The Insula, Read Subregion by Subregion, Tracks Bipolar Disorder Across Treatment
- The insula was split into 12 anatomical subregions rather than treated as one block, and at baseline six of those subregions in 82 patients with bipolar disorder showed altered connectivity to the dorsolateral prefrontal cortex, orbitofrontal cortex, thalamus, default mode network, and visual network compared with 88 healthy controls.
- After three months of treatment in the 43 patients who completed follow-up, connectivity between the insula and the superior frontal gyrus partly normalised, while connectivity with the default mode and visual networks rose further and a new link to the superior temporal gyrus emerged.
- A support vector machine trained on insular subregion connectivity separated patients from controls with 73.75 percent accuracy, suggesting these features carry a usable signal rather than noise.
- The pattern of partial normalisation and partial divergence implies that some connectivity changes track the clinical state while others may be more trait-like and persist despite symptom improvement.
Bipolar disorder has long resisted a clean neural signature. Studies disagree, partly because the brain regions implicated are large and functionally mixed, and the insula is a textbook example. It is not one structure doing one job; its anterior portion sits at the hub of salience and emotional appraisal, while posterior zones handle interoception and sensory integration. Treating it as a single region of interest, as much earlier imaging work did, averages these distinct roles together and can wash out exactly the alterations a study is hunting for.
The team from the Second Xiangya Hospital of Central South University took the obvious but underused step of parcellating the insula into 12 subregions using the Human Brainnetome Atlas, then mapping whole-brain voxel-based functional connectivity from each subregion separately. With 82 patients and 88 controls scanned at rest, and a subset of 43 patients rescanned after three months of treatment, the design allowed two questions that a single cross-sectional snapshot cannot: which connectivity differences are present at baseline, and which ones move as the patient improves.
The baseline picture is consistent with the broader literature but sharper. Six subregions showed abnormal connectivity, and the targets were the usual suspects of mood regulation: dorsolateral prefrontal cortex, orbitofrontal cortex, thalamus, plus nodes of the default mode and visual networks. What makes the follow-up interesting is that the changes did not simply reverse. Insula-to-superior-frontal-gyrus connectivity partly normalised, which is what one would expect if it indexes acute symptom burden. But insular coupling with the default mode and visual networks increased further over treatment, and a connection to the superior temporal gyrus appeared that had not been flagged at baseline.
That dissociation is the substantive message. A purely state-dependent biomarker should track recovery in one direction; here, part of the system moved toward the control pattern while part moved away from it. The most parsimonious reading is that insular connectivity in bipolar disorder is a mixture of state and trait components, and that lumping the insula together had been hiding this by averaging opposing trends. The 73.75 percent classifier accuracy adds weight: the subregion features are structured enough for a machine to learn a boundary, even if that accuracy is well short of clinical decision-making.
For the field, the practical takeaway is methodological as much as clinical. If the insula is going to earn its place in bipolar neuroscience, it has to be read at the resolution at which its functions actually separate.
Why subregion resolution matters
Functional neuroimaging in psychiatry has a reproducibility problem, and coarse anatomy is one quiet contributor. When a region with several jobs is summarised by a single average time series, opposing effects in its parts cancel, and the result looks null or unstable across samples. By segmenting the insula and following each piece longitudinally, this study shows how a region can look both abnormal and treatment-responsive and trait-stable at once, depending on which part you look at.
What treatment does and does not touch
The split between normalising and diverging connections is a reminder that symptom remission is not the same as neural return to baseline. A patient can feel better while parts of the network continue to drift. That gap is where relapse risk and residual vulnerability may live, and it argues for following imaging markers past the point of clinical response rather than declaring victory at discharge.
Read subregion by subregion, the insula in bipolar disorder is part state marker and part trait marker, and averaging the two together had been hiding both.
This is a single-site sample with 82 patients and only 43 completing the three-month follow-up, so the longitudinal findings rest on a modest number and need external replication. Treatment was naturalistic rather than a controlled protocol, so the connectivity shifts cannot be tied to any specific medication. The 73.75 percent classifier accuracy is a research-grade signal, not a diagnostic test.