PHASE: A Physiology-Guided Hierarchical Foundation Model for Intracranial EEG

Yipeng Zhang; Chenda Duan; Yuanyi Ding; Tianyi Wang; Atsuro Daida; Masaki Izumi; Yuta Tanoue; Naoto Kuroda

cs.LGSep 28, 2026

PHASE: A Physiology-Guided Hierarchical Foundation Model for Intracranial EEG

Authors: Yipeng Zhang, Chenda Duan, Yuanyi Ding, Tianyi Wang, Atsuro Daida, Masaki Izumi, Yuta Tanoue, Naoto Kuroda, +5 more

Organizations: UCLA Samueli School of Engineering · UCLA Mattel Children’s Hospital, David Geffen School of Medicine · Children’s Hospital of Michigan, Wayne State University School of Medicine · University of Pennsylvania

Abstract

Clinicians and neuroscientists have long analyzed intracranial electroencephalography (iEEG) through directly measurable physiological characteristics, which carry much of the information that downstream tasks depend on. Recent iEEG foundation models learn by reconstructing or predicting their inputs, which leaves the retention of these characteristics implicit. They are also evaluated mainly on cognitive decoding and a narrow clinical task, i.e., seizure detection. On a broad, clinically relevant benchmark such as Omni-iEEG, they remain below task-specific models when used frozen. We introduce PHASE, a physiology-guided foundation model that makes these characteristics explicit learning targets, pairing them with masked latent prediction in a temporal stage (PHASE-T) within each channel and a spatiotemporal stage (PHASE-ST) across synchronized channels. PHASE is pretrained on heterogeneous recordings from 222 participants at nine clinical sites. On all five Omni-iEEG clinical tasks, frozen PHASE-T outperforms every evaluated foundation model by up to 31%, and fine-tuned PHASE-T surpasses the task-specific models, setting a new state of the art. PHASE-T benefits from physiological supervision, outperforming variants trained with latent prediction alone or auxiliary waveform reconstruction on every task in matched ablations. PHASE-T generalizes to unseen institutions, outperforming the compared models with few or no local labels. PHASE-ST further improves seizure-onset-zone identification over PHASE-T and, when frozen, decodes sound volume and pitch on BrainTreebank better than published models. Beyond task performance, PHASE learns to encapsulate the physiological characteristics clinicians recognize, from seizure onset and its propagation to anatomical region identity, even though its pretraining contains no ictal recordings or anatomical labels.

Figures & tables

Figure 1: PHASE separates shared temporal encoding from cross-channel composition. a , PHASE-T, the temporal encoder shared across channels, whose refinement decoder combines fine- and coarse-scale features into temporal states $Z_{c}$ . b , Pretraining with masked latent prediction and physiological targets. c , PHASE-ST, which updates frozen PHASE-T features across synchronized channels; rows are channels and columns are time steps. The bottom row shows downstream use.

	Event	Window (60 s)		Anatomy		Pathol. brain region
Model	HFO	Ictal	Sleep	Lobe-5	Region-12	Channel	Outcome
	F1	F1	F1	F1	F1	AUC	AUC
PyHFO-Omni	0.8061	–	–	–	–	0.7351	0.7438
CLAP	–	0.9245	0.7225	0.4750	0.3540	0.7684	0.6770
TimeConv-CNN	–	0.8533	0.7118	0.4788	0.3087	0.8061	0.7380
SEEG-NET	–	0.7526	0.6773	0.2520	0.1081	0.7850	0.5952

Table 1: One representation across five clinical tasks, from HFO events to surgical outcome. Results are on the official Omni-iEEG test participants. Frozen PHASE-T and the three foundation models are evaluated with linear probes; fine-tuned PHASE-T is compared with the specialist models published by Omni-iEEG.

C{=}1

marks multichannel models run on one channel at a time. F1 is macro-F1 and AUC is ROC-AUC. Dashes denote tasks that a specialist model does not support. Bold marks the highest performance per column, and underline the highest among frozen encoders.

