Functional MRI (fMRI) measures blood oxygenation rather than neural firing. The signal peaks about five seconds after the neurons fire, so one brain image blurs together several words of a spoken story. Intracranial recordings from epilepsy patients have millisecond timing and separate cleanly into frequency bands, but they come from a few dozen damaged brains with electrodes placed wherever the surgeon needed them. This proposal uses the two together. We fit a model on intracranial data that predicts, word by word, how much power each frequency band should carry in response to language. We then run that model on the stories heard by hundreds of healthy fMRI participants, and convert each predicted band into a slow blood-flow signal using a band-specific delay and sign. The question is whether fMRI responses in healthy people are better predicted from these intracranial "virtual band" signals than from language-model features pushed directly through the usual delay filter. The advantage should be largest when fMRI data are scarce.
The idea comes from a monkey finding that has never been tested in humans or during language. In simultaneous recordings, gamma power set the size of the blood-flow response, a 18–35 Hz band set its timing, and together they carried 30% more information than gamma alone (Different LFP frequency bands convey complementary information about the BOLD signal). If that structure holds for human language, an fMRI model built on band-separated, electrophysiologically shaped signals should beat one that treats the blood-flow signal as a generic smoothed copy of word features. It should also be less fooled by speech rate, which is a known confound when word features are averaged down to the fMRI sampling rate (Interpretable multi-timescale models for predicting fMRI responses to continuous natural speech).
Core Claim
A stimulus-level bridge can move information from clinical intracranial recordings into models of healthy fMRI, and it needs no paired recordings and no matching of electrodes across people. It works in three steps. First, we fit an intracranial model that maps language-model features to band-limited power: high gamma, the 18–35 Hz band, alpha, and theta. Second, we apply it to new stimuli to produce predicted band time courses at millisecond resolution. Third, we model each fMRI voxel as a learned weighted sum of these virtual bands, each convolved with its own estimated hemodynamic response, meaning the delayed blood-flow curve that follows neural activity.
Three outcomes would make this a contribution. The virtual-band model should beat a matched direct-feature model on held-out healthy fMRI. The gain should depend on keeping the mid-frequency and alpha bands, not only high gamma. And the gain should be largest when fMRI training data are small, which is the regime of most clinical and 7-Tesla studies.
The bridge also gives a direct test of whether the clinical origin matters. We can remove channels showing pathological high-frequency oscillations and see whether healthy-fMRI prediction improves.
Where this sits in prior work
Several papers motivate the gap. Language Reconstruction with Brain Predictive Coding from fMRI Data (PREDFT) handles the hemodynamic delay inside fMRI using a delay filter and a predictive side network, and reaches 27.8% BLEU-1 on Narratives. It treats the blood-flow signal as one undivided quantity and never sees electrophysiology. Scaling laws for language encoding models in fMRI shows fMRI encoding gains 122% per order of magnitude of fMRI data but only about 4.4% per order of magnitude of model size, with a plateau near 30B parameters. That pattern says the limiting factor is data, not feature richness, and it is exactly where a prior from another modality might help.
Two other papers define the intracranial side. The Temporal Structure of Language Processing in the Human Brain Corresponds to The Layered Hierarchy of Deep Language Models maps language-model layers onto high-gamma timing in nine patients but models only high gamma and never connects to fMRI. Neuroprobe: Evaluating Intracranial Brain Responses to Naturalistic Stimuli supplies our intracranial data and our method baseline. It shows that a tuned linear model on time-frequency spectrograms beats foundation models (within-session AUROC 0.660). It also shows that cross-subject transfer collapses to about 0.539. We take that collapse as a constraint: our bridge never transfers electrode-level models across people, only stimulus-driven predictions.
The closest multimodal work is Multi-modal Cross-domain Self-supervised Pre-training for fMRI and EEG Fusion (MCSP). It distills between fMRI and scalp EEG for diagnostic classification. Scalp EEG cannot resolve high gamma, and MCSP does not model continuous language, so it neither tests band-specific hemodynamic coupling nor uses intracranial data.
