Research map / Probabilistic and scientific modeling / Dynamical systems
BrainLM: A foundation model for brain activity recordings
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces BrainLM, a Transformer masked autoencoder pretrained on about 6,700 hours of fMRI from UK Biobank and HCP for reconstruction, fine-tuning, and zero-shot analyses of brain activity.
Problem
Existing fMRI analysis methods are narrow, task-specific, and trained on small datasets, so they fail to learn generalizable representations of spatiotemporal brain activity.
Approach
Pretrain a masked-autoencoder Transformer on a very large corpus of parcel-wise fMRI time series using spatiotemporal masking, then reuse the encoder for clinical regression, future-state forecasting, attention-based network discovery, and gradient-based in silico perturbation without task-specific training.
Key results
- Fine-tuned BrainLM reduced age MSE to 0.485 versus 2.0 for raw-data SVM, and PCL, GAD7 and neuroticism MSE also improved.
- Masked-reconstruction R2 reached 0.402 on UKB test and 0.316 on HCP with 100% data and 75% masking.
- Attention-based kNN classified parcels into 7 functional networks at 58.8% accuracy, above raw data (39.2%) and VAE (49.4%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune the pretrained BrainLM model | BrainLM · UKB age prediction | MSE: 0.857 ± 0.1135 -> 0.485 ± 0.0252 | Helped |
| Fine-tune the pretrained BrainLM model | BrainLM · UKB PTSD prediction (PCL) | MSE: 0.022 ± 0.0019 -> 0.018 ± 0.0008 | Helped |
| Fine-tune the pretrained BrainLM model | BrainLM · UKB anxiety prediction (GAD7) | MSE: 0.094 ± 0.0079 -> 0.074 ± 0.0053 | Helped |
| Fine-tune the pretrained BrainLM model | BrainLM · UKB neuroticism prediction | MSE: 0.086 ± 0.0047 -> 0.072 ± 0.0049 | Helped |
| Add pretraining (ablation) | BrainLM · UKB future-state forecasting | MSE: 0.695 -> 0.648 | Helped |
| Add pretraining (ablation) | BrainLM · HCP future-state forecasting | MSE: 0.583 -> 0.568 | Helped |
| Use 100% rather than 1% of training data (ablation) | BrainLM with 75% masking · UKB masked reconstruction (MR=0.2) | R2: 0.361 -> 0.402 | Helped |
| Use 100% rather than 1% of training data (ablation) | BrainLM with 75% masking · HCP masked reconstruction (MR=0.2) | R2: 0.300 -> 0.316 | Helped |
| Use 75% rather than 90% masking (ablation) | BrainLM trained on 100% data · UKB masked reconstruction (MR=0.2) | R2: 0.221 -> 0.402 | Helped |
| Use 75% rather than 90% masking (ablation) | BrainLM trained on 100% data · HCP masked reconstruction (MR=0.2) | R2: 0.176 -> 0.316 | Helped |
| Use BrainLM self-attention weights instead of raw parcel time series | Functional network identification · UKB functional network identification | accuracy: 39.2 -> 58.8 | Helped |
Surprising findings
- The pretrained encoder without fine-tuning already outperformed raw-data SVMs on clinical regression, and the non-pretrained model struggled on forecasting.
Limitations
- Clinical and forecasting evaluations use only a subset of disorders and variables, with broader psychiatric and neurological conditions left to future work.
- Code, weights and hyperparameters are stated only as 'will be made publicly available', and the paper gives no compute details.
- Functional network evaluation covers only 7 networks from a single atlas.
Experimental setup
Datasets: UK Biobank Human Connectome Project
Benchmarks: UK Biobank held-out test set Human Connectome Project
Metrics: R2 MSE Pearson correlation accuracy
Baselines: SVM on raw data SVM on pretrained embeddings LSTM Neural ODE Latent ODE Variational Autoencoder Graph Convolutional Network k-NN on raw data
Builds on: masked autoencoder BERT Vision Transformer
Closest papers by meaning
- Brain-OF: An Omnifunctional Foundation Model for fMRI, EEG and MEG (NeurIPS 2026 · similarity 0.90)
- NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals (ICLR 2025 · similarity 0.89)
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.88)
- Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens (NeurIPS 2025 · similarity 0.87)
- BrainWorld: A Structural-Prior-Conditioned Generative Model for Whole-Brain 4D fMRI Dynamics (NeurIPS 2026 · similarity 0.87)
- fMRI predictors based on language models of increasing complexity recover brain left lateralization (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.