Data Integration with Fusion Searchlight: Classifying Brain States from Resting-state fMRI

TL;DR

This work presents Fusion Searchlight (FuSL), a framework that fuses multiple resting-state fMRI metrics—regional homogeneity (ReHo), fractional amplitude of low frequency fluctuations (fALFF), and functional connectivity efficiency (FCE)—within a searchlight decoding pipeline to improve brain-state classification and pharmacological effect detection. It demonstrates that joint metric integration enhances decoding performance and spatial specificity, aided by SHAP-based explainability that localizes metric contributions and reveals how alprazolam modulates different rs-fMRI markers across networks. The approach is validated on both synthetic data and a pharmacological rs-fMRI study, showing notable gains when combining metrics and providing interpretable maps of metric-driven effects. FuSL offers a flexible, model-agnostic, and potentially multimodal tool for data fusion in neuroimaging with broad applicability to clinical and cognitive neuroscience research.

Abstract

Resting-state fMRI captures spontaneous neural activity characterized by complex spatiotemporal dynamics. Various metrics, such as local and global brain connectivity and low-frequency amplitude fluctuations, quantify distinct aspects of these dynamics. However, these measures are typically analyzed independently, overlooking their interrelations and potentially limiting analytical sensitivity. Here, we introduce the Fusion Searchlight (FuSL) framework, which integrates complementary information from multiple resting-state fMRI metrics. We demonstrate that combining these metrics enhances the accuracy of pharmacological treatment prediction from rs-fMRI data, enabling the identification of additional brain regions affected by sedation with alprazolam. Furthermore, we leverage explainable AI to delineate the differential contributions of each metric, which additionally improves spatial specificity of the searchlight analysis. Moreover, this framework can be adapted to combine information across imaging modalities or experimental conditions, providing a versatile and interpretable tool for data fusion in neuroimaging.

Figures (17)

• Figure 1: Individual steps of the FuSL framework. (A) In resting-state fMRI we can observe a BOLD signal exhibiting complex spatial and temporal dynamics. (B) Using different metrics we can characterize different aspects of these neural dynamics. Regional homogeneity (ReHo) quantifies the coherence of the BOLD signal within local neighborhoods across the cortex. Fractional amplitudes of low frequency fluctuations (fALLF) are defined as the ratio of the low-frequency power to the power of the entire frequency range for each vertex. Functional connectivity can be used to characterize functional connection efficiency (FCE) of individual ROIs in the global brain network. (C) In our FuSL framework these different metrics are concatenated to identify brain regions that are informative for decoding a brain state. (D) We used Shapley additive explanations (SHAP) to retrospectively reconstruct the impact of each metric on the decoding at a specific location.
• Figure 2: Investigating basic properties of the FuSL framework using an artificial dataset. We induce artificial signals in two ROIs (A), whereby signals of source 1 and 2 overlap in ROI 1 and signals in source 3 are only present in ROI 2 (B). (C) We then generate samples by randomly varying the signal amplitude and adding Gaussian noise. (D) We observed higher decoding test accuracy values in ROI 1 than in ROI 2. (E) Impact values of source 1 and 2 are higher in ROI 1 than ROI 2, and of source 3 higher in ROI 2 than ROI 1. (F) The feature-weighted impact displays increased values in source 1 and 3 and the decrease of values in source 2. (G) and (H) show decoding accuracies for all combinations of sources in ROI 1 and 2 respectively. (I) Decoding accuracies averaged across ROI 1 and 2. Error bars represent 95% confidence intervals across folds. Significant differences of accuracies in figure (G), (H) and (I) are indicated with: *: $p \leq 0.05$, **: $p \leq 0.01$, ***: $p \leq 0.001$, ****: $p \leq 0.0001$, ns: not significant, (ns): not significant after false discovery rate correction.
• Figure 3: Statistical power of the decoding analysis in ROI 1. Adding the informative source 2 to source 1 increased the power, especially when the sample size is smaller, whereas adding the uninformative source 3 led to a decrease in statistical power. Errorbars represent 95% confidence intervals across dataset subsampling repetitions.
• Figure 4: Increasing spatial specificity of searchlight decoding using Shapley values. (A) We induce a signal in one vertex only. (B) The area of significant decoding accuracy is considerably larger than the original signal. (C) Analyzing the feature impact maps allows us to reconstruct again the exact location of an informative vertex.
• Figure 5: Decoding performance of FuSL in dependency of rs-fMRI metric combinations. The combinations fALFF$||$FCE, fALFF$||$ReHo and ReHo$||$fALFF$||$FCE achieved highest average decoding test accuracies (A), highest threshold-free cluster enhanced accuracy (B) and largest number of significant vertices (C). (D) Local differences in decoding test accuracies between ReHo$||$fALFF$||$FCE and all other combinations in the left (L) and right (R) cortex. (E) Locations where only ReHo$||$fALFF$||$FCE yielded significant decoding accuracies are highlighted in red. Locations where only the respective other combination is able to decode are depicted in blue. Locations where ReHo$||$fALFF$||$FCE and the respective other combination can both significantly decode are marked in white. Significant differences of accuracies in figure (A) are indicated with: *: $p \leq 0.05$, **: $p \leq 0.01$, ***: $p \leq 0.001$, ns: not significant.