Brain Subtyping Enhances The Neuroanatomical Discrimination of Schizophrenia

Abstract

Identifying distinctive subtypes of schizophrenia could ultimately enhance diagnostic and prognostic accuracy. We aimed to uncover neuroanatomical subtypes of chronic schizophrenia patients to test whether stratification can enhance computer-aided discrimination of patients from control subjects. Unsupervised, data-driven clustering of structural MRI (sMRI) data was used to identify 2 subtypes of schizophrenia patients drawn from a US-based open science repository (n = 71) and we quantified classification improvements compared to controls (n = 74) using supervised machine learning. We externally validated the unsupervised and supervised learning models in a heterogeneous German validation sample (n = 316), and characterized symptom, cognition, and longitudinal symptom change signatures. Stratification improved classification accuracies from 68.5% to 73% (subgroup 1) and 78.8% (subgroup 2), respectively. Increased accuracy was also found when models were externally validated, and an average gain of 9% was found in supplementary analyses. The first subgroup was associated with cortical and subcortical volume reductions coupled with substantially longer illness duration, whereas the second subgroup was mainly characterized by cortical reductions, reduced illness duration, and comparatively less negative symptoms. Individuals within each subgroup could be identified using just 10 clinical questions at an accuracy of 81.2%, and differential cognitive and symptom course signatures were suggested in multivariate analyses. Our findings suggest that sMRI-based subtyping enhances the neuroanatomical discrimination of schizophrenia by identifying generalizable brain patterns that align with a clinical staging model of the disorder. These findings could be used to improve illness stratification for biomarker-based computer-aided diagnoses.

Introduction

Developing objective biological tests to identify individuals with psychotic symptoms is a pressing need that has not been met.¹ A promising direction is the application of machine learning techniques that make predictions at the level of an individual patient using magnetic resonance imaging (MRI) measurements of brain structure.^2–5 However, these predictions are limited by the symptomatic and neuroanatomical heterogeneity introduced by broad clinical definitions, such as schizophrenia or first-episode psychosis.⁶ Reducing this heterogeneity by using unbiased methods to identify individuals with different brain subtypes may increase predictive accuracy and inform the development of psychiatric classification⁷ and stratification⁸ systems.

Clinical heterogeneity of schizophrenia has been recognized since the establishment of a single nosological entity.^9,10 A unifying aspect of early clinical divisions has been the recognition of a subtype of individuals who present with more severe negative symptoms, cognitive deficits, and a deteriorating course that is putatively representative of an endogenous, biological aetiology; for example, Kraepelin’s concept of dementia praecox simplex, Leonhard’s classification of systematic schizophrenia, Crow’s “Type II” schizophrenia, or Carpenter’s deficit schizophrenia.¹⁰ In the simplest 2-class formulation, individuals with these symptoms are were compared to individuals with a remitting course and more positive symptoms, such as hallucinations, delusions, and formal thought disorder.¹¹ Later statistical approaches that cluster individuals based on psychiatric questionnaires supported the 2-class grouping¹² and reified other subgroupings, particularly the “hebephrenic” or disorganized subtype.¹³

Brain differences associated with clinical subtypes of schizophrenia have also been postulated since the inception of the diagnosis,¹⁰ with the individuals experiencing a deficit syndrome hypothesized to exhibit more severe abnormalities.^11,14 This hypothesis has been supported by recent research showing more widespread abnormalities in subgroups with predominantly negative symptoms when compared with positive or disorganized subtypes.^4,15–18 With growing recognition that clinical subtyping is limited by the measurement of subjective symptoms and signs,¹⁹ current research has also highlighted brain heterogeneity by detecting subgroups of individuals using cognitive variables,^20–22 genes,²³ functional MRI,²⁴ or combinations of electrophysiology and cognition.²⁵ These subgroups often do not respect clinical boundaries, or classification schemes, and demonstrate subgroups of individuals with unique symptom and brain profiles.²⁵