	Event	Window (60 s)		Anatomy		Pathol. brain region
Pretraining variant	HFO	Ictal	Sleep	Lobe-5	Region-12	Channel	Outcome
	F1	F1	F1	F1	F1	AUC	AUC
Full PHASE-T	0.8138	0.8958	0.7413	0.4822	0.3398	0.8394	0.6369
Latent-only	0.6276	0.8597	0.5398	0.3271	0.1996	0.7570	0.6285
Latent+Recon	0.6298	0.8532	0.6405	0.3126	0.1954	0.7745	0.6156
No-skip	0.7093	0.7684	0.5904	0.3640	0.2371	0.7415	0.6354

Table 2: Ablation Studies. Full PHASE-T is the unablated model, and each variant changes one component. All rows are evaluated under the same pretraining and linear probe conditions; the Full PHASE-T row repeats the frozen scores of Table 1 . F1 is macro-F1 and AUC is ROC-AUC. Bold marks the highest performance per column.

Model	Tohoku	NCNP
CLAP	0.7663	0.5021
SEEG-NET	0.7636	0.5528
PHASE-T, frozen	0.7775	0.6292

Table 3: Cross-institution generalization with no local labels or labels from four participants. (a) Pathological-channel identification at Tohoku and NCNP: the PHASE-T probe and supervised comparators are trained on Omni-iEEG and applied unchanged. (b) Pathology classification at MAYO and FNUSA after fine-tuning on four local participants; comparator values are published results. Bold marks the highest performance per column.

Figure 2: Frozen PHASE-T embeddings carry seizure dynamics and regional identity across institutions. a , Embedding distance from the pre-ictal baseline for each channel on one stereo-EEG shaft before, during, and after one ictal period; orange traces are SOZ channels, and circles mark departures. b , The same distance for SOZ and other channels: the per-seizure median over channels, averaged over each participant’s seizures and then over participants; shading shows 95% participant-bootstrap confidence intervals. c , The proportion of each true region’s channels (rows) voted into each region (columns), pooled over both directions between UCLA and Detroit and shown relative to chance (red, above; blue, below). Each channel is voted on by its ten most similar participants at the other institution; boxes group regions by lobe.

Method	Sentence onset	Speech	Volume	Pitch
PopT + BrainBERT	0.90	0.93	0.87	0.74
BaRISTA	0.91	0.92	0.85	0.73
MVPFormer	0.87	0.90	0.88	0.83
NeuroCLUS	0.96	0.99	0.92	0.83
PHASE-ST, frozen	0.9176	0.9703	0.9450	0.8822
PHASE-ST, fine-tuned	0.9245	0.9786	0.9515	0.8886

Table 4: Auditory decoding on BrainTreebank. Results are mean test ROC-AUC across the seven evaluation participants. Literature rows are published results. Bold marks the highest and underline the second-highest performance per column.

Figure 3: PHASE-ST uses a participant’s concurrent montage, learns this from relation targets, and carries context into auditory decoding. a , Per-participant mean change in SOZ ROC-AUC relative to scoring the target channel alone, for each set of companion channels; shading shows the 95% participant-bootstrap confidence interval. Glyphs depict the target channel (black) and its companions, with each condition’s pooled ROC-AUC below. b , The same change as a function of the number of co-recorded channels. c , Change in auditory ROC-AUC on BrainTreebank relative to the stage without cross-channel attention, averaged over four tasks and seven participants.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Source	Participants	Recs	Channel-hours	Sampling	Recording context	Evaluation use
Omni-iEEG	151	399	5,235	50.67%	Clinical monitoring; official training partition	Test participants and tasks
Hatano	129	144	10,315	43.29%	Clinical monitoring; Detroit non-REM sleep	Tohoku and NCNP cohorts
NYU Podcast	8	8	632	2.68%	Naturalistic listening	Pretraining only
BrainTreebank	10	17	5,434	3.36%	Naturalistic listening; BaRISTA pretraining split	7 test recordings

Appendix

Table 5: Pretraining sources. Recs, recordings. Sampling shares are proportional to each source’s participant count, and channel-hours are counted before artifact rejection. Omni-iEEG and Hatano share 76 participants, so the corpus includes 222 distinct participants.

