A novel adaptive sampling technique for mapping brain function
Lead Research Organisation:
King's College London
Department Name: Neuroimaging
Abstract
Diagnostic cognitive tests are used in many clinical situations; for example, cognitive tests assessing memory, attention and other processes can be used to detect early signs of dementia and monitor cognitive impairment after a traumatic brain injury. Such diagnostic tests are thought to activate different brain systems. However, the tests have not been designed for this purpose, and so there may be better, more sensitive tests. We will use a new functional brain imaging technique that "automatically" finds the best cognitive tests that activate specific brain systems.
The new technique is called neuroadaptive Bayesian optmisation. It uses neuroimaging (functional MRI) to assess brain function in real-time, guided by a powerful machine learning algorithm. This approach can explore across many cognitive tests and find the ones that are most strongly associated with a specific brain system. This is a "closed-loop" approach: the algorithm suggests a cognitive task, analyses how this affects the brain, and uses this knowledge to update what it "knows" about which cognitive task is best for activating that brain region, all automatically; by taking the human scientist "out-of-the-loop", it is possible to be much more efficient and unbiased and search over many more types of cognitive task very quickly, as our pilot data shows.
We will use this technique to develop a new diagnostic battery that we hope will be more sensitive than tests currently in use to distinguish between two important brain systems that have been associated with neurological and psychiatric disorders. If the project is successful, we would work with clinicians to use the tool to improve cognitive testing in a range of clinical situations and help diagnose cognitive problems better in patients.
The new technique is called neuroadaptive Bayesian optmisation. It uses neuroimaging (functional MRI) to assess brain function in real-time, guided by a powerful machine learning algorithm. This approach can explore across many cognitive tests and find the ones that are most strongly associated with a specific brain system. This is a "closed-loop" approach: the algorithm suggests a cognitive task, analyses how this affects the brain, and uses this knowledge to update what it "knows" about which cognitive task is best for activating that brain region, all automatically; by taking the human scientist "out-of-the-loop", it is possible to be much more efficient and unbiased and search over many more types of cognitive task very quickly, as our pilot data shows.
We will use this technique to develop a new diagnostic battery that we hope will be more sensitive than tests currently in use to distinguish between two important brain systems that have been associated with neurological and psychiatric disorders. If the project is successful, we would work with clinicians to use the tool to improve cognitive testing in a range of clinical situations and help diagnose cognitive problems better in patients.
Technical Summary
Patients with a range of neurological and psychiatric disorders have cognitive impairments associated with abnormal activity and connectivity within fronto-parietal brain networks. Cognitive tasks (performed either behaviourally or in the scanner) that can sub-classify patients based on different types of brain network dysfunction have great potential value in clinical research and assessment. However, achieving this aim is non-trivial. The proposed project will develop and apply a novel neuroadaptive approach for identifying the optimal tasks that differentially engage two important frontoparietal brain networks, with the aim of developing more sensitive cognitive tasks.
There are four work packages. WP1, we will use meta-analytic approaches based on the large existing literature of FMRI studies to identify cognitive tasks that form an ordered experimental parameter space. This will be the starting point for optimisation. For demonstration purposes, we will seek to find tasks that differentiate between two important functional networks involved in cognitive control. WP2 will use closed-loop neuroadaptive Bayesian optimisation with real-time functional MRI to find the tasks that maximally dissociate between these two functional brain networks. WP3 will involve a further fine-tuning round of Bayesian optimisation to refine the task parameters of the best candidate cognitive tasks from WP2. WP4 will validate the cognitive tasks behaviourally, using internet-delivered cognitive testing on a large number of participants followed by factor-analysis. The neuroadaptive tools developed in the project and the optimised cognitive tasks will be shared with the research community and form the basis for subsequent clinical translational grant applications.
There are four work packages. WP1, we will use meta-analytic approaches based on the large existing literature of FMRI studies to identify cognitive tasks that form an ordered experimental parameter space. This will be the starting point for optimisation. For demonstration purposes, we will seek to find tasks that differentiate between two important functional networks involved in cognitive control. WP2 will use closed-loop neuroadaptive Bayesian optimisation with real-time functional MRI to find the tasks that maximally dissociate between these two functional brain networks. WP3 will involve a further fine-tuning round of Bayesian optimisation to refine the task parameters of the best candidate cognitive tasks from WP2. WP4 will validate the cognitive tasks behaviourally, using internet-delivered cognitive testing on a large number of participants followed by factor-analysis. The neuroadaptive tools developed in the project and the optimised cognitive tasks will be shared with the research community and form the basis for subsequent clinical translational grant applications.
Planned Impact
The research outputs from this project will be of benefit to the research community, clinicians, participants and the general public.
Research community: Scientists from several different fields will benefit from this research. The most immediate benefit will be to researchers who are looking to use cognitive assessment tools. Some may wish to use the set of cognitive tasks that we have optimised during the course of this project. Others may wish to use the Bayesian optimization framework to optimise their own assessment tools based on their preferred target brain states.
