Apply for AI Grants India

Electroencephalography (EEG) is widely used in clinical research, brain–computer interfaces, mental-health studies and neurotechnology. Yet EEG datasets collected across hospitals, laboratories or devices rarely align automatically. Differences in hardware, electrode layouts, sampling rates, reference choices, event markers, participant populations and preprocessing can make two recordings difficult to compare—even when they measure the same underlying brain activity.

EEG data harmonization is the systematic process of making heterogeneous EEG recordings comparable while preserving biologically meaningful variation. It combines metadata standards, signal processing, quality control, statistical adjustment and validation. For Indian research teams working across institutions, harmonization is especially valuable because studies may involve varied equipment, multilingual populations, uneven connectivity and different clinical workflows.

What Is EEG Data Harmonization?

EEG data harmonization aligns recordings collected under different conditions so they can be analysed together. The goal is not to make every file identical. Instead, harmonization separates technical variation from genuine neurophysiological variation.

A robust harmonization programme typically addresses:

• Data structure: file formats, channel names, event codes and metadata
• Acquisition: sampling rate, amplifier settings, filter configuration and impedance
• Electrode geometry: montage, sensor positions and missing channels
• Reference: linked mastoids, average reference, REST, Cz or other choices
• Preprocessing: filtering, bad-channel detection, artefact correction and epoching
• Quality: noise levels, rejected trials, line interference and recording duration
• Population: age, language, medication, diagnosis and task differences
• Analysis: feature definitions, statistical models and site effects

Harmonization is different from simple preprocessing. Preprocessing cleans and transforms an individual recording. Harmonization creates a reproducible framework for comparing many recordings or cohorts.

Why EEG Datasets Become Incompatible

EEG is highly sensitive to acquisition and analysis decisions. Common sources of variation include:

Hardware and acquisition settings

Amplifiers differ in dynamic range, input impedance, anti-aliasing filters and noise characteristics. A dataset sampled at 250 Hz cannot be treated identically to one sampled at 1,000 Hz without resampling and documenting the implications. Hardware-triggered events may also have different timing precision than software markers.

Electrode montages and channel labels

One laboratory may use a 64-channel 10–20 montage, while another records 32 channels using custom names. Labels such as T3, T4, A1 and A2 may correspond to older conventions, whereas modern naming may use T7, T8, TP9 and TP10. Incorrect label mapping can corrupt topographic analyses.

Reference choices

EEG measures voltage differences, not absolute voltage. Re-referencing changes the spatial distribution and amplitude of the signal. Comparing average-referenced data with linked-mastoid data without accounting for the difference can create artificial group effects.

Participant and protocol differences

Task instructions, stimulus timing, language, alertness, medication, sleep, age and clinical status all affect EEG. Technical harmonization cannot compensate for an unrecorded protocol difference. Metadata is therefore as important as signal processing.

Pipeline variation

Different choices for high-pass filtering, independent component analysis (ICA), bad-channel interpolation, epoch rejection and baseline correction can materially change outcomes. Reproducibility requires recording these choices rather than relying on informal lab practice.

Core Standards for EEG Data Harmonization

Use a common data model

The Brain Imaging Data Structure (BIDS) and its EEG extension provide a strong foundation for organizing EEG studies. A BIDS-compatible dataset should include participant information, session and task identifiers, recording parameters, channel metadata, electrode locations and event information.

Useful files and fields include:

• *_eeg.* for signal data in a supported format
• *_channels.tsv for channel names, types and units
• *_electrodes.tsv for electrode coordinates where available
• *_events.tsv for event onset, duration and trial labels
• *_eeg.json for sampling frequency, reference and hardware details
• participants.tsv for controlled demographic and clinical variables

BIDS does not solve harmonization by itself, but it makes differences visible, machine-readable and auditable.

Define a controlled vocabulary

Create a data dictionary before merging datasets. Standardise terms for channel types, task names, event labels, diagnoses, medication status, handedness and missing values. Preserve the original value in a separate field when transformation is necessary.

For example, event labels such as target, TARGET_STIM, stimulus_target and go_target should be mapped to a canonical label while retaining the source label for traceability.

Record provenance

Every transformed file should have a documented lineage. Record:

• Source file and acquisition site
• Software versions and container image
• Processing date and operator or pipeline ID
• Parameters for every transformation
• Channels removed or interpolated
• Epochs rejected and rejection reasons
• Quality-control metrics

A Git-based configuration repository, DataLad workflow or containerised pipeline can help research teams reproduce processing across sites.