Evidence of putative schizophrenia brain subtypes suggests that this heterogeneity impairs current neurodiagnostic or prognostic brain classification systems⁶ based on clinical diagnostic categories; that is, because each clinical subgroup putatively has its own unique neuroanatomical “fingerprint.” Methods to reduce heterogeneity to increase predictive accuracy have been developed, but have mostly been applied to functional neuroimaging, cognitive, or multimodal data in healthy control subjects,²⁶ attention-deficit hyperactivity disorder,^20,27 depression,²⁸ and Alzheimer’s disease.²⁹ This research has also shown mixed results, with functional imaging studies finding increases in classification accuracy after subtyping,²⁸ while the only structural neuroimaging study to-date was conducted in Alzheimer’s disease and did not demonstrate improvements.²⁹ Existing research therefore highlights a need to determine if neuroanatomical clustering in schizophrenia increases neurodiagnostic predictions.

In this proof-of-concept study, we developed methods to test the hypothesis that neuroanatomical clustering will result in increased accuracy in identifying individuals with schizophrenia when compared with a healthy control group. We further aimed to assess the clinical translation potential by identifying the subgroups using small subsets of easily obtainable clinical variables, and we validated all models in a large geographically and ethnically distinct external sample. We expected that stratification would result in an increased ability to identify an individual with schizophrenia based on the decoding of neuroanatomical heterogeneity.^6,29

Methods

Participants

We included 74 healthy control and 71 schizophrenia patients (table 1) that comprised the publicly available data repository (http://coins.mrn.org/dx) of the Mind Research Network and the University of New Mexico (see http://cobre.mrn.org/). Clinical psychiatrists using the Structured Clinical Interview for DSM-IV determined schizophrenia diagnosis and symptoms were rated using the Positive and Negative Syndrome Scale (PANSS).³⁰ Antipsychotic medication at MRI scan was converted to chlorpromazine equivalents.^31,32 Subjects were excluded if they had any history of neurological disorder, history of mental disability, history of severe head trauma, and history of substance abuse or dependence within the last 12 months. Informed consent was obtained from all subjects according to institutional guidelines required by the Institutional Review Board at the University of New Mexico.

Table 1.

Sociodemographic, Clinical, and Global Anatomical Characteristics of the Patient and Control Groups

	Discovery Sample			Validation Samples
	SCZ	CTRL	Test Statistic / P value	SCZ Matched	CTRL Matched	SCZ	CTRL
Demographic variables
N	71	74		71	74	158	158
Mean age (yrs)	38.1 (14)	35.8 (11.6)	n.s	33.9 (10.2)*	35.8 (11.5)	30.8 (10.0)**	31.1 (9.7)
Handedness (mixed or left)	12 (16.9%)	3 (4.1%)	χ 2(1) = 6.4, P = .01	3 (4.2%)	8 (10.8%)	10 (6.3%)	16 (10.1%)
Sex (male)	57 (80.3%)	51 (68.9%)	n.s	57 (80%)	51 (69%)	117 (74%)	117 (74%)
Years of education/schoolinga	13 (1.8)	14.4 (3.3)	t(132) = 3.1, P = .003	10.5 (1.8)	11.8 (1.4)	10.6 (2.1)	12.1 (1.2)
Mean verbal IQ (WASI)	98.6 (1.8)	105.1 (14.3)	t(132) = 2.6, P = .01	—	—	—	—
Mean performance IQ (WASI)	103.8 (19.1)	114.2 (12.3)	t(132) = 3.8, P < .001	—	—	—	—
IQ (WASI)	100.4 (16.9)	108.4 (22.2)	t(132) = 2.4, P = .02	—	—	—	—
Clinical variables					—
Age at first psychotic symptoms	21.13 (7.5)	na		25.3 (9.4)*	—	25.5 (8.0)**	—
Age at first psychiatric hospitalization	21.9 (8.5)	na		—	—	—	—
Duration of illness (yrs)b	16.84 (12.99)	na		8.6 (8.4)**	—	4.5 (7.0)**	—
Number of hospitalizations	5.3 (7.9)	na		2.9 (2.7)	—	1.9 (2.1)	—
PANSS-positive symptoms score	14.8 (4.7)	na		9.8 (7.3)**	—	11.9 (8.0)*	—
PANSS-negative symptoms score	14.6 (4.7)	na		16.9 (9.2)	—	15.2 (9.7)	—
PANSS general score	28.2 (8.4)	na		25.8 (13.5)	—	25.6 (16.1)	—
Medication dose (CPZ equivalent)	369.2 (306.3)	na		370 (367.7)	—	346.3 (373.4)	—
Global anatomical volumes
Mean global grey matter volume, ml (SD)	604.89 (82.7)	621.13 (65.2)	n.s	585 (66.2)	604.6 (68)	593.5 (71.7)	623.1 (71.1)
Mean global white matter volume, ml (SD)	573.78 (74.3)	557.83 (68.6)	n.s	642.4 (70.4)	635.4 (69)	636.0 (68.7)	638.6 (69.5)
Mean global cerebrospinal fluid volume, ml (SD)	262 (40.9)	240.78 (32.1)	t(143) = −3.49, P < .001	211.4 (32.0)	194.4 (24.1)	205.7 (30.1)	192.9 (25.6)

