BrainMedTech is being developed for researchers exploring EEG around stimulation sessions. The proposed system combines defined measurements, assisted review, and traceable drafts. Researchers remain responsible for interpretation.
The public site and interactive dashboard are available now. All sample signals, band powers, and quality flags are synthetic. No real EEG is uploaded or processed, and no AI model runs in the demo. Its report is assembled from predefined data and calculations.
Working demonstration
Available in this showcase
·An EEG PCB image, helmet concept, and research information.
·Two synthetic sessions, channel and frequency-band comparisons.
·A synthetic report preview and plain-text download.
·A collaboration inquiry you can draft and download locally.
Not implemented
Planned research workspace
·Accounts, private recording uploads, and study storage.
·Real EEG processing and automated quality suggestions.
·AI pattern models and language-model report drafts.
·Reviewer approval, audit history, and approved study reports.
The proposed AI system
Four layers, with a clear job for each.
Signal measurements, artifact suggestions, experimental pattern models, and written summaries answer different questions. The planned design keeps their outputs and limitations separate. The tools below are proposed components or research candidates, not capabilities delivered by this website.
01 / Defined signal processing
Measure the signal
The proposed Python analysis service uses MNE-Python for recording preparation and spectral measurements. Filtering, reference choice, excluded segments, and the power-estimation method must be explicit. These calculations do not require a language model.
An accepted EEG recording, acquisition settings, and a versioned analysis protocol.
Output
Quality-reviewed segments, frequency spectra, band-power measurements, and a baseline/post-session comparison with units.
Human review
A researcher checks the recording, processing choices, and whether the comparison periods are equivalent.
02 / Assisted artifact review
Suggest quality labels
ICLabel is a candidate for labeling components after independent component analysis (ICA). It estimates component classes such as brain, eye, muscle, and noise. A label is a review suggestion, not proof that a recording is clean.
An ICA decomposition from a suitable recording. ICLabel documents common-average referencing, 1–100 Hz filtering, and extended-infomax ICA assumptions; suitability for this hardware must be tested.
Output
Candidate component labels and probabilities, with proposed exclusions kept visible.
Human review
A qualified reviewer accepts or rejects exclusions and records reasons. A model probability is not the percentage of usable recording time.
03 / Experimental research models
Investigate patterns
EEGPT, a pretrained transformer for EEG representations, and EEGNet, a compact convolutional network for EEG classification, are candidates for later experiments. They are research starting points. Neither is installed, trained, or validated in this website.
Suitable recordings and independently defined study labels or outcomes, with participant-separated training and evaluation data.
Output
Task-specific research features or predictions, only after the target task and validation criteria are defined.
Human review
Researchers assess errors, uncertainty, population and device differences, and performance on held-out participants. Published results do not establish performance on BrainMedTech recordings.
04 / Language assistance
Draft a traceable summary
A future language model would turn approved measurements and protocol context into a readable draft. Each numerical statement should point back to a measurement and processing version. No language-model provider or model has been selected or connected.
A labeled draft that separates observations, uncertainty, and researcher interpretation.
Human review
A researcher checks every claim, corrects unsupported text, and approves the final report. The model must not invent measurements, diagnose a condition, or infer treatment benefit from band-power changes.
The planned study workflow
From recording to a reviewed report.
A future protected workspace would connect study records and private recordings to a background analysis service. It would store original files separately from derived measurements and reports. Accounts, storage, processing jobs, and approvals still need to be built.
01
Define the study
Set the question, endpoints, protocol version, consent requirements, operator, and reviewer before recording.
02
Capture and register
Keep the original recording with its sample rate, channel positions, reference, units, device settings, timing, and session conditions.
03
Check quality
Review contact, clipping, movement, and contamination. Record retained segments and every exclusion before comparison.
04
Run the comparison
Apply the same versioned method to comparable baseline and post-session periods. Retain measurements and processing provenance.
05
Prepare the draft
Assemble measurements, quality notes, and limitations. Optional AI suggestions remain visibly separate from reviewed findings.
06
Review and release
An authorized researcher checks the evidence and approves an identified report version. Preserve the decision history.
A traceable result needs a recording identifier, segment times, channels, reference, filters, measurement units, processing and model versions, exclusions, and reviewer decisions. Input file formats and acquisition-device support will be published after real recordings are validated.
Validation and interpretation
A change needs evidence and context.
What the comparison means
Band power describes activity in a chosen frequency range. Its value depends on the recording, reference, processing method, and retained segments. In the demo, alpha changes from 20 to 24 µV² by design, giving +20%. That is an arithmetic example. It is not a measured treatment effect or a statistical finding.
What it cannot establish
Movement, alertness, electrode contact, recording conditions, and other confounders can affect a comparison. A before-and-after difference alone does not establish causation, a diagnosis, or therapeutic benefit. Outcomes and appropriate controls must be defined independently of the EEG changes. Classifying an active versus control session would not, by itself, demonstrate benefit.
What must be validated
Before real use, the team must verify acquisition accuracy, timing, processing reproducibility, artifact-review agreement, and performance for the intended population and hardware. Learned models need held-out participant evaluation, error analysis, and predefined conditions for withholding an unreliable result. No BrainMedTech model accuracy or clinical validation is reported here.
Recording around stimulation
The initial proposal focuses on standardized baseline and post-session recordings. Simultaneous electrical or magnetic stimulation requires a separate assessment of interference and signal integrity before interpretation. This demo does not establish suitability for that use.
You can explore every sample without an account or equipment. The guide explains the controls, units, quality indicators, and report. For a research collaboration, prepare a local inquiry draft. There is no connected submission service or published study-enrollment process yet.