---
license: cc-by-nc-nd-4.0
---

# BrainLM model

<!-- Provide a quick summary of what the model is/does. -->

The pretrained model of Brain Language Model (BrainLM) aims to achieve a general understanding of brain dynamics through self-supervised masked prediction. It is introduced in [this paper](https://www.biorxiv.org/content/10.1101/2023.09.12.557460v1) and its code is available at [this repository](https://github.com/vandijklab/BrainLM)

## Model Details

### Model Description

We introduce the Brain Language Model (BrainLM), a foundation model for brain activity dynamics trained on 6,700 hours of fMRI recordings. Utilizing self-supervised masked-prediction training, BrainLM demonstrates proficiency in both fine-tuning and zero-shot inference tasks. Fine-tuning allows for the prediction of clinical variables and future brain states. In zero-shot inference, the model identifies functional networks and generates interpretable latent representations of neural activity. Furthermore, we introduce a novel prompting technique, allowing BrainLM to function as an in silico simulator of brain activity responses to perturbations. BrainLM offers a novel framework for the analysis and understanding of large-scale brain activity data, serving as a “lens” through which new data can be more effectively interpreted.

- **Developed by:** [van Dijk Lab](https://www.vandijklab.org/) at Yale University
- **Model type:** ViTMAE
- **License:** [![Preprint License: CC BY-NC-ND 4.0](https://img.shields.io/badge/License-CC_BY--NC--ND_4.0-lightgrey.svg)](https://creativecommons.org/licenses/by-nc-nd/4.0/)

### Model Sources [optional]

<!-- Provide the basic links for the model. -->

- **Repository:** https://github.com/vandijklab/BrainLM
- **Paper:** https://www.biorxiv.org/content/10.1101/2023.09.12.557460v1
- **Demo [optional]:** [More Information Needed]

## Uses

BrainLM is a versatile foundation model for fMRI analysis. It can be used for:

- Decoding cognitive variables and mental health biomarkers from brain activity patterns
- Predicting future brain states by learning spatiotemporal fMRI dynamics
- Discovering intrinsic functional networks in the brain without supervision
- Perturbation analysis to simulate the effect of interventions on brain activity

### Out-of-Scope Use

Currently, this model has been trained and tested only on fMRI data. There are no guarantees regarding its performance on different modalities of brain recordings.  

## Bias, Risks, and Limitations

- The model was trained only on healthy adults, so may not generalize to other populations
- The fMRI data has limited spatial-temporal resolution and BOLD signals are an indirect measure of neural activity
- The model has only been evaluated on reconstruction and simple regression/classification tasks so far
- Attention weights provide one method of interpretation but have known limitations

### Recommendations

- Downstream applications of the model should undergo careful testing and validation before clinical deployment.
- Like any AI system, model predictions should be carefully reviewed by domain experts before informing decision-making.

## How to Get Started with the Model

Use the code below to get started with the model.


## Training Details

### Data

Data stats: 
- UK Biobank (UKB): 76,296 recordings (~6450 hours)
- Human Connectome Project (HCP): 1002 recordings (~250 hours)

Preprocessing Steps: 
- Motion Correction
- Normalization
- Temporal Filtering
- ICA Denoising

Feature Extraction: 
- Brain Parcellation: AAL-424 atlas is used to divide the brain into 424 regions.
- Temporal Resolution: ~1 Hz with 0.735s for UKB and 0.72s for HCP.
- Dimensionality: 424-dimensional time series per scan.

Data Scaling
- Robust scaling was applied, involving the subtraction of the median and division by the interquartile range across subjects for each parcel.

Data split: 
- Training data: 80% of the UKB dataset 
- Validation data: 10% of the UKB dataset
- Test data: 10% of the UKB dataset and HCP dataset

### Training Procedure 

BrainLM was pretrained on fMRI recordings from the UK Biobank and HCP datasets. Recordings were parcellated, embedded, masked, and reconstructed via a Transformer autoencoder. The model was evaluated on held-out test partitions of both datasets.

Objective: Mean squared error loss between original and predicted parcels

Pretraining:
- 100 epochs
- Batch size 512
- Adam optimizer
- Masking ratios: 20%, 75% and 90%

Downstream training: Fine-tuning on future state prediction and regression/classification clinical variables


#### Metrics

In this work, we use the following metrics to evaluate the model's performance:
- Reconstruction error (MSE between predicted and original parcel timeseries)
- Clinical variable regression error (e.g. age, neuroticism scores)
- Functional network classification accuracy

[More Information Needed]

### Results

[More Information Needed]

#### Summary



## Model Examination [optional]

<!-- Relevant interpretability work for the model goes here -->

[More Information Needed]

## Technical Specifications [optional]

### Model Architecture and Objective

[More Information Needed]


**BibTeX:**

```bibtex
 @article{ortega2023brainlm,
  title={BrainLM: A foundation model for brain activity recordings},
  author={Ortega Caro, Josue and Oliveira Fonseca, Antonio Henrique and Averill, Christopher and Rizvi, Syed A and Rosati, Matteo and Cross, James L and Mittal, Prateek and Zappala, Emanuele and Levine, Daniel and Dhodapkar, Rahul M and others},
  journal={bioRxiv},
  pages={2023--09},
  year={2023},
  publisher={Cold Spring Harbor Laboratory}
}
```