Note: SCZ, schizophrenia sample; CTRL, healthy control sample; SCZ matched, external validation sample matched to the age, sex, and illness duration of the discovery sample; CTRL matched, healthy control sample matched to age and sex of the discovery sample.

^aTotal education years of the discovery sample, total years in primary/secondary school of the validation sample.

^bDuration of illness calculated as the time between the age at first psychotic symptoms and current age.

^* P < .05 and

^** P < .001 when compared with the corresponding group in the discovery sample.

Structural Image Acquisition and Preprocessing

Participants were scanned on a 3T SIEMENS TIM scanner with a 12-channel radio-frequency coil at the Mind Research Network. Structural images were obtained using a 5-echo MPRAGE sequence with the following parameters: TR = 2530 ms, TE = (1.64, 3.5, 5.36, 7.22, 9.08 ms), TI = 900 ms, flip angle = 7°, FOV 256 × 256 mm², slice thickness = 176 mm, matrix size 256 × 256 × 176. Structural MRI data were preprocessed using the VBM8 toolbox (http://dbm.neuro.uni-jena.de/vbm8/), including the application of a Blockwise Nonlocal-Means Filter, segmentation into grey, white, and cerebrospinal fluid images, and the application of a hidden Markov Random Field model to produce modulated grey matter images. The modulated images were normalized to Montreal Neurological Institute (MNI) structural template (3 mm isotropic) using the DARTEL algorithm,³³ and smoothed using a full-width-at-half-maximum kernel of 3 mm isotropic. Each image was masked using a binary image of the MNI template and proportionally scaled to each individual’s global grey matter volume.

Identification of Schizophrenia Subgroups

Patients’ images were corrected for age and sex using a regression model computed in the healthy controls, following a previously validated strategy that preserves illness effects that are collinear with age³⁴ (supplementary material). Image dimensionality was then reduced using principal component analysis (PCA), and eigenvariates were retained explaining 80% of the variance; additionally, we also tested the stability of results by retaining different component solutions and repeating the pipeline (supplementary material). In this reduced data space, we identified subgroups with the fuzzy c-means algorithm (FCM),³⁵ which is an unsupervised machine learning technique that detects hidden structure in data by organizing it into subsets of similar elements (clusters or subgroups) without any prior knowledge of group membership and does not assume definitive boundaries between clusters (ie, it allows fuzzy boundaries). This technique has increased face validity compared with hard clustering techniques (eg, k-means) because it allows for expected overlap between subgroups.⁷ We applied the MATLAB 2014a (MathWorks Inc.) implementation of FCM with default settings for the main analysis (m = 2, 1000 iterations)³⁵ and also tested the stability of results across different fuzziness values of m (supplementary materials). Our main analysis was limited to the discovery of 2 clusters due to the need to retain an adequate sample size in each subgroup for the downstream classification systems to successfully “learn” predictive signatures, but we also conducted supplementary analysis retaining 3 clusters (supplementary materials). Subjects were assigned to either subgroup based on their highest fuzzy membership value (ie, individual, i, placed in group j, if w_ij1 > w_ij2, where w is the cluster membership weight). Because FCM clustering solutions can change each time the algorithm is run due to random starting points, we employed a consensus-based clustering technique for FCM in order to identify the final subgroup partitioning^36–38(supplementary material).

