Project Details
Description
This project will develop a personalized whole-brain modeling framework, including a machine learning model and a training pipeline that leverages structural and functional MRI data to predict individual brain activity patterns. Subsequent studies will use this framework to personalize neuromodulation therapies for neuropsychiatric disorders. Currently, clinicians target treatments such as transcranial magnetic stimulation according to population-level empirical associations between brain regions and functions, resulting in widely varying efficacy across patients. Personalized simulation of how local stimulation affects individual whole-brain dynamics will better predict responsiveness and enable optimization in-silico.
Biophysical models, including neural mass models (NMMs), which represent the brain as a network of regions with predefined low-dimensional dynamical differential equations, fail to replicate both the complex behavior within and the heterogeneity between brain regions from functional MRI (fMRI) data. By contrast, deep neural networks (DNNs) can learn complex patterns from data but lack interpretability and must train on large datasets. A DNN trained on one population may need extensive retraining to transfer to another, but data from an individual or a population with a rare disease may be too scarce. Furthermore, we need a foundational model that can simulate healthy and disease states and transition between them.
Our model will combine the strengths of both approaches by representing the brain as a network of brain region nodes implemented as recurrent neural networks (RNNs) trained on fMRI data. A separate structure-function mapping DNN will predict intrinsic properties of and interconnectivity between nodes from heterogeneous structural MRI features, including cortical thickness and myelination. We will first train a foundational model on healthy young adults from the Human Connectome Project (HCP) 1200 Subjects (S1200) dataset so that the common node RNN will learn universal rules of dynamics while the mapping DNN learns to induce region- and individual-specific patterns. We will then fine-tune the model to optimally fit individual fMRI data. To test that the model reproduces cognition-relevant features, we will predict participants’ psychometric scores from time series generated by their personalized models. We will also generalize the model to the HCP Development and Aging datasets to prepare for future modeling of Schizophrenia and Alzheimer’s Disease (AD) patients, who show abnormalities in the networks that change most during development and aging.
Our personalized whole-brain modeling framework will empower clinicians to predict patient responses to existing therapies and translational researchers to test new approaches in simulation before trying them on human subjects.
Biophysical models, including neural mass models (NMMs), which represent the brain as a network of regions with predefined low-dimensional dynamical differential equations, fail to replicate both the complex behavior within and the heterogeneity between brain regions from functional MRI (fMRI) data. By contrast, deep neural networks (DNNs) can learn complex patterns from data but lack interpretability and must train on large datasets. A DNN trained on one population may need extensive retraining to transfer to another, but data from an individual or a population with a rare disease may be too scarce. Furthermore, we need a foundational model that can simulate healthy and disease states and transition between them.
Our model will combine the strengths of both approaches by representing the brain as a network of brain region nodes implemented as recurrent neural networks (RNNs) trained on fMRI data. A separate structure-function mapping DNN will predict intrinsic properties of and interconnectivity between nodes from heterogeneous structural MRI features, including cortical thickness and myelination. We will first train a foundational model on healthy young adults from the Human Connectome Project (HCP) 1200 Subjects (S1200) dataset so that the common node RNN will learn universal rules of dynamics while the mapping DNN learns to induce region- and individual-specific patterns. We will then fine-tune the model to optimally fit individual fMRI data. To test that the model reproduces cognition-relevant features, we will predict participants’ psychometric scores from time series generated by their personalized models. We will also generalize the model to the HCP Development and Aging datasets to prepare for future modeling of Schizophrenia and Alzheimer’s Disease (AD) patients, who show abnormalities in the networks that change most during development and aging.
Our personalized whole-brain modeling framework will empower clinicians to predict patient responses to existing therapies and translational researchers to test new approaches in simulation before trying them on human subjects.
| Status | Not started |
|---|---|
| Effective start/end date | 1/01/25 → 31/12/27 |
Fingerprint
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.