:35:58.330 5.0, 0.0, 9.906994982622564, -70.9234627597034, -57.358845620416105, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158336272E9, 0.0, 2024-08-27 22:35:58.336 5.0, 0.0, 20.404894158244133, -2.898422465659678, 23.232258524745703, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158336272E9, 0.0, 2024-08-27 22:35:58.336 6.0, 0.0, 31.871427422389388, 21.590442444197834, 48.091373530216515, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158342368E9, 0.0, 2024-08-27 22:35:58.342 6.0, 0.0, 28.370881063863635, -50.466210002079606, -44.037321978248656, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158342368E9, 0.0, 2024-08-27 22:35:58.342 7.0, 0.0, 32.29777601733804, -26.639307384379208, -4.110150051303208, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158348886E9, 0.0, 2024-08-27 22:35:58.348 7.0, 0.0, 50.97483639605343, 68.9226162834093, 69.85385137237608, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158348886E9, 0.0, 2024-08-27 22:35:58.348 8.0, 0.0, 33.161692907102406, -22.56281643267721, -21.956952639855444, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.72481615835535E9, 0.0, 2024-08-27 22:35:58.355 8.0, 0.0, 33.651619801297784, -22.413220434449613, -18.632181579247117, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.72481615835535E9, 0.0, 2024-08-27 22:35:58.355 9.0, 0.0, 47.01802224293351, 57.1755905225873, 73.6648094272241, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158367885E9, 0.0, 2024-08-27 22:35:58.367 9.0, 0.0, 27.40598687529564, 4.514059246517718, 16.95296649914235, 0.12800000607967377, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158367885E9, 0.0, 2024-08-27 22:35:58.367 10.0, 0.0, 24.653420507907867, -43.315521286800504, -29.855621346272528, 0.12800000607967377, 0.0, 0.7840000372380018, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158373942E9, 0.0, 2024-08-27 22:35:58.373 10.0, 0.0, 40.15530582424253, 31.78914962336421, 52.31746048014611, 0.12800000607967377, 0.0, 0.7840000372380018, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158373942E9, 0.0, 2024-08-27 22:35:58.373 11.0, 0.0, 45.15929196495563, 54.078953359276056, 68.27187369111925, 0.12800000607967377, 0.6720000319182873, 0.7840000372380018, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158383254E9, 0.0, 2024-08-27 22:35:58.383 11.0, 0.0, 30.270750241354108, -37.208264659158885, -35.562708678655326, 0.12800000607967377, 0.6720000319182873, 0.7840000372380018, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158383254E9, 0.0, 2024-08-27 22:35:58.383 12.0, 0.0, 47.11151974182576, 8.108103103935719, 25.913766792975366, 0.11200000531971455, 0.6720000319182873, 0.7840000372380018, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158398514E9, 0.0, 2024-08-27 22:35:58.398 12.0, 0.0, 64.27392063848674, 74.67084251530468, 89.66784133762121, 0.11200000531971455, 0.6720000319182873, 0.7840000372380018, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158398514E9, 0.0, 2024-08-27 22:35:58.398 13.0, 0.0, 37.69445165339857, -31.665732924826443, -26.355074987746775, 0.11200000531971455, 0.6720000319182873, 0.7840000372380018, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158404891E9, 0.0, 2024-08-27 22:35:58.404 13.0, 0.0, 39.92717192694545, -20.322616359218955, -5.778145431540906, 0.11200000531971455, 0.6720000319182873, 0.7840000372380018, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158404891E9, 0.0, 2024-08-27 22:35:58.404 14.0, 0.0, 54.42676405515522, 58.503255006857216, 69.89125037193298, 0.11200000531971455, 0.6720000319182873, 0.7840000372380018, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158429897E9, 0.0, 2024-08-27 22:35:58.429 14.0, 0.0, 51.748995686881244, 4.192427850328386, 22.387041134759784, 0.11200000531971455, 0.6720000319182873, 0.7840000372380018, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158429897E9, 0.0, 2024-08-27 22:35:58.429 15.0, 0.0, 35.84320117533207, -37.42143895663321, -20.490911857225, 0.11200000531971455, 0.6720000319182873, 0.7840000372380018, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158434747E9, 0.0, 2024-08-27 22:35:58.434 15.0, 0.0, 55.335559744387865, 38.22925734706223, 44.94237776752561, 0.11200000531971455, 0.6720000319182873, 0.7840000372380018, 0.0, 0.0, 0.0, 0.0, 0.0, 1.724816158434747E9, 0.0, 2024-08-27 22:35:58.434

In my research lab, we have acquired several EmotiBit sensors for studies involving human subjects. Initially, the EmotiBit EDA signals seemed to work fine, as shown in the example below:

In this image, the EDA signals exhibit good quality, with a visible tonic component and identifiable phasic peaks.

However, we have increasingly been registering poor quality signals, which mainly consist of noise or lack the expected tonic and phasic variations, as shown in the images below:

These images represent the raw EDA signals from different experiments. We have visually confirmed that data collection appears to be functioning correctly through the EmotiBit Oscilloscope.

Interestingly, we do occasionally register good quality signals, though this is inconsistent. Additionally, the EmotiBit sensors respond to movements in the bracelet or artificial manipulation, indicating that they are reading and transmitting data, but not reliably capturing physiological EDA responses.

In order to solve this bad quality EDA signal problem, we have tested several solutions, including:
• Changing the EmotiBit device (Two different Adafruit Feather M0 WiFi, and one Adafruit Feather ESP32 Huzzah).
• Replacing EmotiBit electrodes with new ones
• Updating and reinstalling EmotiBit software (versions 1.11.1, 1.8.4, 1.8.1)
• Measuring EDA signals from different individuals
• Adjusting the EmotiBit bracelet (tightening and loosening)
• Measuring EDA signals from different parts of the arm and hands
• Thoroughly cleaning the area where the EmotiBit is placed
• Trying to elicit high arousal and low arousal stimuli
• Using EmotiBit with low battery (30%, 10%) and high battery (70%, 90%)
• Directly feeding the EmotiBit with a power bank through the USB connection.

None of these attempts have resolved the issue.

Has anyone encountered this problem with EmotiBit? If so, were you able to resolve it, and what steps did you follow?

Now the best accuracy I can achieve if I take one file and divide it into test and validation sets (so that all external factors are the same) is approximately 70% (in the best-case scenario), but it is not enough for rehabilitation.
I have also tried to use OpenBCI GUI time series through LSL, but it did not make accuracy better, and in GUI at all no difference can be seen between rest and action when looking at graphs.

Is it possible to achieve better accuracy with this EEG headset? If yes, then what am I missing? What should I change, or add? What should be the bandpass filter range? How should the graph of rest and action states look?

Raw data:

All electrodes after preprocessing (with pink line showing rest - 0 and action - 1 states):

Does spikes before 1500 frame influences data overall? Should I remove them? Can they be detected already in raw data?
C3 electrode after preprocessing (yellow - action, purple - rest):