Diagnostic Pattern Classification

We established a machine learning pipeline using our in-house tool NeuroMiner (http://www.pronia.eu/neurominer) to extract grey matter volume patterns that separate control subjects from: (a) the whole patient sample; (b) patients in the first subgroup; and (c) patients in the second subgroup. The modulated and smoothed grey matter images entered the analysis pipeline. Models were trained within a repeated, nested cross-validation framework.^34,39 In the inner loop of cross-validation, the training data were separated into 10 folds and preprocessed using the same settings employed during clustering analysis, that is, corrected for age and sex, PCA (80% of variance retained), and scaling [0,1]. For each training set, this procedure was repeated by permuting the data 10 times within each study group. Training samples were analyzed with a linear support vector machine (SVM; LIBSVM 3.12⁴⁰; http://www.csie.ntu.edu.tw/~cjlin/libsvm) and we selected models that maximized the balanced accuracy (average of sensitivity and specificity) across a range of SVM hyperparameters (geometric progression from C = 2^–3 to C = 2⁵). To account for classification bias due to uneven sample sizes of the SCZ and HC groups,⁴¹ we employed class-weighting of the C hyperparameters using the inverse ratio of the training group sizes multiplied across a second set of hyperparameters (from 1 to 2 in steps of 0.1; supplementary material). To guard against overfitting, we retained 5% of the top-performing SVM models with different hyperparameter settings (ie, 5 models) for each of the 100 (10 folds × 10 permutations) inner cross-validation folds and applied the retained analysis chains to the outer-loop validation data. For each outer CV fold, this created 500 retained models (5 models × 100 inner folds). As the outer CV folds were permuted 10 times, this procedure produced an ensemble of 5000 decision scores reflecting the degree to which an individual was classified as control or patient. We determined each participant’s class membership through majority voting across models in which given participant was not involved as training instance. Permutation-based P-values were calculated to assess classification accuracy differences compared to random subdivisions of the data (supplementary materials) and group differences were assessed with McNemar’s tests.⁴² Visualization of the multivariate patterns used in classification was achieved by back-projecting the feature weight vector of each SVM model from PCA to MNI space. Voxel probability maps (VPMs) were then visualized for each comparison reflecting the degree of consistency and reliability of the subgroup signatures following Koutsouleris et al. (2015) (supplementary material).

External Validation of Subgroups

We externally validated the models obtained in the COBRE discovery sample in an independent cohort consisting of 158 patients with first-episode psychosis and chronic schizophrenia, and 158 age- and sex- matched controls, examined at the Department of Psychiatry, Ludwig-Maximilian-University, Munich (supplementary material) with a 1.5T MR scanning system. We evaluated performance at 2 levels: (1) in a subgroup of the Munich cohort consisting of individuals with chronic presentations who were matched with the discovery sample for age, sex, and illness duration; and (2) the full and more heterogeneous Munich sample containing individuals with illness durations ranging from 15 days to 44 years (table 1). In brief, structural MR images from the independent sample were first corrected for scanner differences and then processed using the identical preprocessing pipeline as in the discovery sample (supplementary material). To obtain a completely unbiased estimate of generalizability, we applied all models obtained from our unsupervised and supervised discovery analysis without any in-between re-training steps to the external validation cohort; for example, correction, PCA, scaling, and determination of distance from discovery cluster centroids for the clustering analysis (supplementary material).

