Does brain structure record the inflammation that blood misses?

Inflammation inside the brain often looks nothing like inflammation in the blood. The brain sits behind barriers, so immune signals can build up in the cerebrospinal fluid (CSF, the liquid that bathes the brain and spinal cord) while blood tests look normal. In three patients with SARS-CoV-2 encephalitis, the CSF held a local "cytokine storm" of the inflammatory messenger proteins IL-6, IL-8 and IL-10, and the patients were making antibodies inside the nervous system. Yet no virus was found in the CSF, and several patients had normal CSF protein (Encephalopathy and Encephalitis Associated with Cerebrospinal Fluid Cytokine Alterations and Coronavirus Disease, Atlanta, Georgia, USA, 2020). Measuring this central inflammation requires a lumbar puncture, an invasive procedure that is rarely done. Structural MRI, by contrast, is routine. This proposal asks a narrow question: do routine structural scans carry information about the part of CSF inflammation that blood cannot explain? We call that part the intrathecal excess, meaning inflammation produced inside the central nervous system. The question is not about CSF inflammation in general, much of which simply leaks in from the blood.

The tool is source-based morphometry (SBM). SBM applies independent component analysis, a method that splits many brain scans into a small set of spatial patterns that vary independently across people. Each person then gets one "loading" per pattern, which says how strongly that pattern is expressed in their brain. We use a constrained version, in which a prior map steers one pattern toward a chosen anatomy. Here the prior covers the choroid plexus, the tissue inside the brain's fluid-filled cavities (the ventricles) that forms the blood–CSF barrier, together with the ventricles themselves and the thalamus, hippocampus and putamen. The claim is that a person's loading on this barrier-anchored pattern predicts CSF cytokine excess after paired blood levels and barrier leakage are accounted for. It should not predict blood cytokines once CSF levels are known. If both halves hold, structural MRI becomes a non-invasive read-out of compartmentalized neuroinflammation, something the immunology literature says does not yet exist.

What is new relative to existing work

The closest work predicts CSF amyloid and tau status from brain volumes. It segments clinical T1-weighted scans with SynthSeg+ into regions, builds group-level structural covariance networks, and reaches an area under the ROC curve (AUROC) of 0.795 for amyloid with ridge regression and 0.694 for tau with a random forest on the Mass General Brigham (MGB) cohort (Brain MRI signatures across sex and CSF Alzheimer's disease biomarkers). That paper motivates us, because it shows that brain structure encodes CSF chemistry. It differs from our proposal in three ways. It targets neurodegenerative proteins rather than immune markers. It never pairs CSF with blood, so it cannot separate local from systemic signal. And it ignores the choroid plexus.

Two choroid plexus studies supply the anatomical hypothesis. In schizophrenia-spectrum patients, average choroid plexus volume did not differ from controls, but its spread across people did. A subgroup with large choroid plexus and large ventricles had the worst cognition and higher blood monocyte counts. Larger choroid plexus also went with larger hippocampus, putamen and thalamus, but only in patients (Investigation of choroid plexus variability in schizophrenia-spectrum disorders—insights from a multimodal study). A second study of antipsychotic-naïve first-episode patients found a modest mean enlargement (1630.65 vs 1498.21 mm³, p = 0.027) and no link to symptoms (Choroid plexus volume enlargement in first-episode antipsychotic-naïve schizophrenia).

Together these suggest that the useful signal is a coordinated pattern that varies from person to person, running from the choroid plexus through the ventricles to subcortical structures, rather than a single volume. Neither study measured CSF immune markers, and neither decomposed the pattern with multivariate methods. The methods themselves come from two sources. Constrained independent component analysis follows Constrained Source-Based Morphometry Identifies Structural Networks Associated with Default Mode Network, which used a functional-network prior to extract coupled gray-matter and white-matter patterns. The federated version follows dcSBM: A federated constrained source-based morphometry approach for multivariate brain structure mapping, which matched pooled-data loadings across sites in UK Biobank, fBIRN and COBRE while sharing only mixing matrices and source maps. So far it has been tested only against age, sex and diagnosis.

Core Claim

The target is defined once. For subject $i$ and cytokine $k$, we fit a reference model on the log CSF concentration $C^{csf}_{ik}$ using the log serum concentration $C^{ser}_{ik}$ and the log CSF/serum albumin ratio $Q_i$. The albumin ratio is the standard index of barrier leakage. The intrathecal excess is the residual:

$$ e_{ik} = \log C^{csf}_{ik} - \left(a_k + b_k \log C^{ser}_{ik} + c_k \log Q_i\right). $$

Here $a_k, b_k, c_k$ are fit only within training folds. This mirrors the logic of the classical IgG index used to detect antibody production inside the central nervous system. The claim is directional. The barrier-anchored SBM loading $\ell_i$ should explain out-of-sample variance in $e_{ik}$ for IL-6, IL-8 and IL-10, the cytokines implicated in the Atlanta series. Conversely, $\ell_i$ should explain little variance in $\log C^{ser}_{ik}$ once $\log C^{csf}_{ik}$ is known. This asymmetry is the signature of compartmentalization. A generic "inflamed brain" signal would predict blood and CSF alike.