Clinicians: The assessment battery may be of use as a clinical diagnostic tool because it will allow cognitive impairments to be differentially assessed in a manner that relates directly to the underlying network organisation of the human brain. We already know that many patient populations that suffer from cognitive impairments tend to show abnormal function within one or other of these networks, but we currently lack behavioural tools for differentiating between them. Being able to do this is important because impairments with different underlying aetiologies will likely respond best to different treatments. A set of tools for determining these differences is particularly important in clinical populations that have heterogeneous aetiologies in order to enable the development and individualisation of therapeutic strategies. Indeed, our optimisation approach can also be applied to generate tasks that maximally differentiate between populations and the server technology will be of interest to clinical researchers who are focused on developing diagnostic and prognostic tools.
Participants: Individuals who contribute to this research will have the option to receive updates that are targeted at the lay audience when results are published in peer-reviewed journals.
General Public: We have a strong track record of taking part in public engagement events and with the media, which we will continue going forwards. For example, RL recently took part in the Wellcome Trust funded Science Streetfair that ran in parallel to the Festival of Neuroscience 2013 in the Barbican centre and the 2014 Royal Society Summer Exhibition. Similarly, AH has a strong track record of engagement with the media, with high-profile collaborations with the BBC and New Scientist, engaging 100,000s of members of the public in his research.
Research community: Scientists from several different fields will benefit from this research. The most immediate benefit will be to researchers who are looking to use cognitive assessment tools. Some may wish to use the set of cognitive tasks that we have optimised during the course of this project. Others may wish to use the Bayesian optimization framework to optimise their own assessment tools based on their preferred target brain states.
Clinicians: The assessment battery may be of use as a clinical diagnostic tool because it will allow cognitive impairments to be differentially assessed in a manner that relates directly to the underlying network organisation of the human brain. We already know that many patient populations that suffer from cognitive impairments tend to show abnormal function within one or other of these networks, but we currently lack behavioural tools for differentiating between them. Being able to do this is important because impairments with different underlying aetiologies will likely respond best to different treatments. A set of tools for determining these differences is particularly important in clinical populations that have heterogeneous aetiologies in order to enable the development and individualisation of therapeutic strategies. Indeed, our optimisation approach can also be applied to generate tasks that maximally differentiate between populations and the server technology will be of interest to clinical researchers who are focused on developing diagnostic and prognostic tools.
Participants: Individuals who contribute to this research will have the option to receive updates that are targeted at the lay audience when results are published in peer-reviewed journals.
General Public: We have a strong track record of taking part in public engagement events and with the media, which we will continue going forwards. For example, RL recently took part in the Wellcome Trust funded Science Streetfair that ran in parallel to the Festival of Neuroscience 2013 in the Barbican centre and the 2014 Royal Society Summer Exhibition. Similarly, AH has a strong track record of engagement with the media, with high-profile collaborations with the BBC and New Scientist, engaging 100,000s of members of the public in his research.
Organisations
Publications
Cole J
(2018)
Active Acquisition for multimodal neuroimaging
in Wellcome Open Research
Cole JH
(2018)
Active Acquisition for multimodal neuroimaging.
in Wellcome open research
Da Costa P
(2020)
Bayesian optimization for automatic design of face stimuli
Da Costa P
(2020)
Bayesian optimization for automatic design of face stimuli
Dafflon J
(2020)
An automated machine learning approach to predict brain age from cortical anatomical measures.
in Human brain mapping
Dafflon J
(2022)
A guided multiverse study of neuroimaging analyses.
in Nature communications
Dezhina Z
(2023)
Establishing brain states in neuroimaging data.
in PLoS computational biology
Fagerholm E
(2019)
Dynamic causal modelling of phase-amplitude interactions
Fagerholm ED
(2021)
Fine-tuning neural excitation/inhibition for tailored ketamine use in treatment-resistant depression.
in Translational psychiatry
Fagerholm ED
(2020)
Neural diffusivity and pre-emptive epileptic seizure intervention.
in PLoS computational biology
Title | Neuroimaging evidence for a network sampling theory of human intelligence |
Description | This zip file represents the data to support our paper in Nature communication titled "Neuroimaging evidence for a network sampling theory of human intelligence" |
Type Of Material | Database/Collection of data |
Year Produced | 2020 |
Provided To Others? | Yes |
URL | https://figshare.com/articles/dataset/Neuroimaging_evidence_for_a_network_sampling_theory_of_human_i... |
Title | Active Acquisition for multimodal neuroimaging |
Description | Software/analysis pipeline for real time multimodal neuroimaging analysis |
Type Of Technology | Software |
Year Produced | 2018 |
Impact | NA given recent development of the software |
URL | https://wellcomeopenresearch.org/articles/3-145/v1#ref-12 |
Title | FaceFit |
Description | Code neuroadaptively generating face stimuli based on closed loop neural activity or behavioural data using a generalised adversarial network |
Type Of Technology | Software |
Year Produced | 2021 |
Open Source License? | Yes |
Impact | Technique is being adapted for use in large scientific and potentially clinical trials, in typical neurodevelopment and related to epilepsy and neurodevelopmental disorders |
Title | Mind-the-Pineapple/into-the-multiverse: |
Description | This release includes the code used to generate all the analyses in the manuscript. |
Type Of Technology | Software |
Year Produced | 2022 |
Open Source License? | Yes |
Impact | A new approach for performing guided multiverse analyses of neuroimaging and other datasets |
URL | https://zenodo.org/record/6036978 |