Organizations: School of Information Science & Engineering, Lanzhou University, Lanzhou 730000, China · the College of Computer Science and Software Engineering, Shenzhen University, Shenzhen 518060, China · the School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou 510006, China · School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China · Cloud and AI BU, Huawei, Shenzhen, Guangdong 518129, China · the School of Computing, National University of Singapore, Singapore 119077
Electroencephalography (EEG) provides a non-invasive means for neurological and mental health assessment, yet existing EEG models are typically developed for individual tasks and are difficult to unify across heterogeneous signal configurations and assessment objectives. Large language models (LLMs) offer a common interface for multiple tasks, but adapting heterogeneous EEG signals to a shared LLM raises a key challenge: how to preserve task-specific signal characteristics while enabling effective language-side parameter sharing. To address this problem, we propose EEG-HealthLLM, a unified and parameter-efficient framework for heterogeneous EEG health assessment. We introduce an Asymmetric Hierarchical Adaptation Strategy: task-specific EEG adapters preserve heterogeneous spatio-temporal dynamics, while a unified shared LoRA is injected into the LLM backbone (e.g., Phi-3) to capture shared high-level assessment semantics across tasks. We further establish EEG-HealthBench by reorganizing five existing EEG resources into paradigms covering epileptic seizure detection, major depressive disorder assessment, anxiety risk assessment, depressive symptom risk assessment, and emotional state assessment under a unified subject-independent protocol. Extensive experiments show that EEG-HealthLLM achieves an average balanced accuracy (BAC) of 69.58%, ranking first among the evaluated methods in both average BAC and macro-F1. Compared with independent single-task LoRA adaptation, it improves average BAC from 67.70% to 69.58% while reducing the aggregate language-side adapter footprint by 80.0%. Further analyses reveal task-dependent cross-task interactions and trade-offs, supporting the proposed hierarchical adaptation strategy. The implementation is publicly available at https://github.com/Marui-88/EEG-HealthLLM.
Figures & tables
Fig. 1: Overall framework of EEG-HealthLLM. (a) Overall architecture. (b) Task-specific EEG adapter for continuous EEG token generation. (c) Hierarchical sharing strategy with task-specific signal-side adaptation and shared language-side LoRA. The medical-image pathway is shown for architectural context, while the experiments focus on EEG.
Fig. 2: Closed-set generative QA formulation for the five EEG health assessment tasks in EEG-HealthBench. (a) Unified answer spaces across the five tasks. (b) Task-specific questions under a shared closed-set generative interface.
Paradigm
EEG resource
Cohort
Split
Channels / fs
Answer space
Epilepsy [ 66 ]
CHB-MIT
23 subjects / 24 cases
15/4/5 cases
21 / 256 Hz
yes / no
Mumtaz [ 67 ]
Mumtaz EEG
63 subjects
38/13/12 subjects
19 / 256 Hz
yes / no
GAD-7 [ 68 ]
GAD-7 EEG
38 subjects
23/7/8 subjects
14 / 128 Hz
yes / no
PHQ-9 [ 70 ]
PHQ-9 EEG
31 subjects
19/6/6 subjects
14 / 128 Hz
yes / no
SEED [ 72 ]
SEED
15 subjects
9/3/3 subjects
62 / 200 Hz
negative / neutral / positive
TABLE I: Cohort and signal characteristics of the five paradigms in EEG-HealthBench.
Stage
Training objective
Trainable module
Role
Stage 0
Unaligned-projector evaluation
None
Pre-alignment reference
Stage 1
Sequence-only projector alignment
Sequence EEG projector
Sequence EEG token alignment
Stage 2
Global-plus-sequence projector alignment
Sequence and global EEG projectors
EEG adapters for downstream settings
Stage 3-S
Single-task LoRA adaptation
Task-specific LoRA
Single-task language-side baseline
Stage 3-U
Unified shared LoRA adaptation
Shared LoRA
Unified adaptation in this paper
TABLE II: Stage definitions for EEG projector alignment and language-side adaptation.
Fig. 3: Training protocol for single-task and unified shared language-side adaptation.