Parameter	Setting
Waveform units and grid	Microvolts at 1 kHz; released reference convention retained
Pretraining crop	60 seconds (60,000 samples)
High-pass filter	Second-order Butterworth, 2-Hz corner
Mains-notch filters	50 and 60 Hz, each with quality factor $Q=30$
Deterministic filtering	Zero-phase product of the squared-magnitude filter responses, applied in the frequency domain after reflection padding
Input amplitude transform	$u_{t}=\operatorname{asinh}(x_{t}/s_{0})$ for filtered waveforms with scale parameter $s_{0}=20\,\mu\mathrm{V}$ , preserving absolute signal amplitudes

Appendix

Table 6: Shared preprocessing for training and inference.

Parameter	Setting
Requested mask fraction	55% of fine-grid tokens
Short mask spans	2–4 tokens, probability 0.35
Medium mask spans	16–62 tokens, probability 0.45
Long mask spans	156–312 tokens, probability 0.20
Optimizer	AdamW; $(\beta_{1},\beta_{2})=(0.9,0.95)$ , $\epsilon=10^{-10}$
Weight decay	0.1 on matrix weights

Appendix

Table 7: Temporal pretraining with multiscale masking and a moving-average teacher. Mask-span lengths are in fine-grid tokens; schedules are in optimizer updates.

	Event	Window (60 s)		Anatomy		Pathol. brain region
Model	HFO	Ictal	Sleep	Lobe-5	Region-12	Channel	Outcome
	F1	F1	F1	F1	F1	AUC	AUC
LSTM+Attention	0.7338	–	–	–	–	–	–
TimesNet	0.7652	–	–	–	–	–	–
PatchTST	0.7726	–	–	–	–	–	–
eHFO	–	–	–	–	–	0.6611	0.4521

Appendix

Table 8: Clinical prediction from events to participant outcomes: full comparison. Rows above the rule are specialist models published by Omni-iEEG; dashes denote tasks that a specialist does not support. F1 is macro-F1 and AUC is ROC-AUC. Bold marks the highest performance per column, and underline the highest among frozen encoders.

	Event	Window (60 s)		Anatomy		Pathol. brain region
Representation	HFO	Ictal	Sleep	Lobe-5	Region-12	Channel	Outcome
	F1	F1	F1	F1	F1	AUC	AUC
Full PHASE-T	0.8138	0.8958	0.7413	0.4822	0.3398	0.8394	0.6369
Direct targets (no encoder)	0.6223 ∗	0.8112	0.3089	0.2431	0.1425	0.7293	0.5962

Appendix

Table 9: Direct physiological-target control across five clinical tasks. F1 is macro-F1 and AUC is ROC-AUC. ∗ Duration-matched patch-local target. Bold marks the highest performance per column.

Model	Training regime	Tohoku	NCNP
TimeConv-CNN	Supervised source model	0.7434	0.4716
CLAP	Supervised source model	0.7663	0.5021
SEEG-NET	Supervised source model	0.7636	0.5528
PHASE-T	Frozen encoder + source probe	0.7775	0.6292

Appendix

Table 10: Cross-institution transfer without local labels: full comparison. Pathological-channel ROC-AUC at Tohoku and NCNP. Bold marks the highest performance per column.

Model	MAYO	FNUSA
NeuroCLUS	0.40	0.51
MVPFormer	0.36	0.46
PopT	0.34	0.43
BaRISTA	0.30	0.45
PHASE-T	0.7337	0.6656

Appendix

Table 11: Four-participant adaptation: full comparison. Binary F1 on held-out participants at MAYO and FNUSA. Bold marks the highest performance per column.

Quantity	MAYO	FNUSA
Training participants	4	4
Training clips	21,665	33,805
Test participants	20	10
Test clips	133,517	159,313
Binary F1	0.7337	0.6656
ROC-AUC	0.9571	0.8625

Appendix

Table 12: Four-participant adaptation: held-out cohort metrics.

