Brain disorder prediction with dynamic multivariate spatio-temporal features: Application to Alzheimer's disease and autism spectrum disorder

Abstract

The dynamic functional connectivity (dFC) in functional magnetic resonance imaging (fMRI) is beneficial for the analysis and diagnosis of neurological brain diseases. The dFCs between regions of interest (ROIs) are generally delineated by a specific template and clustered into multiple different states. However, these models inevitably fell into the model-driven self-contained system which ignored the diversity at spatial level and the dynamics at time level of the data. In this study, we proposed a spatial and time domain feature extraction approach for Alzheimer's disease (AD) and autism spectrum disorder (ASD)-assisted diagnosis which exploited the dynamic connectivity among independent functional sub networks in brain. Briefly, independent sub networks were obtained by applying spatial independent component analysis (SICA) to the preprocessed fMRI data. Then, a sliding window approach was used to segment the time series of the spatial components. After that, the functional connections within the window were obtained sequentially. Finally, a temporal signal-sensitive long short-term memory (LSTM) network was used for classification. The experimental results on Alzheimer's Disease Neuroimaging Initiative (ADNI) and Autism Brain Imaging Data Exchange (ABIDE) datasets showed that the proposed method effectively predicted the disease at the early stage and outperformed the existing algorithms. The dFCs between the different components of the brain could be used as biomarkers for the diagnosis of diseases such as AD and ASD, providing a reliable basis for the study of brain connectomics.

Keywords: deep learning; dynamic functional connectivity; fMRI; multivariate; spatio-temporal features.

FIGURE 1

dICA-LSTM pipeline diagram. (A) Group ICA decomposes resting-state data into 20 components at the group level. The individual-level independent components are estimated by the back reconstruction. (B) Extracting time series for 15 spatially independent components after removing noise components. A sliding window with a window width of 50TR and a step size of 1TR splits the time series into m windows (W₁–W_m). The correlation of the time series within each window (FC₁–FC_m) is calculated. (C) The dynamic functional connections between independent components are used as the input of the deep learning network. The network includes an LSTM layer, two fully connected layers, a ReLu layer, a softmax layer, and a classification layer.

FIGURE 2

Reliability assessment in the independent component decomposition process of two datasets. (A) The reliability assessment of the ADNI dataset. (B) The reliability assessment of the ABIDE dataset. The black dots represent the estimated ICs in single run. The numbers of the clusters are labeled by the stability index, and minimum and maximum cluster size. The compact and isolated clusters indicate the reliable estimation. The different colors in the right bar indicate the similarity between different clusters. The darker red color (0.9–1) indicates the higher similarity.

FIGURE 3

Spatial components obtained by the GICA for two databases. (A) The spatial components of the ADNI dataset. (B) The spatial components of the ABIDE dataset. There are 15 meaningful independent components in the GICA process after eliminating obvious noise components.

FIGURE 4

Standard deviation comparison of classification accuracy in 10-fold cross-validation for the dynamic ROI and dynamic ICA methods. The standard deviations of the accuracy with dynamic ICA methods are smaller than that of the dynamic ROI methods during cross-validation.

FIGURE 5

ROC curves for different groups and different methods. (A) The ROC curves for different groups of the ADNI dataset. The highest area under the curve (AUC) is obtained in the AD vs. NC task. (B) The ROC curves for different methods of the ABIDE dataset. The dICA-LSTM method obtained the best AUC in the classification task.

FIGURE 6

Classification performance comparison of three methods including random forest, dynamic ROI levels, and dynamic ICA levels. The performance of the dynamic features outperforms the static features. Moreover, the performance of the dynamic features at the ICA level outperforms the dynamic features at the ROI level. (RF: random forest, dROI-LSTM: dynamic features at the ROI level combined with LSTM, dICA-LSTM: dynamic features at the ICA level combined with LSTM).

FIGURE 7

Visualization of the dynamic functional connectivity with width 50 and step size of 1. The functional connections between different resting-state networks in the sliding window are dynamic which have depth-time characteristics. (From left to right: the 10th, 20th, 30th, 40th, 50th, 60th, 70th, and 80th window).

FIGURE 8

Brain mapping between the independent components and the ROIs on the AAL template. (A) Brain activation areas of the ADNI dataset. (B) Brain activation areas of the ABIDE dataset. The resting-state networks obtained by blind source analysis are mapped to the delineated ROIs, indicating the reliability of the functional networks obtained by ICA. The resting-state network ensures the homogeneity within the network and the heterogeneity between the networks compared to the specific ROI.

FIGURE 9

Effect of different parameters on the accuracy. (A) Classification performance with different number of components. (B) Classification performance with different window widths. (C) Classification performance with different step lengths. The best performance is obtained with an IC of 20, a window width of 50, and a step length of 1.