Machine Learning Reveals How Brain Asymmetry Breaks Down in Schizophrenia

For more than a century, scientists have suspected that schizophrenia leaves a fingerprint on the most distinctive feature of the human brain: its asymmetry. The left hemisphere is typically slightly larger in language-related regions, while the right dominates in areas governing attention and spatial processing, and this handedness of neural organization is considered a hallmark of typical neurodevelopment. Now, one of the largest investigations ever conducted on the topic has confirmed that this symmetry is measurably disrupted in schizophrenia, but with a twist that may explain decades of contradictory findings: the pattern of disruption depends on which neuroanatomical subtype of the illness a patient belongs to, and possibly on their ancestry.

The study, published in BMC Medicine, analyzed structural brain scans from 1,761 individuals with schizophrenia and 1,513 healthy controls, drawn primarily from East Asian cohorts across multiple research centers in China. Led by Lu Jin, Zhaoyun Liu, and Xiao Chang, with corresponding author Yuchao Jiang of Shenzhen University, the research team applied a standardized image-processing pipeline to T1-weighted magnetic resonance imaging data, quantifying regional asymmetry in three distinct anatomical measures: cortical thickness, cortical surface area, and subcortical volume. This multi-metric approach matters because asymmetry is not a single property of the brain; a region can show typical leftward bias in thickness while remaining symmetric in volume, and different measures may capture different developmental or degenerative processes.

To compute asymmetry, the researchers calculated an asymmetry index for each brain region, essentially comparing the value of a metric in the left hemisphere against its counterpart on the right, scaled to the total of the two. When they compared patients against controls across the entire sample, three robust differences survived the strictest statistical correction, Bonferroni correction at p < 0.001. Patients showed reduced leftward asymmetry in the rostral middle frontal gyrus, a region in the frontal lobe involved in executive control and working memory; reduced rightward asymmetry in the medial orbitofrontal cortex, which participates in emotion regulation and decision-making; and enhanced rightward asymmetry in the amygdala, the deep temporal structure central to threat processing and emotional salience.

These three findings are notable because they touch regions that map plausibly onto core symptoms of the illness. The frontal deficits align with the cognitive and negative symptoms that have long been linked to frontal lobe dysfunction, while altered amygdala asymmetry resonates with the emotional and perceptual disturbances characteristic of psychosis. Yet the effect sizes, the researchers emphasize, are small. This is a crucial point for the field: individual asymmetry differences in schizophrenia are subtle enough that small studies, which have dominated the literature for decades, could easily find them, miss them, or find them by chance, producing the inconsistent results that have plagued asymmetry research since the era of postmortem studies.

The inconsistency problem became even clearer when the team compared their results with previous findings from Western cohorts covering more than 5,000 schizophrenia patients, work associated with the large-scale ENIGMA consortium effort. The comparison revealed only moderate similarity in cortical surface area asymmetry, with a correlation coefficient of r = 0.36 that reached statistical significance at p = 0.02. For cortical thickness and subcortical volume, however, there were no significant correlations at all between the East Asian and Western asymmetry maps. In other words, the brain asymmetry signature of schizophrenia appears to be partly population-specific, a result that carries substantial implications for how neuroimaging biomarkers are developed and generalized across the globe.

Ancestry is not the only source of heterogeneity the study addressed. Schizophrenia is increasingly understood not as a single brain disease but as a spectrum of neurobiologically distinct conditions that happen to share a clinical label. To capture this, the researchers turned to a machine learning algorithm called Subtype and Stage Inference, or SuStaIn, which had previously been used to derive neuroanatomical subtypes of schizophrenia from patterns of brain structural variation. SuStaIn works by modeling how different patterns of brain change can coexist within a dataset, inferring latent subgroups and the stage of progression along each subgroup’s trajectory, using techniques including Markov Chain Monte Carlo estimation to quantify uncertainty. When the schizophrenia patients in this study were stratified according to these previously derived subtypes, new asymmetry differences emerged that were invisible in the pooled case-control analysis.

The subtype comparisons, corrected at p < 0.05 with Bonferroni correction, revealed differences in the temporal and parietal lobes, the precentral gyrus, and the caudate nucleus. The temporal lobe findings are particularly provocative given that temporal structures, including the superior temporal gyrus and auditory association cortex, have been repeatedly implicated in auditory verbal hallucinations, the most emblematic symptom of schizophrenia. The caudate, a key node of the striatal circuitry that modulates dopamine signaling, sits squarely within the dominant neurochemical models of psychosis. That asymmetry in these regions differs between machine-learned subtypes suggests that the neurodevelopmental deviation underlying asymmetry disruption may follow different trajectories in different patients.

Perhaps the most clinically intriguing result concerns the anterior cingulate cortex, a region straddling the medial frontal surface that regulates conflict monitoring, motivation, and error detection. The researchers found that the relationship between asymmetry in anterior cingulate volume and the severity of negative symptoms, the amotivational and socially withdrawn features of schizophrenia that respond poorly to current medications, differed significantly across the neuroanatomical subtypes, as confirmed by a permutation test at p < 0.05. In practical terms, the same degree of cingulate asymmetry may carry different prognostic or symptomatic meaning depending on which subtype a patient belongs to. This kind of interaction between brain structure and clinical presentation, conditioned by latent neurobiological grouping, is exactly what precision psychiatry aims to exploit.

The study’s conclusions extend a hypothesis with deep historical roots. The idea that schizophrenia involves abnormal cerebral asymmetry was championed by researchers who noted links between the illness, handedness, and language lateralization, arguing that schizophrenia might be the price Homo sapiens pays for the lateralized, language-dominant brain. The new findings neither fully vindicate nor demolish that idea; instead, they refine it. Asymmetry alterations in schizophrenia are real, statistically robust in specific frontal, orbitofrontal, and amygdala regions when tested at scale, but their magnitude is modest and their expression is modulated by both neuroanatomical subtype and ancestral background. A single universal asymmetry marker for schizophrenia, the data suggest, does not exist.

For the field, the methodological lessons may be as important as the biological ones. Large, harmonized samples with standardized preprocessing pipelines can resolve effects that smaller studies have argued over for decades. Machine learning subtyping can unmask heterogeneity that pooling conceals, turning apparent contradictions into structured, interpretable variation. And cross-population comparisons, though still rare in psychiatric neuroimaging, are essential if findings from predominantly Western cohorts are to be translated into clinical tools for the rest of the world. As neuroimaging datasets grow and subtyping algorithms mature, the asymmetry of the human brain, once a niche curiosity of neuropsychology, is emerging as a measurable axis along which the neurobiology of schizophrenia, and perhaps of psychiatric illness more broadly, can be mapped with new precision.

Subject of Research: Brain structural asymmetry alterations in schizophrenia across neuroanatomical subtypes identified by machine learning

Article Title: Revisiting brain asymmetry in schizophrenia: insights from machine learning subtypes from 1,761 patients

Article References: Revisiting brain asymmetry in schizophrenia: insights from machine learning subtypes from 1,761 patients. (n.d.). https://doi.org/10.1186/s12916-026-05318-5

Image Credits: AI Generated

DOI: 10.1186/s12916-026-05318-5

Keywords: schizophrenia, brain asymmetry, MRI, machine learning, SuStaIn, neuroanatomical subtypes, cortical thickness, amygdala, anterior cingulate cortex, negative symptoms, psychiatry, neuroimaging