Figure 4: SOZ identification against panel size. Pooled SOZ ROC-AUC as a function of the number of co-recorded channels from the same participant; the dashed line is PHASE-T.

context given to each channel	probe fitted on	pooled	mean $\Delta$ [95%]	median $\Delta$
PHASE-T, no context	—	0.841	—	—
1 channel	same	0.843	0	0
$\leq$ 8 channels, same minute	same	0.853	$+0.018$ [ $+0.007$ , $+0.032$ ]	$+0.003$
$\leq$ 32 channels, same minute	same	0.859	$+0.041$ [ $+0.022$ , $+0.061$ ]	$+0.013$
$\leq$ 64 channels, same minute	same	0.862	$+0.048$ [ $+0.026$ , $+0.071$ ]	$+0.016$
same channels, another minute	another minute	0.850	$+0.015$ [ $+0.002$ , $+0.029$ ]	$+0.001$

Appendix

Table 13: SOZ identification under different contexts. Pooled ROC-AUC and paired per-participant differences (

\Delta

) from the one-channel condition; the

\leq

32-own-plus-32-other-participant row is compared with the

\leq

32-channel same-minute condition. Brackets are 95% participant-bootstrap confidence intervals.

Explore similar work

EEG-JEPA: Structured Latent Prediction for EEG Foundation Models
Jul 31, 2026Jinhao Li, Zhiyuan Ma, Xueqiao Han +8Electroencephalography Foundation Models Electroencephalography
Tsinghua Laboratory of Brain and Intelligence, Tsinghua University · School of Basic Medical Sciences, Tsinghua Medicine, Tsinghua University · School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University +2
What Do EEG Foundation Models Capture from Human Brain Signals?
May 12, 2026Ling Tang, Qian Chen, Jilin Mei +6Electroencephalography Foundation Models Electroencephalography
Shanghai Artificial Intelligence Laboratory · Shanghai Jiao Tong University · Tongji University +2
A foundation model with multi-variate parallel attention to generate neuronal activity
Jun 25, 2025Francesco Carzaniga, Michael Hersche, Abu Sebastian +2Multivariate Time-Series Temporal Attention
IBM Research – Zurich, Rüschlikon, Switzerland · Department of Neurology, Inselspital, Sleep-Wake-Epilepsy-Center, Bern University Hospital, Bern University, Bern, Switzerland

Jul 31, 2026, eess.SPJ/K move · Enter open · S save

EEG-JEPA: Structured Latent Prediction for EEG Foundation Models

Jinhao Li, Zhiyuan Ma, Xueqiao Han, Zhongye Xia+7

Tsinghua Laboratory of Brain and Intelligence, Tsinghua University · School of Basic Medical Sciences, Tsinghua Medicine, Tsinghua University · School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University+2

Electroencephalography Foundation Models Electroencephalography Latent Prediction Electrocardiogram Foundation Models Pretraining

Open paper AskPDF arXiv

Electroencephalography (EEG) foundation models aim to learn reusable representations from large-scale unlabeled recordings. A common pretraining strategy is masked waveform reconstruction, but applying supervision directly to noisy EEG may encourage models to recover predictable background activity, acquisition effects, and artifacts rather than neural structure that transfers across tasks. This raises a central question: what should an EEG foundation model predict to learn transferable representations? We introduce EEG-JEPA a structured latent-prediction framework for EEG foundation modeling. Rather than reconstructing masked voltage samples, a masked context encoder and predictor infer contextual latent states produced by an exponential-moving-average target encoder that observes the complete input. EEG-JEPA organizes target design along three complementary dimensions: target content specifies what representation is predicted, target support specifies where prediction occurs over structured electrode--time regions through Neurotopology-Aware Multi-scale Electrode-Temporal Masking (N-MET), and target depth specifies at which encoder layers supervision is applied. Together, these designs shift EEG pretraining from recovering missing measurements to inferring latent states from structured electrode--time context. We evaluate EEG-JEPA through controlled objective comparisons, frozen multitask transfer, and full fine-tuning. Under the same backbone, pretraining corpus, and training duration, EEG-JEPA improves the 14-task frozen macro balanced accuracy from 40.49% to 50.42% over CBraMod-style masked waveform reconstruction. Multi-source continuation further raises this result to 52.94%, the highest average among the EEG foundation models evaluated on EEG-FM-Bench. Under protocol-matched full fine-tuning, EEG-JEPA also improves the nine-task average balanced accuracy from 68.98% to 70.65%.

Abstract

Explore similar work

EEG-JEPA: Structured Latent Prediction for EEG Foundation Models

What Do EEG Foundation Models Capture from Human Brain Signals?

A foundation model with multi-variate parallel attention to generate neuronal activity