Two neurostructural subtypes: results of machine learning on brain images from 4,291 individuals with schizophrenia

Abstract

Machine learning can be used to define subtypes of psychiatric conditions based on shared clinical and biological foundations, presenting a crucial step toward establishing biologically based subtypes of mental disorders. With the goal of identifying subtypes of disease progression in schizophrenia, here we analyzed cross-sectional brain structural magnetic resonance imaging (MRI) data from 4,291 individuals with schizophrenia (1,709 females, age=32.5 years±11.9) and 7,078 healthy controls (3,461 females, age=33.0 years±12.7) pooled across 41 international cohorts from the ENIGMA Schizophrenia Working Group, non-ENIGMA cohorts and public datasets. Using a machine learning approach known as Subtype and Stage Inference (SuStaIn), we implemented a brain imaging-driven classification that identifies two distinct neurostructural subgroups by mapping the spatial and temporal trajectory of gray matter (GM) loss in schizophrenia. Subgroup 1 (n=2,622) was characterized by an early cortical-predominant loss (ECL) with enlarged striatum, whereas subgroup 2 (n=1,600) displayed an early subcortical-predominant loss (ESL) in the hippocampus, amygdala, thalamus, brain stem and striatum. These reconstructed trajectories suggest that the GM volume reduction originates in the Broca's area/adjacent fronto-insular cortex for ECL and in the hippocampus/adjacent medial temporal structures for ESL. With longer disease duration, the ECL subtype exhibited a gradual worsening of negative symptoms and depression/anxiety, and less of a decline in positive symptoms. We confirmed the reproducibility of these imaging-based subtypes across various sample sites, independent of macroeconomic and ethnic factors that differed across these geographic locations, which include Europe, North America and East Asia. These findings underscore the presence of distinct pathobiological foundations underlying schizophrenia. This new imaging-based taxonomy holds the potential to identify a more homogeneous sub-population of individuals with shared neurobiological attributes, thereby suggesting the viability of redefining existing disorder constructs based on biological factors.

Keywords: ENIGMA; artificial intelligence; brain gray matter; schizophrenia; structural MRI; subtype.

Extend Data Fig 1.. Pathophysiological progression trajectories in first-episode population and medication-naïve population.

Trajectories are repeated based on the subsample data from the first-episode schizophrenia patients whose illness duration was less than two years (N=1,112, 513 females, mean age=25.4±12.4 years), and another subsample data from medication-naïve patients with schizophrenia (N=718, 353 females, mean age=23.7±12.1 years).

Extend Data Fig 2.. Comparisons of morphological z-score between the two subtypes.

A larger positive z-score indicates a larger deviation of reduction relative to healthy control group. Two sample t test is conducted to examine inter-subtype difference for the (a) averaged cortical volume (t=9.36, p<10e-16, Cohen’s d=0.446); (b) averaged cortical area (t=8.09, p<10e-16, Cohen’s d=0.386); (c) averaged cortical thickness (t=1.29, p=0.198, Cohen’s d=0.061); (d) thalamus volume (t=−4.28, p=1.97e-5, Cohen’s d=−0.205); (e) brain stem volume (t=−9.79, p<10e-16, Cohen’s d=−0.469); (f) hippocampus volume (t=−9.25, p<10e-16, Cohen’s d=−0.449); (g) amgydala volume (t=−7.83, p=8.44e-15, Cohen’s d=−0.379); (h) accumbens volume (t=−6.40, p=1.94e-10, Cohen’s d=−0.305); (i) caudate volume (t=−9.82, p<10e-16, Cohen’s d=−0.468); (j) putamen volume (t=−8.14, p<10e-16, Cohen’s d=−0.389).

Extend Data Fig 3.. Hippocampus subregional morphological z-score for the two subtypes.

A larger positive z-score indicates a larger deviation of reduction relative to healthy control group.

Extend Data Fig 4.. Amygdala subregional morphological z-score for the two subtypes.

A larger positive z-score indicates a larger deviation of reduction relative to healthy control group.

Extend Data Fig 5.. Geographic map of included datasets.

Figure 1.. Two pathophysiological progression trajectories in schizophrenia.

(a) Dice coefficient indicates that K=2 is the optimal number of subtypes with best consistency of the subtype labeling between two independent schizophrenia populations using non-overlap 2-folds cross-validation procedure. Data are presented as mean values +/− SD. (b) The proportion of individuals whose subtype labels keep consistent by non-overlap cross-validation procedure. (c) Sequences of regional volume loss across seventeen brain rFegions for each ‘trajectory’ via SuStaIn are shown in y-axis. The heatmap shows regional volume loss in which biomarker (y-axis) in a particular ‘temporal’ stage (T0-T16) in the trajectory (x-axis). The Color bar represents the degree of gray matter volume (GMV) loss in schizophrenia relative to healthy controls (i.e., z score). (d) Spatiotemporal pattern of pathophysiological ‘trajectory’. The z-score images are mapped to a glass brain template for visualization. Spatiotemporal pattern of gray matter loss displays a progressive pattern of spatial extension along with later ‘temporal’ stages of pathological progression, that is distinct between trajectories. (e-f) Pathological stages of SuStaIn are correlated with reduced gray matter volume of Broca’s area and hippocampus. (g-i) Pathological stages of SuStaIn are correlated with longer disease duration, worse negative symptoms and worse cognitive symptoms.

Figure 2.. Trajectories are reproducibility for samples from different locations of the world.

Two sets of ‘trajectories’ are separately derived from two non-overlapping location cohorts, that are (a) East Asian ancestry (EAS) cohort, and (b) European ancestry (EUR) cohort. The Color bar represents the degree of gray matter volume (GMV) loss in schizophrenia relative to healthy controls (i.e., z score). (c) The similarity of the spatiotemporal pattern of each ‘trajectory’ between any two of cohorts is shown by the heatmap. The color bar of the heatmap represents the similarity, which is quantified via the Spearman correlation coefficient between the trajectories from two cohorts. A total of six location cohorts are classified by where the sample locate at, including the EAS, EUR, China, Japan, Europe and North American. The whole sample is labelled as a cross-ancestry cohort.

Figure 3.. Subtype-specific signatures in neuroanatomical pathology.

Regional Morphological z-scores (i.e., normative deviations from healthy control group) for each subtype are mapped to a brain template for visualization. Effect size of inter-subtype difference is quantified using Cohen’s d.

Figure 4.. Symptomatic trajectories across three stages of disease duration.

Individuals of each subtype are divided into three subgroups according to their illness durations (early stage: ≤2 years; middle stage: 2–10 years; late stage: >10 years). Data are presented as mean values +/− se. * p<0.05.