Clinical Comparison of Subgroups

We aimed to identify multivariate patterns that separated the subgroups with the smallest possible subsets of 41 demographic, illness course, and psychometric variables (supplementary table S1). The variables entered an SVM nested leave-one-per-group-out (ie, leaving one person from each group out) cross-validation pipeline—this strategy was employed instead of a 10 × 10 framework as used for neuroimaging to maximize the training set data in the presence of the wrapper procedure below. In the inner cross-validation, the training data were first scaled [0,1], followed by nearest neighbor-based imputation to fill in missing values (0.33% in total).⁴³ Each processed training sample entered a stepwise forward variable selection process⁴⁴ using a linear SVM at each selection step to iteratively grow a predictive group of variables up to 50% of the variable set. The analysis chains were then applied to the outer-loop validation data and each validation patient’s outcome class was determined through majority voting. The percentage of times each variable was selected by the SVM models across 600 training partitions of the cross-validation setup was calculated to assess variable importance (>50% selection probability across partitions). To assess the clinical translation potential by further reducing the variable set, we then identified the top 10 variables in the discovery analysis and repeated the processing steps in the discovery and validation cohorts.

Supplementary Characterization of Cognition and Illness Outcomes

To further characterize the subgroups, we conducted supplementary analyses of cognition with a neuropsychological test battery with available data from the discovery cohort, and 1-year illness outcomes in a subset of the external validation cohort (supplementary materials).

Results

Sociodemographic, Clinical, and Global Anatomical Results

Discovery sample patients with schizophrenia did not differ from healthy controls with regard to age and sex, but they exhibited increased left-handedness, and decreased years of education and IQ (table 1). The external validation patients were younger, with shorter illness duration, less positive symptoms, and a later age of onset compared with the discovery sample (table 1).

Brain SVM Classification Analysis

Classification Performance.

Classifying patients compared with controls based on brain volume differences using the full schizophrenia sample resulted in a balanced accuracy of 68.3% (table 2). Classification performance for both subtype groups exceeded this for the first subtype group by 4.7% at trend levels of significance (P = .07) and 10.5% (P = .02) in the second subtype group. Performance gains were associated with increases in specificity for the first subgroup (+14.9%, P = .006), and both sensitivity and specificity for the second subgroup at trend levels (+10.8, P = .07; +10.2%, P = .06). Classification performance was significantly increased compared with random subdivisions for subgroup 1 (P < .01) and subgroup 2 (P < .001; supplementary materials). Deriving subgroups using different PCA solutions or fuzziness values extended and supported these findings with an average increase in balanced accuracy of 9% that was significant for all comparisons (P < .05; supplementary materials). A 3-cluster solution also supported the findings by demonstrating the division of the sample into 2 main groups (combined n = 60) showing significant increases in balanced accuracy (subgroup 1, +12.5%, P = .01; subgroup 2, +11.2, P = .002) and a third, smaller group (n = 11) (supplementary materials).

Table 2.

Diagnostic Performance of Brain Classifiers Before and After Neuroanatomical Subgrouping in the Discovery (n = 145) and Validation (n = 316) Samples

	Group	TP	FP	TN	FN	Sens (%)	Spec (%)	BAC (%)	FPR (%)	PPV (%)	NPV (%)	DOR
Discovery sample	WG	50	25	49	21	70.4	66.2	68.3	33.8	66.7	70.0	4.3
	SG1	26	14	60	14	65.0	81.1*	73.0	18.9	65.0	81.1	11.8
	SG2	25	17	57	6	80.6	77.0	78.8*	23.0	59.5	90.5	12.3
Validation 1	WG	48	24	50	23	67.6	67.6	67.6	32.4	66.7	68.5	4.3
	SG1	16	11	63	15	51.6	85.1**	68.4	14.9	59.3	80.8	12.1
	SG2	30	17	57	10	75.0	77.0*	76.0*	23.0	63.8	85.1	10.7
Validation 2	WG	95	52	106	63	60.1	67.1	63.6	32.9	64.6	62.7	3.3
	SG1	24	16	142	22	52.2	89.9**	71.0**	10.1	60.0	86.6	26.5
	SG2	78	55	103	34	69.6	65.2	67.4	34.8	58.6	75.2	4.0