A Practical EEG Harmonization Workflow

1. Audit datasets before processing

Start with an inventory rather than immediately concatenating files. For each recording, inspect sampling frequency, channel count, channel labels, reference, filter history, event timing, recording duration and file integrity.

Create a compatibility matrix that identifies which datasets can be harmonized directly, which require transformation and which should remain separate.

2. Convert to a stable intermediate format

EEGLAB, BrainVision and EDF are common EEG formats, but a project should select one internally consistent representation. Conversion must preserve event timing, channel metadata, units and original sampling information.

Never discard the raw files. Store converted data separately and verify that signal duration, channel count and event counts match expectations after conversion.

3. Standardise channel names and types

Map source labels to a canonical naming scheme. Separate EEG, EOG, ECG, EMG, reference and auxiliary channels. Do not silently treat an EOG channel as EEG merely because it has a similar label.

Where channel locations are unavailable, document the limitation. Do not infer precise electrode coordinates from a label unless the montage is known.

4. Align sampling rates

Choose a target sampling rate based on the highest frequency needed for the analysis. For ERP studies, a lower rate may be adequate; high-frequency oscillation studies require more bandwidth.

Apply proper anti-aliasing before downsampling. If recordings have already been filtered, inspect the filter history and avoid applying incompatible operations. Keep the original rate in metadata.

5. Harmonise filters and line-noise treatment

Use consistent filter specifications, including filter type, order, cutoff frequencies and phase behaviour. Zero-phase filtering can reduce phase distortion but may introduce edge effects and leakage if applied carelessly.

Indian laboratories may encounter 50 Hz mains interference and harmonics. Notch filters can help, but they should not replace good grounding, shielding and impedance control. Evaluate whether line-noise removal affects the frequencies used in downstream analyses.

6. Resolve reference differences

Select a project reference strategy and apply it consistently where technically defensible. Average reference is often useful when channel coverage is sufficiently broad and spatially distributed. With sparse or uneven montages, average referencing may introduce bias.

Document the original reference and the new reference. Reference conversion cannot recover information that was removed during acquisition, especially when the original reference electrode was noisy or unavailable.

7. Detect and repair bad channels

Use automated metrics such as excessive variance, flat-line duration, correlation with neighbouring channels and abnormal line-noise power. Automated flags must be reviewed because muscle activity and legitimate task-related changes can resemble artefacts.

Interpolate bad channels only when enough neighbouring spatial information exists. Preserve a count of interpolated channels per recording; a participant with extensive interpolation may not be comparable to a clean recording.

8. Remove ocular, cardiac and muscle artefacts

ICA, regression, canonical correlation and signal-space projection are common approaches. ICA is not automatically superior: it depends on sufficient data length, channel count and suitable preprocessing.

Component removal should be based on reproducible criteria using component topography, time course and auxiliary channels. Save component classifications and avoid deleting components solely because they have high amplitude.

9. Harmonise events and epochs

Align event markers to a common schema. Check trigger delays, duplicated events, missing markers and differences between stimulus onset and recorded response.

Define common epoch windows, baseline intervals and rejection rules. If protocols use different windows, analyse the overlapping interval rather than padding incompatible segments without justification.

10. Apply statistical harmonization cautiously

After technical preprocessing, site effects may remain. Methods such as regression adjustment, mixed-effects models, ComBat-style empirical Bayes correction and domain-adaptation models can reduce systematic site differences.

Statistical harmonization should be trained without leaking outcome information. Include biological covariates such as age, sex, diagnosis and task condition where appropriate. Never remove site effects blindly if site is confounded with the clinical variable of interest.

Quality Control Metrics That Matter

A harmonization pipeline should produce quantitative reports for every recording. Useful metrics include:

• Total duration and number of valid epochs
• Percentage of rejected epochs
• Number of removed and interpolated channels
• Median and maximum electrode impedance, if available
• Peak-to-peak amplitude and robust signal variance
• Signal-to-noise ratio by task or frequency band
• Power spectral density and 50 Hz harmonic power
• Fraction of samples marked as artefact
• Event count and timing consistency
• ICA component count and removed-component count
• Connectivity or coherence stability, if relevant

Visual summaries are valuable: raw traces, channel-level heatmaps, power spectra, event-alignment plots and topographic maps can reveal failures that aggregate metrics miss. Establish exclusion thresholds before examining outcome differences to reduce researcher degrees of freedom.