A tight relationship between BOLD fMRI activation/deactivation and increase/decrease in single neuron responses in human association cortex supports one design choice. In two patients, negative blood-flow responses tracked real suppression of firing. That justifies letting band weights be signed instead of forcing them to be positive.
Two papers bear on clinical artifacts. Omni-iEEG: A Large-Scale, Comprehensive iEEG Dataset and Benchmark for Epilepsy Research provides a detector for pathological high-frequency oscillations, reported at event F1 0.8061. The development of aperiodic neural activity in the human brain warns that shifts in the broadband "1/f" background, the aperiodic part of the spectrum, can masquerade as band power. We therefore separate aperiodic and periodic components before defining bands.
Research Program
The intracranial stage uses BrainTreebank, as packaged by Neuroprobe: 10 patients, 40 hours of film watching, and more than 223,000 aligned words. For each patient and electrode, after Laplacian re-referencing as in Neuroprobe, we compute word-locked power in four bands at lags from −100 to +600 ms. Each band is computed both before and after removing a fitted aperiodic component. We fit ridge models from contextual language-model embeddings plus acoustic envelope features to each band's power, and keep only electrode-band pairs that generalize across sessions. This step is deliberately linear, because Neuroprobe shows linear spectrogram models are the strongest choice here.
The bridge stage runs these models on the transcripts and audio of the Narratives stories (The "Narratives" fMRI dataset for evaluating models of naturalistic language comprehension; 345 subjects) and on the podcast stimuli used by Antonello et al. The result is a few thousand virtual band channels at high temporal resolution, all predicted from the stimulus. We compress them with PCA within each band, fit on the intracranial training stimuli only.
For each fMRI voxel or parcel, the predicted response is
$$\hat{y}_v(t) = \sum_{b} \sum_{k} w_{v,b,k}\,\big(h_b * z_{b,k}\big)(t),$$
where the four terms are defined as follows. $z_{b,k}$ is the $k$-th principal component of virtual band $b$. $h_b$ is a hemodynamic delay kernel shared by all voxels for band $b$, parameterized by a small set of finite-impulse-response taps. $*$ denotes convolution followed by resampling to the fMRI sampling interval. $w_{v,b,k}$ is a signed ridge weight. Sharing $h_b$ across voxels is the band-specific part of the hypothesis. If the monkey result transfers, the fitted 18–35 Hz kernel should differ in latency from the gamma kernel, and the alpha weights should be predominantly negative.
Baselines
The naive reference is a model with only word rate, phoneme rate and the acoustic envelope, passed through a standard delay filter. Any language model has to beat this to show it does more than track speech density, the confound documented in Interpretable multi-timescale models for predicting fMRI responses to continuous natural speech.
The strongest existing encoding method is the Antonello et al. pipeline from Scaling laws for language encoding models in fMRI: the same language-model embeddings, Lanczos-downsampled to the fMRI sampling rate, then a voxelwise ridge model with delays. It is the strongest competitor because it uses exactly the information our bridge starts from. Any gain must therefore come from the intracranial reshaping, not from better features. The file gives that paper's scaling slopes but no reported voxelwise encoding correlation on Narratives, so we report "no reported result found on Lacuna" for an absolute number and re-run the method ourselves.
The fairness rules are as follows. Both models use the same language-model layer, chosen by nested cross-validation separately for each model. Both use the same story-level train/test splits and the same delay range. Both get an identical regularization search grid and are given the same input dimensionality.
For decoding, the relevant reference is PREDFT's 27.8% BLEU-1. We do not claim to beat it with an encoding model. Instead we use a secondary segment-identification test: rank the true 60-second window among distractors using predicted responses.
Three ablations isolate the key ingredient. The first, high-gamma-only, keeps only the gamma virtual band; this is the Goldstein-style convention. The second, shuffled-bridge, refits the intracranial models on circularly shifted neural data, keeping the nonlinear-looking reshaping of features but destroying real brain content. The third, random-filter, replaces the intracranial ridge weights with random projections matched in band count and dimension.
