Transferring fMRI Fingerprints from Big to Small Data Improves Trait Prediction
Poster No:
1435
Submission Type:
Abstract Submission
Authors:
Qianwen Wu1, Mark Woolrich2, Stephen Smith1, Rezvan Farahibozorg1
Institutions:
1FMRIB, WIN, Nuffield Dept. of Clinical Neuroscience, University of Oxford, Oxford, UK, 2OHBA, Wellcome Centre for Integrative Neuroimaging, Department of Psychiatry, University of Oxford, Oxford, UK
First Author:
Co-Author(s):
Mark Woolrich
OHBA, Wellcome Centre for Integrative Neuroimaging, Department of Psychiatry, University of Oxford
Oxford, UK
OHBA, Wellcome Centre for Integrative Neuroimaging, Department of Psychiatry, University of Oxford
Oxford, UK
Rezvan Farahibozorg
FMRIB, WIN, Nuffield Dept. of Clinical Neuroscience, University of Oxford
Oxford, UK
FMRIB, WIN, Nuffield Dept. of Clinical Neuroscience, University of Oxford
Oxford, UK
Introduction:
Fingerprints extracted from resting-state functional MRI (rfMRI) can be used to predict personalised traits and diseases. Recent advances in big data neuroimaging, e.g. UK Biobank (UKB) with expected 100,000 subjects, are facilitating this substantially. However, clinical studies still rely on small data, where building accurate and generalisable predictive models for patient groups remains highly challenging. Recent techniques, such as meta-matching (He, 2022), aimed to address this for scenarios where trait similarities can be used to transfer trait prediction weights between big and small data. However, such similarities may not always exist. Here we instead propose transfer learning techniques that directly transfer fingerprints, i.e. brain features that are independent of target traits, from big to small data. We show that while our approach is unaware of specific traits, it still yields improvement for trait prediction in small datasets.
Methods:
Data: Both big and small datasets were selected from UKB data. The big dataset consisted of 20,000 subjects. To mimic small studies, we randomly selected groups of 15, 50, 100, 200, and 1000 subjects, that were treated as new subjects independent of big data in the analyses. The random selection was repeated 100 times to obtain confidence intervals.
fMRI Analysis: The framework of Probabilistic Functional Modes (PFMs)(Farahibozorg, 2021) was used to extract individual-specific resting state fMRI networks (RSNs). Spatial maps of RSNs were used as bases for fingerprint extraction.
Fingerprint identification: RSN modelling and feature extraction were conducted by 4 different methods: 1) PFMs were trained independently on small data, spatial maps were dimension reduced to subject-specific fingerprints using SVD followed by ICA; 2) PFMs were trained on big data, and small data PFMs were modelled based on priors from these "pre-trained" PFMs. SVD-ICA were applied to small data for fingerprint extraction; 3) PFMs and SVD-ICA were trained on big data, and both pretrained PFMs and SVD-ICA sources were transferred (Fig 1b); 4) PFMs were trained on big data, and FLICA was used for fingerprint extraction (Gong, 2021). FLICA is a Bayesian ICA approach that finds consensus components across all RSNs instead of simple concatenation done in ICA, and also doing dimension reduction. Pretrained PFMs and pretrained FLICA were transferred (Fig 1b).
Trait prediction: The 4 sets of fingerprints were used as input to unregularized linear regression with 5-fold cross validation to predict age, sex, bone health, cardiovascular measures, mental health, alcohol, tobacco and cognition. Unpaired t-tests were used to compare the prediction accuracy of methods.
fMRI Analysis: The framework of Probabilistic Functional Modes (PFMs)(Farahibozorg, 2021) was used to extract individual-specific resting state fMRI networks (RSNs). Spatial maps of RSNs were used as bases for fingerprint extraction.
Fingerprint identification: RSN modelling and feature extraction were conducted by 4 different methods: 1) PFMs were trained independently on small data, spatial maps were dimension reduced to subject-specific fingerprints using SVD followed by ICA; 2) PFMs were trained on big data, and small data PFMs were modelled based on priors from these "pre-trained" PFMs. SVD-ICA were applied to small data for fingerprint extraction; 3) PFMs and SVD-ICA were trained on big data, and both pretrained PFMs and SVD-ICA sources were transferred (Fig 1b); 4) PFMs were trained on big data, and FLICA was used for fingerprint extraction (Gong, 2021). FLICA is a Bayesian ICA approach that finds consensus components across all RSNs instead of simple concatenation done in ICA, and also doing dimension reduction. Pretrained PFMs and pretrained FLICA were transferred (Fig 1b).
Trait prediction: The 4 sets of fingerprints were used as input to unregularized linear regression with 5-fold cross validation to predict age, sex, bone health, cardiovascular measures, mental health, alcohol, tobacco and cognition. Unpaired t-tests were used to compare the prediction accuracy of methods.
·Figure 1: Method summary flowchart.
Results:
Method 3&4, where both pretrained-PFMs and fingerprints were transferred, showed improvement over small-data-only for age, sex, bone and cardiovascular health prediction (Fig 2). This was especially evident in big transferred FLICA, where age prediction accuracy was doubled in some cases. The combination of hierarchical population modelling in PFMs and FLICA's ability for capturing linked information between RSNs likely contributes to this result. Method 2, where only population priors but no fingerprints were transferred, yielded similar performance to independent small data modelling.
Comparing across group sizes, smaller groups were found to benefit more from big-transfer methods, especially for N=50 and 100. The improvement was still significant even when the group size is up to 1000. For small data with N=15, where a breakdown in cross-validated regression is seen generally, methods 3&4 improve the mean prediction accuracy across random trials, but the variance is large.
Comparing across group sizes, smaller groups were found to benefit more from big-transfer methods, especially for N=50 and 100. The improvement was still significant even when the group size is up to 1000. For small data with N=15, where a breakdown in cross-validated regression is seen generally, methods 3&4 improve the mean prediction accuracy across random trials, but the variance is large.
·Figure 2: Prediction accuracy comparison between different feature extraction methods.
Conclusions:
We propose new transfer learning techniques to extract fMRI fingerprints in small datasets using knowledge learnt from big fMRI data. Crucially, while our transfer methods are fully unsupervised (uninformed of target traits), they still significantly improved trait prediction accuracies.
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Connectivity (eg. functional, effective, structural)
fMRI Connectivity and Network Modeling 1
Methods Development
Novel Imaging Acquisition Methods:
BOLD fMRI
Keywords:
FUNCTIONAL MRI
Machine Learning
Modeling
Other - functional connectivity, big data fMRI, transfer learning, PROFUMO, phenotype prediction
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
Gong, W.(2021). Phenotype discovery from population brain imaging. Medical Image Analysis, 71, 102050.
He, T. (2022). Meta-matching as a simple framework to translate phenotypic predictive models from big to small data. Nature Neuroscience, 25(6), 795–804.