Validation: How to Know Harmonization Worked

Successful harmonization is not demonstrated merely by similar-looking plots. Validate at multiple levels.

Technical validation

Confirm that files open correctly, units are consistent, events are aligned and channel mappings are accurate. Compare signal distributions before and after each major transformation.

Biological validation

Test whether known effects remain detectable—for example, an expected ERP component, alpha reactivity or task-related spectral change. A method that removes all site variation by destroying biological signal is not successful.

Cross-site validation

Train models using data from some sites and test on an unseen site. Compare performance before and after harmonization. Report confidence intervals, calibration and subgroup performance rather than accuracy alone.

Negative-control validation

Use features or time windows where no meaningful effect is expected. If harmonization creates strong group separation in a negative-control setting, residual confounding or overcorrection may be present.

Common Mistakes to Avoid

• Merging files without checking reference and sampling rate
• Renaming channels without verifying electrode positions
• Using a single rejection threshold for all hardware and populations
• Treating missing metadata as equivalent to a known acquisition setting
• Applying ComBat or another correction before separating biological covariates
• Interpolating many channels and reporting the result as raw data
• Removing all site information from a model when site reflects real clinical context
• Failing to preserve raw data and intermediate outputs
• Allowing test-set information to influence preprocessing decisions
• Reporting only final model accuracy instead of QC and external validation

Tools and Implementation Options

Python ecosystems such as MNE-Python, NumPy, SciPy, PyBIDS and pandas support programmable EEG harmonization. MATLAB users commonly combine EEGLAB with BIDS-compatible plugins. FieldTrip is widely used for MATLAB-based analysis and source-level workflows.

For reproducible deployment, package dependencies with Docker or Apptainer and execute pipelines through Snakemake, Nextflow or a comparable workflow manager. Automated pipelines should still support manual QC review and an audit trail.

For Indian institutions, plan for secure data transfer and governance from the beginning. EEG may be linked to health information, video, voice or cognitive assessments. Use role-based access, encryption, de-identification or pseudonymisation, institutional ethics approvals and clear data-sharing agreements. Under India’s evolving digital and health-data environment, consult the institution’s ethics committee, data-protection officer and applicable regulations before cross-site exchange.

A Recommended Data-Harmonization Checklist

Before analysis, confirm that:

• All files have stable participant, session and site identifiers
• BIDS or an equivalent documented structure is used
• Sampling rates, units, references and filters are known
• Channel labels and locations are mapped correctly
• Events have been validated against the experimental protocol
• Bad channels and rejected epochs are logged
• Artefact-removal decisions are reproducible
• QC reports are generated automatically
• Raw and intermediate data are retained securely
• Statistical correction avoids biological signal leakage
• External-site or held-out validation is performed
• The final dataset includes a complete provenance record

FAQ: EEG Data Harmonization

What is the difference between EEG preprocessing and harmonization?

Preprocessing cleans and transforms individual recordings. Harmonization standardises those transformations and addresses differences between sites, devices, protocols and populations so datasets can be compared.

Can EEG data from different devices be combined?

Often yes, but only after auditing hardware settings, channel layouts, references, sampling rates and noise characteristics. Some datasets may require separate analyses if their protocols or metadata are fundamentally incompatible.

Should all EEG recordings use average reference?

Not necessarily. Average reference depends on electrode coverage and assumptions about the spatial mean. The reference should be selected based on the montage, research question and available metadata, then applied consistently and documented.

Is ComBat suitable for EEG?

ComBat-style methods can reduce systematic site effects in extracted EEG features, but they require careful modelling of biological covariates and validation against overcorrection. They are not a substitute for acquisition metadata or signal-level QC.

What should a multi-site Indian EEG project prioritise first?

Begin with a shared acquisition protocol, BIDS-compatible metadata, controlled event and channel vocabularies, secure data governance and a central QC pipeline. Standardising these foundations is usually more valuable than adding complex machine-learning correction later.

Apply for AI Grants India

If you are an Indian AI founder building tools for EEG data harmonization, neurotechnology, clinical research or responsible health AI, apply through AI Grants India. Explore the programme and submit your application at https://aigrants.in/.