Disentangling global brain states from local activity for better classification accuracy
Poster No:
1947
Submission Type:
Abstract Submission
Authors:
Eloy Geenjaar1, Vince Calhoun2
Institutions:
1Georgia Institute of Technology, Atlanta, GA, 2GSU/GATech/Emory, Decatur, GA
First Author:
Co-Author:
Introduction:
Resting-state magnetic resonance imaging (rs-fMRI) is often used to track whole-brain connectivity dynamics [1]. These connectivity dynamics are often computed by dividing the signal into (overlapping) windows and computing the correlation between brain regions over that window. These connectivity matrices effectively summarize the activity in each window. We introduce a more general summarization of the activity in each window using a variational autoencoder (VAE) that summarizes the activity in a window into a low-dimensional representation for the whole window (global) and each timestep in the window (local) [2]. We show that latent space of the global representations is more easily separable into windows belonging to schizophrenia patients and controls than models that only use local information.
Methods:
As shown in Figure 1, our model consists of two main sub-models: a global network that summarizes a window of fMRI activity into a low-dimensional global representation, and a local network that uses the global representation and the fMRI activity to summarize each individual timepoint. Just as with dynamic connectivity analyses, we expect that the rs-fMRI signal is non-stationary over time and that this non-stationarity affects the individual timepoints within the window. In the same way, we expect the global representations to capture the underlying brain state a person is in, which influences the latent representation for individual timesteps. By disentangling global and local representations, we expect to find brain states (global representations) that are more common for schizophrenia patients than controls. To explore this research question, we train our model on 53 NeuroMark [4] ICA component timeseries from the fBIRN [5] dataset with 40 second windows (20 seconds of overlap) and compare it to dynamic functional connectivity (dFNC) and a model that only encodes local representations. The local and global representations in our model are both 2-dimensional, and we compare against a model with only 2-dimensional local representations. We train our model and the local-only model across 4 seeds so we can perform a statistical analysis. To evaluate our model against the baselines, we calculate a classification accuracy across windows in the test set to see which method can separate individual windows the best into belonging to schizophrenia patients or controls. Classification is done using a support vector machine with an RBF kernel. To get a representation for a window in the local-only model, we average the latent timesteps in each window, we only use the global representation for the classification in our proposed model and reduce the lower-triangular part of each dFNC matrix to a 2-dimensional vector with PCA. Thus, classifications for our model and the baselines are done in a 2-dimensional space.
·An abstract depiction of our model, each recurrent neural network is a GRU [3]
Results:
Our experiment in Figure 2a shows that our model significantly (p=2.16E-5) outperforms the local-only baseline, and in Figure 2b, we show how well our model separates out schizophrenia patient windows from control windows. The global and local representations in Figure 2b are both derived from our model.
·Figure 2a shows the window classification results, and figure 2b shows an example of global and local representations results for our model
Conclusions:
Our results conclude that disentangling local and global representations is a potentially fruitful direction to understand how brain states (global representations) are different for people with a schizophrenia diagnosis as opposed to controls. We compare our model against a baseline, and we find significant differences. In the future, we want to visualize and interpret what these differences look like in the brain. Given that our model is generative, we can ask the question: what would activity in a certain window have looked like if this person was diagnosed with schizophrenia?
Disorders of the Nervous System:
Psychiatric (eg. Depression, Anxiety, Schizophrenia) 2
Modeling and Analysis Methods:
Activation (eg. BOLD task-fMRI)
Classification and Predictive Modeling
Connectivity (eg. functional, effective, structural)
Methods Development 1
Keywords:
Computational Neuroscience
Data analysis
DISORDERS
FUNCTIONAL MRI
Modeling
1|2Indicates the priority used for review
Provide references using author date format
[2] Yingzhen, L., & Mandt, S. (2018, July). Disentangled sequential autoencoder. In International Conference on Machine Learning (pp. 5670-5679). PMLR.
[3] Chung, J., Gulcehre, C., Cho, K., & Bengio, Y. (2014). Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555.
[4] Du, Y., Fu, Z., Sui, J., Gao, S., Xing, Y., Lin, D., ... & Alzheimer's Disease Neuroimaging Initiative. (2020). NeuroMark: An automated and adaptive ICA based pipeline to identify reproducible fMRI markers of brain disorders. NeuroImage: Clinical, 28, 102375.
[5] Keator, D. B., van Erp, T. G., Turner, J. A., Glover, G. H., Mueller, B. A., Liu, T. T., ... & Potkin, S. G. (2016). The function biomedical informatics research network data repository. Neuroimage, 124, 1074-1079.