Neural Dynamics-Informed Pre-trained Framework for Personalized Brain Functional Network Construction

TL;DR

A neural dynamics-informed pre-trained framework is proposed for personalized brain functional network construction that extracts personalized representations of neural activity patterns in heterogeneous scenarios and achieves superior performance in heterogeneous scenarios.

Abstract

Brain activity is intrinsically a neural dynamic process constrained by anatomical space. This leads to significant variations in spatial distribution patterns and correlation patterns of neural activity across variable and heterogeneous scenarios. However, dominant brain functional network construction methods, which relies on pre-defined brain atlases and linear assumptions, fails to precisely capture varying neural activity patterns in heterogeneous scenarios. This limits the consistency and generalizability of the brain functional networks constructed by dominant methods. Here, a neural dynamics-informed pre-trained framework is proposed for personalized brain functional network construction. The proposed framework extracts personalized representations of neural activity patterns in heterogeneous scenarios. Personalized brain functional networks are obtained by utilizing these representations to guide brain parcellation and neural activity correlation estimation. Systematic evaluations were employed on 18 datasets across tasks, such as virtual neural modulation and abnormal neural circuit identification. Experimental results demonstrate that the proposed framework attains superior performance in heterogeneous scenarios. Overall, the proposed framework challenges the dominant brain functional network construction method.

Figures (6)

• Figure 1: Overview of the Personalized Brain Functional Network Construction Framework.a, Personalized brain functional network construction framework can be adapted to fMRI acquired across variable and heterogeneous scenarios such as subjects' age groups, brain disorder types, linguistics, and image acquisition strategies. b, The proposed framework begins with a foundation model pretrained on large-scale fMRI data. The model is then fine-tuned with neural dynamics information to extract personalized representations from neural activities. These representations are used to guide the brain region parcellation and the correlation estimation to construct the personalized brain functional network.
• Figure 2: Consistency evaluation of personalized brain functional networks.a–c, Box plots of PDiv distributions comparing the personalized brain functional network constructed by our method (blue) with brain functional network constructed by the baseline Micapipe method (pink) across heterogeneous scenarios. Results demonstrate that our method achieves significantly higher median PDiv values across (a) varying age groups , (b) image acquisition strategies , and (c) brain disorder types, consistently outperforming the baseline method. d, Regression analysis of inter-representation similarity versus brain functional network consistency. Results demonstrate that personalized representations are the underlying mechanism for enhanced brain functional network consistency. e, t-SNE visualization of personalized brain functional network features for different disorders (ASD, AD, MDD, ADHD). Results demonstrate that the high consistency of the personalized brain functional network is achieved with ignoring the distinct discriminative features of neural activity patterns.
• Figure 3: Evaluation of personalized brain functional networks in multiple diagnosis and prediction tasks.a, Box plots comparing diagnostic performance metrics (Accuracy, Sensitivity, Precision, F1) across multiple brain disorder (AD, PD, MDD, ADHD, ASD). Results demonstrate that our method consistently outperforms the baseline method across all diagnosis tasks. b, c, Comparison of $R^2$ for physiological index prediction. Results demonstrate that our method achieves significantly higher predictive accuracy for (b) age and (c) BMI. d, e, Performance evaluation for (d) motor imagery decoding and (e) individual identification. Results demonstrate that our method significantly surpasses the baseline method.
• Figure 4: Evaluation of personalized brain functional networks via virtual neural modulation across three brain disorders.a, Schematic illustration of the virtual neural modulation. The process begins by simulating brain activity states using a neural network. Next, perturbations are applied to the optimal targets identified by SHAP analysis. Finally, a "Virtual Doctors" classifier is applied to diagnose whether patients have recovered. b, Visualization of neural modulation targets for PD based on the TaoWu dataset. The heatmap on the left shows the importance rankings of different brain regions as neural modulation targets. The key high-priority targets are highlighted in red. The brain surface maps on the right illustrate the spatial distribution differences between the core targets identified by our method and the baseline method. Results demonstrate that our method identifies targets more consistent with empirical clinical evidence compared with the baseline method. c, The box plots show recovery rates after virtual perturbation across nine independent datasets for PD, ASD, and MDD. Results demonstrate that the proposed method achieves significantly higher recovery rates across diverse datasets compared with the baseline method. d, The scatter plot shows the change in correlation of mean functional connectivity (Mean FC) between the MDD patient group and healthy controls. The left panel depicts the state before perturbation, and the right panel shows the state after perturbation. Results demonstrate that virtual modulation effectively transform pathological brain activity towards healthy patterns.
• Figure 5: Abnormal neural circuit identification and consistency analysis across datasets with different image acquisition strategies.a, Box plots illustrating cosine similarity between abnormal neural circuits identified in AD datasets with different acquisition strategies. Results demonstrate that the abnormal neural circuits identified by our method exhibit significantly higher consistency across AD datasets with different acquisition strategies compared to the baseline method. b, Box plots illustrating cosine similarity between abnormal neural circuits identified in ADHD datasets with different acquisition strategies. Results demonstrate that the abnormal neural circuits identified by our method exhibit significantly higher consistency across ADHD datasets with different acquisition strategies compared to the baseline method. c, d, Topographical visualization of the top 5 highest-weighted abnormal connections for (c) AD and (d) ADHD datasets with different acquisition strategies. Results demonstrate the robust visual consistency of neural circuits identified by our method across datasets with different acquisition strategies. e, Comparative analysis of neural circuit complex network metrics as a function of connectivity threshold, focusing on characteristic path length, clustering coefficient, and global efficiency. Results demonstrate highly consistent trends between datasets with different acquisition strategies.