Extracting Root-Causal Brain Activity Driving Psychopathology from Resting State fMRI

Neuroimaging studies of psychiatric disorders often correlate imaging patterns with diagnostic labels or composite symptom scores, yielding diffuse associations that obscure underlying mechanisms. We instead seek to identify root-causal maps – localized BOLD disturbances that initiate pathological cascades – and to link them selectively to symptom dimensions. We introduce a bilevel structural causal model that connects between-subject symptom structure to within-subject resting-state fMRI via independent latent sources with localized direct effects. Based on this model, we develop SOURCE (Symptom-Oriented Uncovering of Root-Causal Elements), a procedure that links interpretable symptom axes to a parsimonious set of localized drivers. Experiments show that SOURCE recovers localized maps consistent with root-causal BOLD drivers and increases interpretability and anatomical specificity relative to existing comparators.

💡 Research Summary

The paper tackles a fundamental limitation in psychiatric neuroimaging: most studies correlate whole‑brain functional MRI patterns with diagnostic categories or composite symptom scores, producing diffuse associations that do not pinpoint the neural origins of specific psychopathology dimensions. To address this, the authors propose a “root‑causal” framework that seeks localized BOLD disturbances that act as primary drivers of downstream network abnormalities and that can be linked selectively to symptom dimensions.

Bilevel Structural Causal Model
The authors formalize a two‑level linear non‑Gaussian structural causal model (SCM). At the between‑subject level, a vector of latent drivers (S \in \mathbb{R}^K) influences voxel‑level summaries (X \in \mathbb{R}^p) both directly (through matrix (\Gamma)) and indirectly via voxel‑to‑voxel propagation matrix (B). A subset of voxels then directly feeds into symptom measures (Y \in \mathbb{R}^q) via matrix (\Phi). The within‑subject level assumes that each subject’s voxel time series are generated by a low‑dimensional set of independent latent sources (F(t) \in \mathbb{R}^K) that act on voxels through spatially localized maps (\Gamma_i). Voxel‑to‑voxel interactions are captured by (B) but are treated as a nuisance; the key causal effect is the direct source‑to‑voxel mapping. After algebraic manipulation the model reduces to a mixing form (X(t)=F(t)M+\epsilon(t)) where (M=\Gamma(I-B)^{-1}) is a common template across subjects, enabling ICA‑based recovery of the source time courses.

SOURCE Algorithm
The proposed procedure, named Symptom‑Oriented Uncovering of Root‑Causal Elements (SOURCE), consists of three stages:

1. Latent source recovery – After standard nuisance regression and z‑scoring, all subjects’ time points are concatenated, spatially smoothed, and subjected to ICA. Noise‑like components are discarded, leaving a set of candidate sources (F).

2. Root‑proximal voxel identification – For each voxel (i), the authors compute two correlations: (a) a conditional correlation (\theta_i) between the voxel’s residualized time series and each source after regressing out a Gaussian neighborhood (to approximate local confounding), and (b) an unconditional correlation (\eta_i). The element‑wise product (\eta_i\odot\theta_i) is dense and noisy; to obtain a sparse, spatially coherent map (\zeta_j) for each source (j), they solve a regularized least‑squares problem with total‑variation (TV) and (\ell_1) penalties using iteratively re‑weighted least squares (IRLS). The absolute values (|\zeta_j|) constitute a “root‑proximal” map, highlighting voxels where the source’s direct effect survives local adjustment. Hyper‑parameters ((\lambda_1,\lambda_2)) are selected by maximizing a composite score that rewards both smoothness (low graph Laplacian roughness) and compactness (low (\ell_1/\ell_2) ratio).

3. Symptom axis learning – To link sources to clinical phenotypes, the algorithm seeks a non‑negative symptom loading vector (\alpha) (defining a symptom axis (Y\alpha)) and a sparse source weight vector (\beta) that maximizes the correlation (\text{cor}(S\beta, Y\alpha)) while penalizing the number of selected sources ((|\beta|_0)). Constraints enforce unit variance of the projected source scores and unit (\ell_1) norm of (\alpha) for interpretability. Multiple orthogonal axes are extracted iteratively by deflating the source matrix after each discovery.

Experimental Evaluation
The authors benchmark SOURCE against three representative baselines: Spatially Constrained ICA (scICA), orthogonal projective NMF (opNMF), and rank‑one dictionary learning (r1DL). They also evaluate three ablations of SOURCE (removing root‑proximal maps, removing correlation maximization, and removing both). All methods are tested on two large, publicly available rs‑fMRI datasets:

• Schizophrenia (BSNIP2) – 216 subjects with PANSS item‑level scores. SOURCE discovers a symptom axis dominated by oppositional rigidity and preoccupation (7 non‑zero PANSS items). The associated root‑proximal map localizes to the right dorsolateral prefrontal cortex (rDLPFC), a region implicated in rule maintenance and cognitive flexibility. SOURCE matches the highest held‑out Pearson correlation while substantially improving correlation density (CD) and compactness relative to baselines.

• Major Depression (EMBARC) – 281 subjects with QIDS‑SR items. SOURCE isolates an anergia/neurovegetative axis explained by a single source whose root‑proximal map peaks in the right inferior parietal lobule, consistent with frontoparietal control and attentional effort networks known to be disrupted in depression. Again, predictive accuracy is on par with the best baseline, but CD and compactness are markedly superior.

Ablation studies confirm that both the root‑proximal voxel selection and the sparse symptom‑axis optimization contribute to the gains in CD and compactness. Runtime analysis shows that SOURCE completes within 2–4 hours on standard computing hardware, demonstrating practical scalability.

Interpretation and Impact
The key methodological advance is the explicit separation of direct source‑to‑voxel effects (the causal “roots”) from indirect voxel‑to‑voxel propagation, which is treated as a nuisance. By focusing on the former, SOURCE yields anatomically precise maps that can be directly interpreted as putative neural generators of specific symptom dimensions. The use of a sparsity‑promoting symptom‑axis learning step ensures that only a minimal set of sources drives each clinical dimension, enhancing interpretability for clinicians and researchers. The introduction of quantitative spatial metrics (CD and compactness) provides an objective way to assess how well a method isolates focal neural drivers.

Limitations and Future Directions
The model assumes linear relationships and Gaussian‑like noise after ICA preprocessing, which may not capture all complexities of brain dynamics. The current implementation treats voxel‑to‑voxel propagation as a stationary linear operator; incorporating time‑varying or nonlinear propagation could improve fidelity. Extending the framework to multimodal data (e.g., structural MRI, diffusion imaging, PET) or to longitudinal designs could enable causal inference about disease progression and treatment response.

Conclusion
The paper introduces a bilevel linear non‑Gaussian SCM that links latent rs‑fMRI sources to symptom structure while relegating voxel‑to‑voxel interactions to a nuisance term. The derived SOURCE algorithm efficiently recovers candidate sources, estimates sparse root‑proximal maps, and learns interpretable symptom axes driven by a small subset of sources. Across two major psychiatric datasets, SOURCE produces anatomically localized signatures with strong predictive performance and superior interpretability compared with state‑of‑the‑art ICA, NMF, and dictionary‑learning baselines. This work provides a powerful new tool for uncovering the neural origins of complex psychiatric phenotypes and sets the stage for more precise, causally informed neuropsychiatric research.