Model
Adaptation setting
Unified interface
Avg. BAC (%) ↑
Avg. Acc (%) ↑
Avg. macro-F1 (%) ↑
Mean κ↑
Discriminative EEG models
EEG-DINO [ 55 ]
Task-specific
No
69.53
70.09
68.04
0.429
BIOT [ 34 ]
Task-specific
No
68.40
68.11
66.38
0.406
MGFormer [ 75 ]
Task-specific
No
68.12
69.31
64.71
0.396
MSCFormer [ 76 ]
Task-specific
No
66.98
68.22
65.50
0.373
EEGNet [ 30 ]
Task-specific
No
66.40
64.85
64.03
0.356
TABLE III: Overall task-averaged performance on EEG-HealthBench.
Fig. 4: Task-level BAC comparison of representative methods on EEG-HealthBench.
Task
Stage 0
Stage 1
Stage 2
Stage 3-S
Epilepsy
50.00/50.00
51.41/51.41
64.52/64.52
65.32/65.32
Mumtaz
50.00/52.98
90.28/90.51
91.66/91.95
93.22/93.31
GAD-7
50.00/37.50
50.33/37.13
50.00/37.50
54.89/57.92
PHQ-9
50.00/33.15
52.72/37.88
50.00/33.43
64.47/61.28
SEED
32.27/31.85
60.01/60.42
60.21/60.54
60.59/61.02
TABLE IV: Stage-wise performance of the single-task LoRA pipeline (%). Each cell reports BAC/ACC.
Setting
Epilepsy
Mumtaz
GAD-7
PHQ-9
SEED
Avg. BAC
Language-side footprint
EEG-adapter footprint
Total adapter footprint
Five task-specific LoRA adapters (Stage 3-S)
65.32
93.22
54.89
64.47
60.59
67.70
136.97M
19.08M
156.05M
Unified shared LoRA (Stage 3-U)
72.78
93.91
49.94
70.62
60.67
69.58
27.39M
19.08M
46.47M
TABLE V: Comparison between single-task and unified shared language-side adaptation. Task performance is reported as BAC (%), and parameter footprints correspond to all five tasks.
Fig. 5: BAC changes of the original four tasks after extending unified adaptation from four to five tasks.
Fig. 6: Task-level BAC at different checkpoints during five-task unified adaptation.
Task
Fisher ratio
Macro Silhouette
Permutation percentile
Mumtaz
2.81
0.60
100.0%
GAD-7
1.01
0.30
98.2%
PHQ-9
1.15
0.27
93.3%
TABLE VI: Signal-side representation separability before shared language-side adaptation.
Setting
Epilepsy
Mumtaz
GAD-7
PHQ-9
SEED
Avg.
Single-task LoRA
65.32
93.22
54.89
64.47
60.59
67.70
Unified shared LoRA
72.78
93.91
49.94
70.62
60.67
69.58
SR-LoRA-r2
70.36
93.09
47.50
71.46
60.73
68.63
SR-LoRA-r4
70.97
93.01
48.17
70.62
60.61
68.68
TABLE VII: Task-specific residual LoRA ablation results. All values are BAC (%).
Fig. 7: Gradient cosine similarity across the five EEG health assessment tasks on shared LoRA parameters. Positive and negative values indicate aligned and conflicting gradient directions, respectively.
Electroencephalography (EEG) studies brain-generated electrical activity to understand cognition and support diagnosis, monitoring, and brain-computer interfaces.
School of Artificial Intelligence, The Chinese University of Hong Kong, Shenzhen, China · Language, Intelligence, and Machine Center, Shenzhen Loop Area Institute, Shenzhen, China · School of Computing, National University of Singapore, Singapore +2
State Key Laboratory of Brain-machine Intelligence, Zhejiang University · College of Computer Science and Technology, Zhejiang University · MOE Frontier Science Center for Brain Science and Brain-machine Integration, Zhejiang University
Massachusetts General Hospital, Boston, MA, 02114, USA. · Binghamton University, State University of New York, Binghamton, NY, 13902, USA. · Harvard Medical School, Boston, MA, 02115, USA. +1