Fast temporal dynamics and causal relevance of face processing in the human temporal cortex

The study addresses the longstanding debate between modular and distributed models of face processing in the human brain. Prior work using lesions, functional neuroimaging, EEG/MEG, and non-human primate studies has identified face-selective regions but also patterns suggesting anatomically distributed representations. However, temporal dynamics across large cortical extents within individuals remain underexplored due to methodological limitations. Using electrocorticography (ECoG) to record fast, local signals from many sites within individual temporal cortices, the authors tested the hypothesis that face information is first processed in anatomically distinct, face-selective sites and then appears in non-selective sites over time. The aims were: (1) to map the timing of responses across posterior-to-anterior temporal cortex and assess temporal distribution of face information, and (2) to determine whether stimulation of face-responsive sites causally distorts conscious face perception.

Previous studies have provided causal, spatial, and temporal insights into face processing via lesions, fMRI, EEG/MEG, and primate neurophysiology. Face-selective responses have been found in distinct temporal cortex regions (e.g., fusiform face area), while multivariate patterns from broader non-selective regions can also discriminate faces, supporting distributed coding. Yet these conclusions often relied on low temporal resolution, region-of-interest approaches, averaging across subjects, or limited intracranial sampling (single or paired sites). Recent intracranial recording and stimulation reports advanced understanding but did not directly resolve how anatomical selectivity relates to the temporal distribution of responses across broader cortical territories. This study builds on that body of work by leveraging simultaneous multi-site ECoG within individuals and combining univariate, multivariate, and electrical stimulation approaches.

Participants: Eight neurosurgical patients (6 males, 2 females; 23–68 years) undergoing invasive monitoring for drug-resistant epilepsy. Unilateral implantation: right hemisphere in five, left hemisphere in three. Total temporal cortex (TC) recording sites across subjects: 357. Task and stimuli: Visual categorization task presenting images of faces (human, mammal, bird, marine) and non-face categories (including bodies without faces and other controls). Each image was followed by a blank screen and a behavioral prompt (key press). Recordings: ECoG sampled >1000 Hz (TDT system for seven subjects at 1525.58 Hz; Nihon Kohden for one). Electrodes localized via pre/post-op MRI/CT registration, visualized in native anatomy and transposed to MNI space for group visualization; left-hemisphere coordinates mirrored for display. Preprocessing: Notch filtering for line noise (57–64 Hz) and harmonics (117–123 Hz; 177–183 Hz). Artifact/pathological electrodes excluded based on variance and spike/jump detection criteria. Epochs extracted around stimulus onset; baseline correction (noted windows around −200 to 100/200 ms; specifics varied across descriptions). Time–frequency decomposition performed (Fourier/wavelet), focusing on multiple bands including high-frequency broadband (HFB; ~70–117 Hz in main analyses; high gamma ranges also noted). Power scaled relative to baseline and smoothed (e.g., 50 ms Gaussian). Primary analyses focused on HFB due to highest discriminability and established correlation with local neuronal activity and BOLD. Site categorization (encoding relevance): Sites deemed “active” if HFB in 0–1500 ms post-onset significantly exceeded baseline for any category (permutation tests, p<0.05, FDR-corrected). “Face-selective” sites showed significantly higher responses to all face categories vs all non-face categories. “Human face-selective” sites showed higher responses to human faces vs all non-face stimuli. Remaining active but non-selective sites labeled “task-active.” Univariate analyses: Compared HFB amplitudes across categories and face subcategories; tested for effects of low-level visual features (e.g., spatial frequency, luminance) using ANOVA and permutation tests with corrections (Bonferroni/FDR) to rule out stimulus confounds. Decoding models (within-subject classification of human faces vs non-faces): - Model I: All TC sites. - Model II: Excluding face-selective sites. - Model III: Random subsets of task-active sites removed/included (499 random models per subject) and analyses varying proportion of included face-selective sites to test contribution. - Model IV: Sparse multiple-kernel learning (simple MKL; SVM-based) selecting a subset of sites vs a standard distributed SVM (Model I) to compare sparse vs distributed representations. Performance assessed via balanced accuracy and permutation tests; group comparisons via Wilcoxon signed-rank tests. Temporal analyses: Estimated response onset latency (ROL) at trial level from unsmoothed HFB signals. Related ROL to anatomical position (posterior-to-anterior, MNI y-coordinate). Compared ROL distributions for face-selective vs matched task-active sites. Electrical brain stimulation (EBS): Bipolar and unipolar stimulations applied to pairs/single sites (50 Hz; durations ~1–3 s; pulse width reported as 200 ms in text) in seven subjects during bedside testing while subjects viewed faces (e.g., own face in mirror, cartoons, facial parts). Recorded subjective reports of face percept distortions. Compared proportions of sites in posterior fusiform gyrus (pFG/pFUS) vs mid fusiform gyrus (mFG/mFUS) that elicited distortions (z-tests). Mapped stimulation sites relative to fMRI-defined mFUS and pFUS in a case example (Subject 8).

