Topology Agnostic Deep Learning for Variable Sensor Layouts

41 concepts38 papers1 proposal

In traditional deep learning for neuroimaging and brain-computer interfaces (BCIs), multi-channel electrophysiological signals like EEG, sEEG, or MEG are treated as rigid, fixed-dimensional matrices of shape (Channels × Time). In this paradigm, the first row might always represent the Fp1 electrode, the second Fp2, and so on.

This rigid structuring creates a fundamental engineering bottleneck: sensor layout heterogeneity. As highlighted by Wei et al. in the 2021 BEETL Competition, different clinical sites, consumer headsets, and research protocols utilize varying numbers of electrodes and distinct physical placements. When a model trained on a 64-channel laboratory cap is presented with a 16-channel wearable headset, or when surgical constraints dictate unique intracranial electrode placements for individual patients (Mentzelopoulos et al.), the fixed matrix structure breaks down. Standard convolutional or recurrent architectures cannot natively handle missing rows, permuted channel orderings, or shifting spatial semantics without extensive retraining or lossy data truncation (Xiaoxi Wei et al.).

To build universal foundation models for brain activity, researchers have shifted toward topology-agnostic deep learning. Instead of treating sensor arrays as fixed grids, modern frameworks decouple the neural signal from its hardware-specific index, treating recordings as flexible collections of spatially anchored data points.


Decoupling Channels via Spatial Encodings

The core intuition behind bypassing fixed channel matrices is to treat multi-channel data as a set of independent time-series tokens, where each token is explicitly tagged with its physical coordinates in space. Because architectures like Transformers are inherently permutation-invariant, they process sets of tokens natively. By embedding the physical 3D or 4D coordinates of a sensor directly into its signal representation, the model can understand where a signal comes from geometrically, regardless of how many other sensors are present or what index the sensor occupies.

Traditional Matrix Input:
[ Channel 1 (Fp1) ] ---> [ Fixed Weight Matrix ] ---> Broken if channels change
[ Channel 2 (Fp2) ]

Topology-Agnostic Set Input:
{ Signal Token A + Spatial Coordinate (x1, y1, z1) } ---\
{ Signal Token B + Spatial Coordinate (x2, y2, z2) } ----> [ Attention Mechanism ] ---> Robust to any layout
{ Signal Token C + Spatial Coordinate (x3, y3, z3) } ---/

In Neural decoding from stereotactic EEG: accounting for electrode variability across subjects, Mentzelopoulos et al. leverage this concept to decode behavior from stereotactic EEG (sEEG). Because sEEG electrodes are implanted based strictly on clinical necessity, no two patients share the same layout. The authors introduce seegnificant, a framework that maps individual electrode locations to spatial "stamps." This allows a single transformer model to pool data across an entire heterogeneous patient cohort.

Similarly, El Ouahidi et al. introduce REVE (REVE: A Foundation Model for EEG - Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects), which utilizes a 4D positional encoding scheme to scale pre-training across 60,000 hours of highly varied EEG data. By encoding both the spatial coordinates of the electrodes and the temporal dynamics of the signal, the model seamlessly adapts to arbitrary electrode configurations during inference.

Taking this unification a step further, Xiao et al. develop BrainOmni (BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals), which employs a physics-aware sensor encoder. This architecture maps entirely different modalities—such as electrical potentials from EEG and magnetic fields from MEG—into a shared representation space by conditioning the network on the underlying physical properties and spatial coordinates of the recording hardware.


Geometric Processing with Graph Neural Networks

Another powerful mechanism for handling layout variability is to frame the sensor configuration as a geometric graph. In this formulation, sensors are treated as nodes, and the spatial or functional relationships between them form the edges.

As surveyed by Klepl et al. in Graph Neural Network-based EEG Classification: A Survey, Graph Neural Networks (GNNs) compute convolutions over local neighborhoods rather than fixed matrix dimensions. This means that if a sensor is missing or added, the model simply updates the graph's adjacency matrix or performs message passing over the available nodes, preserving the underlying spatial context without collapsing the network architecture.

For instance, Covert et al. introduce the Temporal Graph Convolutional Network (TGCN) in Temporal Graph Convolutional Networks for Automatic Seizure Detection. By treating EEG data as a structural time series over a graph, their model decouples the feature extraction from rigid institutional hardware setups, allowing clinicians to visualize and detect seizures across varying electrode placements.


Spatial Imputation and Canonical Alignment

When downstream tasks strictly require a uniform input size, researchers utilize spatial alignment or imputation to map heterogeneous layouts into a standardized virtual space. Instead of discarding mismatched channels, these techniques mathematically "fill in the blanks" or reorder inputs dynamically.

Differentiable Channel Reordering

Rather than relying on manual channel mapping, models can learn to identify and sort channels based purely on their signal characteristics. Saeed et al. propose CHARM (Learning From Heterogeneous Eeg Signals with Differentiable Channel Reordering), a differentiable neural network layer that automatically rearranges inconsistent sensor inputs into a canonical order. By learning to recognize a channel by its content rather than its hardcoded index, CHARM enables models to remain robust against missing sensors and shuffled inputs.

Graph Signal Interpolation

If a target dataset contains more or different sensors than the training set, Graph Signal Processing (GSP) can be used to interpolate the missing physical space. Drumetz et al. demonstrate this in Spatial Graph Signal Interpolation with an Application for Merging BCI Datasets with Various Dimensionalities, using a closed-form GSP interpolation to smoothly estimate missing electrode values based on their geometric neighbors. This allows fragmented datasets with completely different dimensionalities to be merged into a unified training set.

Self-Supervised Imputation

Masked autoencoders can also be trained to solve layout mismatches as a pretext task. In Spatial Imputation Drives Cross-Domain Alignment for EEG Classification, Liu et al. introduce IMAC, a framework that reframes cross-dataset alignment as a spatial time-series imputation problem. By training the model to reconstruct missing electrode signals across diverse 32-channel and 64-channel caps, the network naturally learns a hardware-independent representation of brain dynamics.


Why It Works: The Shift to Single-Channel and Set Semantics

Ultimately, topology-agnostic architectures succeed because they break the assumption that multi-channel neural data must be processed holistically from the start. By shifting the atomic unit of processing down to the single-channel level or treating the layout as an unordered set with geometric tags, these models achieve true hardware independence.

As demonstrated by Pradeepkumar et al. with the TFM-Tokenizer (Tokenizing Single-Channel EEG with Time-Frequency Motif Learning), tokenizing brain activity at the single-channel level creates a device-independent vocabulary. When combined with spatial coordinates or graph structures, this allows lightweight models running on low-density consumer wearables to inherit the rich, structural representations learned from high-density clinical systems (Xinxu Wei). By moving away from rigid channel matrices and toward flexible, geometry-aware architectures, the field is paving the way for universal foundation models capable of decoding brain activity across any device, clinic, or subject.

Key Foundational Work

Targeting EEG/LFP Synchrony with Neural Nets (2017) — contributed 1 concept elements to this direction.
Parsing epileptic events using a Markov switching process model for correlated time series (2013) — contributed 1 concept elements to this direction.
Spatiotemporal Beamforming: A Transparent and Unified Decoding Approach to Synchronous Visual Brain-Computer Interfacing (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_sensor_layout_variability_a92e6804,
  title = {sensor layout variability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sensor-layout-variability/rdr_a92e680423f4426cb8a73778e1147ee2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_sensor_layout_variability_a92e6804_v1,
  title = {sensor layout variability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sensor-layout-variability/rdr_a92e680423f4426cb8a73778e1147ee2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake