BDO (Brain Dynamics with Optimal control)

BDO (Brain Dynamics with Optimal control) is a foundation model for functional MRI (fMRI) that recasts the modeling of brain activity as a stochastic optimal control (SOC) problem. Rather than treating fMRI time series as a sequence to be tokenized and processed by a standard transformer, BDO encodes the region-of-interest (ROI) signals observed at discrete time points into an optimal control policy that steers a continuous latent state through time. The pretrained control policy then serves as a transferable representation for a wide range of downstream neuroimaging tasks.

The model was introduced in February 2025 by researchers at KAIST and Yonsei University (with Juho Lee at KAIST and Byung-Hoon Kim at Yonsei as corresponding authors), in the paper "A Foundational Brain Dynamics Model via Stochastic Optimal Control." It addresses a central difficulty in fMRI analysis: the signals are noisy, high-dimensional, and irregularly informative, which makes purely discrete sequence models brittle. By grounding the representation in a continuous-discrete dynamical system, BDO aims to capture the underlying neural dynamics in a more physically principled and noise-robust way.

BDO sits within the emerging class of brain/neuroimaging foundation models that learn general-purpose representations from large population cohorts and then adapt to specific clinical or behavioral prediction problems with little or no fine-tuning.

#Key Features

• Stochastic optimal control formulation: Brain dynamics are modeled as an SOC problem in which observed ROI signals are encoded into a control policy that drives continuous latent state evolution, yielding physically motivated representations.
• Continuous-discrete state space model: A continuous latent stochastic differential equation is observed at discrete fMRI sampling times, allowing the model to handle the intricate and noisy nature of fMRI recordings.
• Self-supervised ELBO objective: An Evidence Lower Bound derived from the SOC formulation is combined with self-supervised learning, promoting robust and transferable representations without task labels.
• Efficient amortized inference: Locally linear approximations and a simulation-free latent dynamics scheme make inference tractable at population scale.
• Frozen-encoder transferability: The pretrained policy produces embeddings that perform strongly across downstream tasks even with the encoder held fixed.

#Technical Details

BDO was pretrained in a self-supervised manner on the large-scale UK Biobank (UKB) dataset, using resting-state fMRI recordings and associated medical records from 41,072 participants. The architecture couples a continuous latent stochastic differential equation (the controlled dynamics) with discrete-time fMRI observations, forming a continuous-discrete state space model. Training optimizes an Evidence Lower Bound derived from the stochastic optimal control objective, integrated with self-supervised learning to encourage transferable representations; locally linear approximations and a simulation-free formulation keep inference efficient. The authors report state-of-the-art results across demographic prediction, trait analysis, disease diagnosis, and prognosis, with external validation on the HCP-A, ABIDE, and ADHD200 cohorts spanning aging, autism spectrum disorder, and ADHD.

#Applications

BDO targets researchers and clinicians working with resting-state fMRI who need general-purpose representations of brain dynamics. Because the pretrained control policy transfers across cohorts, it can be applied to demographic and trait prediction (such as age and behavioral measures), psychiatric and neurodevelopmental disease classification (autism, ADHD), and disease prognosis, often with a frozen encoder and a lightweight task head. This makes it useful as a feature extractor for neuroimaging studies that lack the large labeled datasets needed to train deep models from scratch.

#Impact

By framing fMRI representation learning as stochastic optimal control over a continuous latent process, BDO offers a distinctive, dynamics-grounded alternative to the transformer-centric approaches that dominate biosignal foundation models. Its use of more than 41,000 UK Biobank subjects places it among the larger-scale brain dynamics models, and its reported state-of-the-art transfer performance across multiple independent clinical cohorts illustrates the promise of population-scale self-supervised pretraining for neuroimaging. As a recent preprint, its long-term adoption and independent reproduction remain to be established, and no public code or weights release was identified at the time of writing.