- Coverage and selectivity: Of 357 TC sites, 53.23% (n=190) were task-active; 13.45% (n=48) human face-selective; 10.64% (n=38) face-selective (faces > non-faces). Face-selective sites clustered in fusiform gyrus and lateral occipital gyrus. - Frequency specificity: HFB (70–117 Hz) provided the strongest univariate and multivariate discrimination of faces vs non-faces, outperforming lower-frequency bands and ERPs in MKL models. - Species effects in face-selective sites: HFB responses were largest for human faces (median 5.21 dB, n=1579), then mammal (4.06 dB, n=1609), bird (3.48 dB, n=1666), and marine (3.48 dB, n=1706). Significant differences for all pairs except bird vs marine (permutation tests; human–mammal p<0.0001, mammal–bird p<0.001, human–marine p<0.0001, mammal–marine p=0.9896; Bonferroni-corrected). - Task-active sites’ univariate responses to faces were weaker (e.g., human 0.73 dB, n=5761; mammal 0.91 dB, n=5497; bird 0.67 dB, n=5387; marine 0.79 dB, n=5068). Only mammal > human and mammal > bird were significant (both p<0.001; others n.s. after FDR). - Decoding (human faces vs non-faces): Model I (all sites) achieved significant discrimination in all subjects. Excluding face-selective sites (Model II) significantly reduced accuracy across subjects (Wilcoxon p=0.0078, n=8); only 2/8 subjects remained significant after exclusion, indicating strong reliance on face-selective sites. Removing random task-active subsets (Model III) did not significantly affect performance overall (Wilcoxon p=0.5781); performance increased with the proportion of face-selective sites included in 6/7 subjects, indicating their dominant contribution and possible redundancy/SNR gains with more face sites. - Sparse vs distributed models: Sparse MKL (Model IV) performed as well as or slightly better than distributed SVM (Model I) across subjects (Wilcoxon p=0.225; mean +2.28% accuracy; up to +8.95% in a subject), indicating that a subset of sites—often human face-selective—sufficed for high accuracy. - Temporal gradients: Both face-selective and task-active sites exhibited a posterior-to-anterior gradient in response onset latency (ROL). Face-selective sites in posterior TC responded earlier than more anterior sites; similar anatomical dependence observed in task-active sites. Selectivity did not systematically vary with anatomical position among task-active sites. - Causal relevance via EBS: Stimulation of posterior fusiform (pFUS/pFG) was more likely to distort face perception than mid fusiform (mFUS/mFG). Proportion with distortions: pFG 9/31 vs mFG 0/21 (z=1.833, p=0.034). Considering only face-selective sites: pFG 6/17 vs mFG 0/7 (z=1.815, p=0.035). Considering only selective sites (as described): pFG 3/14 vs mFG 0/6 (z=1.833, p=0.058). Effects were predominantly in the right hemisphere (one left-hemisphere exception). A case study (Subject 8) showed that stimulation of a pFUS site (site 3) alone produced diverse face distortions, whereas nearby mFUS sites did not.

Findings support a model in which face information is encoded strongly within anatomically discrete face-selective patches (notably in fusiform gyrus) while exhibiting temporally distributed processing along a posterior-to-anterior gradient. High-frequency broadband activity provided the most informative signal for both univariate encoding and multivariate decoding. Decoding performance depended critically on including face-selective sites; removing them substantially reduced accuracy, whereas removing task-active (non-selective) sites had limited impact, suggesting that face information in task-active regions is weak and possibly redundant in the analyzed time window. Sparse models matched or exceeded distributed models, indicating that a relatively small subset of informative sites—largely overlapping with human face-selective electrodes—can capture the discriminative pattern. Temporal analyses demonstrated earlier responses in posterior TC, consistent with hierarchical processing streams. Electrical stimulation established a causal role for posterior fusiform (pFUS) in conscious face perception, with minimal effects from mid fusiform (mFUS), aligning with the encoding/decoding results and reinforcing the importance of specific nodes within the face network. Together, the results reconcile modular and distributed accounts by showing anatomically localized nodes with temporally distributed activity patterns along the ventral temporal cortex.

The study demonstrates that a minority of anatomically discrete sites, clustered in ventral temporal cortex (especially fusiform gyrus), carry strong, behaviorally relevant information about faces, while face-related activity is temporally distributed along a posterior-to-anterior gradient. High-frequency broadband signals optimally index this processing. Decoding relies primarily on face-selective sites, and sparse models using these sites perform as well as or better than distributed approaches. Electrical stimulation identifies posterior fusiform as a key causal node for conscious face perception. Future work should expand simultaneous recordings and perturbations across broader visual and associative networks to delineate how diverse aspects of face information (e.g., identity, emotion, memory linkage) are represented and relayed, and to clarify hemispheric lateralization and inter-node interactions with larger samples and refined stimulation paradigms.

- Limited sample size (n=8) and clinical population with epilepsy may limit generalizability. - ECoG provides sparse, clinically determined coverage; anatomical sampling varies across individuals, potentially missing regions and reducing power to detect weak distributed patterns. - Potential methodological inconsistencies/constraints in preprocessing and parameter choices; some reported windows and parameters varied within text. - Task-active site signals may be too weak to detect reliably with the limited number of sites per subject; lack of significant differences in some ROL comparisons may reflect insufficient power and sampling variability. - Electrical stimulation protocols (site selection, parameters, unilateral implants) and subjective report-based outcomes may introduce bias; effects were largely right-hemisphere with one left-hemisphere case. - Analyses focused on relatively short post-stimulus windows; longer timescales and interactions with broader association networks were not fully characterized.