Beating the baseline means a held-out voxelwise correlation gain over the Antonello-style model that survives three conditions. It must be significant in a subject-level permutation test, appear in language-network parcels, and persist after the speech-rate regressors are partialled out. Both the shuffled-bridge and random-filter ablations must lose that gain.
Why not just…?
Why not just collect more fMRI? Antonello et al.'s 122% gain per order of magnitude of data says this is the strongest rival, and with enough hours it may erase our advantage. We treat that as a measurement rather than a threat to argue away. On their three deeply scanned subjects we vary fMRI training data from about one hour to twenty and report how many fMRI hours the intracranial prior is worth. If the curves merge above a few hours, the honest conclusion is that the bridge is a small-sample tool for clinical and limited-scan cohorts. That is still where most studies live.
Why not just use high gamma, the signal intracranial language work already trusts? Two papers suggest it discards something. Magri et al. found timing information in 18–35 Hz, and Spectral Changes in Cortical Surface Potentials during Motor Movement shows that low and high frequencies move in opposite directions and cancel when mixed. The high-gamma-only ablation settles this directly. If it matches the full model, the band-separation hypothesis fails for language, and we will report that.
Why not just let a deep model learn the hemodynamics end to end from fMRI, as PREDFT does? PREDFT still loses words at window boundaries and needs a predictive side network to compensate. A model learned only from fMRI also cannot tell neural timing apart from vascular timing. We will add PREDFT's delay-filter backbone with virtual bands as an extra input and test whether its decoding improves. If it does not, the bridge adds nothing a learned fMRI model lacks.
Why not worry that epileptic brains give the wrong prior? This is a real risk, given altered aperiodic backgrounds and pathological oscillations. We rebuild the bridge three ways: excluding channels flagged by the Omni-iEEG oscillation detector, without aperiodic removal, and from each patient alone. Prediction should not get worse when pathological channels are removed, and the bridges built from different patients should agree on the band delay kernels.
Decisive Experiments
The first experiment is the primary comparison: on Narratives, with whole stories held out and subjects pooled, the full virtual-band model against the Antonello-style direct model, scored on held-out correlation in language parcels.
The second is the speech-rate test. We use variance partitioning against word-rate regressors, and we evaluate on story segments matched for speech rate but differing in content.
The third is the band-kernel test. The fitted 18–35 Hz kernel should peak at a different latency from the gamma kernel, with confidence intervals from bootstrapping over subjects, and alpha weights should be predominantly negative. A null result here falsifies the transfer of Magri et al.'s finding to human language, even if overall prediction improves.
The fourth is the sample-size curve on the deep-phenotyping subjects, which reports the fMRI hours at which the advantage vanishes.
The fifth is the pathology test: the three rebuilt bridges. Together these determine whether intracranial frequency structure carries real, transferable information about healthy hemodynamics, or whether it is only a re-encoding of features fMRI already had.
References
- Different LFP frequency bands convey complementary information about the BOLD signal
- Interpretable multi-timescale models for predicting fMRI responses to continuous natural speech
- Language Reconstruction with Brain Predictive Coding from fMRI Data
- Scaling laws for language encoding models in fMRI
- The Temporal Structure of Language Processing in the Human Brain Corresponds to The Layered Hierarchy of Deep Language Models
- Neuroprobe: Evaluating Intracranial Brain Responses to Naturalistic Stimuli
- Multi-modal cross-domain self-supervised pre-training for fMRI and EEG fusion
- A tight relationship between BOLD fMRI activation/deactivation and increase/decrease in single neuron responses in human association cortex
- Omni-iEEG: A Large-Scale, Comprehensive iEEG Dataset and Benchmark for Epilepsy Research
- The development of aperiodic neural activity in the human brain
- The “Narratives” fMRI dataset for evaluating models of naturalistic language comprehension
- Spectral Changes in Cortical Surface Potentials during Motor Movement