Research Program

The first stage is a methods check on data that already exists. Before touching cytokines, we verify that barrier-anchored constrained SBM on clinical T1 scans is a competent representation. On the MGB-style setting, with clinical T1 and CSF amyloid measured within a year, we compare its loadings against the SynthSeg+ region-plus-ridge pipeline on the published amyloid and tau tasks, using the same subjects and folds. We do not need SBM to beat 0.795 here. We need it to come within the published confidence interval ([0.788, 0.802]) while using a handful of loadings instead of a full set of regional volumes. If it falls far short, the representation is inadequate and the rest of the program would switch to regional volumes.

The second stage builds the paired cohort that the question needs and that the file does not contain. We would assemble patients who had a clinically indicated lumbar puncture with same-day serum and a structural MRI within a short window. Plausible sources are neurology, psychiatry and post-infectious encephalopathy services. We would measure a multiplex cytokine panel and albumin in both fluids. Because each site will contribute few such patients, the decomposition runs federated with dcSBM, and only mixing matrices and source maps leave each site. This turns the privacy barrier from a limitation into part of the design.

With about 300 pooled subjects, a Fisher-z power calculation gives 80% power to detect a correlation near 0.16 at α = 0.05. That is small enough to matter clinically, but only if the effect is not diluted across many tests. We therefore pre-register the three cytokines and a single barrier component.

The third stage asks what the component is. We compare its spatial map with the patients-only choroid plexus–subcortical coupling reported in the schizophrenia study. We test whether the component's loading dispersion reproduces the increased variability reported there. We also test whether splitting subjects by loading reconciles the conflicting enlargement results between the two choroid plexus studies.

Baselines

The naive reference point predicts $e_{ik}$ from age, sex, intracranial volume and site alone. Every imaging model must beat it, because age and head size confound both ventricular size and cytokine levels.

The strongest existing method is the pipeline of Brain MRI signatures across sex and CSF Alzheimer's disease biomarkers: SynthSeg+ regional volumes fed to ridge logistic or linear regression, with a random forest as the second model. Its best reported number is an AUROC of 0.795 for CSF amyloid status on MGB. We found no reported result on Lacuna for predicting CSF cytokines, or intrathecal cytokine excess, from MRI. It is the strongest competitor because it is whole-brain, already tuned to a CSF target, and interpretable. We rerun it on identical subjects and nested folds, give it the same hyperparameter search budget as our model, and add choroid plexus volume to its regional inputs so that it is not handicapped on our anatomy. Beating it means a higher out-of-sample Spearman correlation with $e_{ik}$, and a higher AUROC for classifying the top tertile of excess, with a paired bootstrap interval that excludes zero.

The ablation that isolates the key ingredient, the barrier prior, is unconstrained SBM with the same number of components. A second ablation is the same constrained SBM targeting raw CSF concentration instead of $e_{ik}$. This shows whether the leakage correction changes which component carries the signal.

Why not just…?

Why not just regress CSF cytokines on automated choroid plexus volume? The schizophrenia-spectrum study found no mean volume difference. What it did find was variability and coupling with subcortical structures that appeared only in patients, which is a multi-region pattern that a single volume cannot express ( Investigation of choroid plexus variability in schizophrenia-spectrum disorders—insights from a multimodal study). Choroid plexus volume is also strongly tied to ventricle size (Choroid plexus volume enlargement in first-episode antipsychotic-naïve schizophrenia Choroid plexus volume enlargement in first-episode antipsychotic-naïve schizophrenia), so its signal may simply be atrophy. We include it as a baseline, alone and with ventricular volume. If it matches the component, we report that the cheaper marker suffices.

Why not just use serum cytokines, which are cheap? By construction, serum is a covariate in $e_{ik}$, so this alternative cannot address our target. It is still the right question for clinical utility. We therefore also report how well serum alone predicts raw CSF levels. If that prediction is already near ceiling, the compartmental excess is small and the whole premise weakens. The Atlanta series ( Encephalopathy and Encephalitis Associated with Cerebrospinal Fluid Cytokine Alterations and Coronavirus Disease, Atlanta, Georgia, USA, 2020) argues otherwise, but with n = 3.

Why not just use regional volumes with ridge regression, as the Alzheimer's study did? This is the real threat. On a small cohort, a regularized regional model may match any decomposition. We treat it as a coequal contender. If it ties, the scientific finding, that structure encodes intrathecal excess and not serum, still stands, and SBM's remaining advantage is that it can be federated with dcSBM. We would state that plainly rather than claim an accuracy gain.

Why not just build individual morphometric similarity networks? These networks need several MRI sequences and are sensitive to thresholding. Clinical lumbar-puncture cohorts usually have only a T1 scan. We keep this alternative as a secondary analysis for the subset of subjects with multiple sequences.

Decisive Checks

The idea survives only if three things hold. First, the barrier component predicts $e_{ik}$ out of sample better than the demographic reference. Second, it does not predict serum cytokines given CSF levels, which is the asymmetry test. Third, both results replicate in held-out sites under federated fitting, with loadings correlating highly with a centralized refit, as dcSBM reported for diagnosis.

Two negative controls guard against confounds. The component should not predict CSF amyloid excess any better than the Alzheimer's pipeline does, which checks that it is not just a neurodegeneration axis. Its association with $e_{ik}$ should survive adjustment for total ventricular volume, which checks that it is not just atrophy. If the component predicts serum as well as it predicts CSF excess, the compartmental interpretation fails, and the result reduces to "structure tracks general inflammation."