Direct intracranial recordings from the human brain—such as electrocorticography (ECoG) and stereotactic electroencephalography (sEEG)—provide an unparalleled window into neural dynamics, offering millisecond temporal resolution and high spatial localization. However, a fundamental challenge prevents researchers from using this data to construct a complete, brain-wide picture of cognition: spatial sampling bias.
Because invasive electrode implantation carries inherent surgical risks, electrode placement is dictated entirely by clinical necessity—typically for seizure localization in patients with drug-resistant epilepsy—rather than experimental design. Consequently, researchers can only observe the brain through a fragmented set of "keyholes" that vary drastically from patient to patient, leaving critical functional regions under-sampled or entirely unmonitored.
The Geometry of the Blind Spot
To understand why spatial sampling bias is more than a simple missing-data problem, consider the brain as a complex, directed network. If an experiment aims to map how information flows during a cognitive task, omitting even a single critical node can lead to entirely incorrect causal inferences.
As highlighted by Alia Seedat et al., clinical requirements frequently result in uneven anatomical coverage, leaving researchers unable to systematically sample all regions relevant to a given cognitive theory. This constraint manifests in several distinct ways across literature:
- Hemispheric and Regional Biases: In speech and language decoding, electrodes may be concentrated in a single hemisphere depending on the patient's pathology. As noted by Hemmings Wu et al., focusing recordings on a non-dominant hemisphere due to clinical constraints can lead to an incomplete or skewed understanding of robust neural activity.
- Cortical vs. Subcortical Gaps: Surface-level grids capture superficial cortical dynamics but remain blind to deep-seated structures. Kazuki Sakakura et al. point out that surface arrays fail to capture crucial subcortical structures like the hippocampus or insula, creating a stark spatial sampling gap between the surface and deep brain regions.
- Network Fragmentation: When studying complex behaviors like sequence learning or memory retrieval, the clinical footprint might capture one node of a network (e.g., the temporal lobe) while missing upstream or downstream partners. Radhika Madhavan et al. emphasize that this non-uniform coverage means critical functional regions are often left entirely unobserved, obscuring the broader network coordination.
Clinical Constraints (Pathology-driven placement)
│
▼
Sparse, Fragmented Electrode Coverage
│
├─► Hemispheric Bias (e.g., missing language-dominant regions)
├─► Depth Bias (e.g., cortical surface vs. deep subcortical structures)
└─► Network Blind Spots (e.g., missing unobserved confounding nodes)
Stitching the Mosaic: Cross-Subject Imputation
If any single patient only provides a sparse, biased glimpse of brain activity, how can we infer global neural dynamics? The solution lies in treating data from multiple subjects as complementary views of a shared underlying dynamical system.
One powerful approach is neural inpainting. By leveraging datasets with partially overlapping recording fields, modern machine learning architectures can learn the latent rules governing how different brain regions interact. Ji Xia et al. introduced a transformer-based framework, NeuroPaint, designed to infer the dynamics of unrecorded brain areas by training on multi-animal datasets with overlapping coverage.
The intuition behind this approach is analogous to image inpainting: if a model understands the statistical structure of the world, it can fill in a missing patch of an image based on the surrounding context. In neuroscience, if the model learns the canonical correlation and propagation delays between Region A and Region B from Patient 1, it can infer the unobserved dynamics of Region B in Patient 2 by looking at Patient 2's recorded activity in Region A.
Multimodal Integration and Open Data
Beyond computational imputation, researchers mitigate spatial sampling constraints by combining sparse, high-resolution intracranial data with non-invasive, whole-brain imaging modalities.
For instance, Christine Grützner et al. demonstrated how combining magnetoencephalography (MEG) and functional magnetic resonance imaging (fMRI) helps overcome the restricted spatial coverage of localized protocols, ensuring that critical regions like the anterior temporal and inferior frontal cortices are not omitted from the functional map. While fMRI provides the global spatial context, the sparse intracranial electrodes anchor the high-frequency temporal dynamics.
Simultaneously, the field is moving toward large-scale, standardized data sharing to combat individual sampling biases. Open multi-center datasets, such as the one presented by Alia Seedat et al., aggregate recordings across dozens of patients with diverse clinical implantations. By pooling these fragmented clinical footprints into a unified coordinate space, the community can systematically fill in the spatial gaps, turning a collection of isolated, biased recordings into a comprehensive, brain-wide atlas of human cognition.