The signal and its rhythms.
During pretraining, the model reconstructs the EEG waveform and estimates power in six frequency bands.
INTRODUCING / OPEN-WEIGHT EEG FOUNDATION MODEL
Pratibha
From patterns
to insight.
Pratibha is an EEG foundation model. It learns patterns in brain recordings that researchers can use across tasks, or adapt with their own data.
BRAIN INTELLIGENCEILLUSTRATIVE ANNOTATION
01 / THE FOUNDATION
How Pratibha
learns from EEG.
The encoder is open.
Use it with your own data.
92.8MPARAMETERS
22TRANSFORMER LAYERS
512EMBEDDING DIMENSIONS
Apache 2.0WEIGHTS & CODE
Keeping the reconstruction consistent.
ConRec discourages sudden changes in energy between neighbouring parts of the reconstructed recording.
Use it with your own data.
Extract features with the released encoder, or fine-tune it for your task. It supports different channel counts and recording lengths.
02 / EVALUATIONS
How it performs
on EEG tasks.
Seven tasks, compared
with six other models.
5 / 7TASKS LED IN BALANCED ACCURACY
Select a task to compare the models, or open the full results table. In these experiments, the encoder was kept frozen; only a small classifier was trained.
Cognitive workload
Mental arithmetic and resting-state classification.
0BALANCED ACCURACY (%)100
Published LP-AVG protocol: frozen encoder, average pooling and a lightweight classifier. Mean test scores at the best validation epoch, with reported variation across seeds. EEGPT’s Motor-MV result includes pretraining-corpus overlap. Table 3 · MANAS-2 paper ↗
03 / APPLICATIONS
Research uses
and work in progress.
For EEG researchers.
Clinical tools are in development.
Studying brain activity.
For motor imagery, movement, sleep or workload research, you can start with features from Pratibha instead of training an EEG encoder from scratch.
Try the released encoder on your dataset, then test whether fine-tuning improves performance. Results will depend on the task and recording setup.
Start with the encoderSupport for epilepsy assessment.
An EEG recording is only part of an epilepsy assessment. Clinicians also need the person’s symptoms, medical history and details relevant to a suspected epilepsy diagnosis.
We’re developing tools that bring these details together with EEG analysis in a report clinicians can review. The goal is to help them reach a diagnosis sooner.
Screening for memory concerns.
When someone notices changes in memory or thinking, it can be difficult to know whether they need further assessment.
We’re exploring an affordable EEG-based first screen to help clinicians decide who should be assessed further. This work is still in development.
How the pieces
fit together.
The aim is to give clinicians information they can review alongside the rest of the assessment.
- 01 / EEG ANALYSIS
Find patterns in the EEG recording.
- 02 / CLINICAL CONTEXT
Add symptoms, observations and relevant medical history.
- 03 / CLINICIAN REVIEW
Bring the findings into a report for the treating clinician.
From EEG
to a task-specific model.
Pick an application to see the steps.
Record EEG. Record electrical activity through electrodes on the scalp.
A simplified example of model adaptation, not a live analysis.
VALIDATION STUDIES
Validation studies conducted at
AIIMS Delhi
All India Institute of Medical Sciences
GB Pant
Govind Ballabh Pant Hospital (GIPMER)
IHBAS
Institute of Human Behaviour and Allied Sciences
RESEARCH COLLABORATION
Working
with EEG?
If you have a dataset or a project you’d like to discuss, we’d like to hear about it.
04 / OPEN RESOURCES