Note: Validation 1, validation sample matched to age, sex, and illness duration; Validation 2, unmatched validation sample; WG, whole group; SG1, first subgroup; SG2, second subgroup; TP, true positive; FP, false positive; TN, true negative; FN, false negative; Sens, sensitivity; spec, specificity; BAC, balanced accuracy; FPR, false-positive rate; PPV/NPV, positive/negative predictive values; DOR, diagnostic odds ratio.

^* P < .05.

^** P < .001.

External validation mainly supported the main findings (table 2). When individuals in the validation sample were partially matched with the discovery group using age, sex, and illness duration, neuroanatomical stratification resulted in an increase in specificity in the first subgroup (+22.9% P = .002) and increases in balanced accuracy (+9.4; P = .04) and specificity (+10%; P = .04) in the second subgroup. In the more heterogeneous, full schizophrenia sample, increased balanced accuracy was also observed for the first subgroup (+7.4%, P = 1.66 × 10^–6) with increased specificity (+22.8%; p = 2.67 × 10^–4), but no differences were found in the second subgroup.

We observed significant age differences between the subgroups of all samples, despite statistically controlling for age prior to subtyping (supplementary tables S2–S4). Thus, we used 2 strategies to exclude the possibility that clustering performance increases were due to sociodemographic age and not illness factors: (1) we applied the clustering pipeline to the MRI images of the healthy control discovery sample (n = 74); (2) we replaced the unsupervised clustering steps by an age-defined median split of the SCZ sample (supplementary materials). Healthy control clustering did not result in an age difference between subgroups (mean difference = 0.4 years, P = .9), and accuracy increases were not observed following a medial split of age in the schizophrenia sample with downstream classification (supplementary material).

Neuroanatomical Differences Between Subgroups.

Neuroanatomical subgroup signatures are presented in figure 1. Without stratification, the discriminative signature consisted of reduced grey matter volume in the frontal, insula, temporal, hippocampal, and striatal regions. The signature also included increased grey matter volume in the temporo-parieto-occipital junction, medial parietal lobe, and posterior cingulate cortex. Both subgroups were separated from controls by decreased volume in the medial prefrontal lobes and left temporal cortex, but 2 distinctive profiles were also evident. The first subgroup was additionally characterized by insula, striatal, thalamic, hippocampal, and right-hemispheric superior temporal volume reductions, and a pattern of increased volume in the medial and lateral parietal lobes. The second subgroup involved volume reductions in the lateral prefrontal, medial parietal, and temporal cortices and volume increments in cerebellar structures.

Fig. 1.

Voxel probability maps (VPMs) showing effect of neuroanatomical subtyping on classification of schizophrenia. A. VPM associated with the separation of schizophrenia patients compared with controls using the entire sample for cortical (upper images) and subcortical/cerebellar regions (lower images). B. Subtype 1 showing involvement of the insula, medial frontal, temporal, and parietal lobes. C. Subtype 2 showing a more diffuse pattern associated with the medial frontal, lateral frontal, and temporal cortex. Scale represents the degree of reliability in which the voxels were either negatively (red-yellow) or positively (blue-purple) associated with the discrimination of schizophrenia subjects from controls. Further details presented in supplementary materials.

Clinical Analyses

Main Analyses.

Univariate tests in the discovery sample indicated that subgroup 1 patients were significantly older with a later age of onset, longer duration of illness, and more hospitalizations (supplementary tables S2–S4). Notably, despite a mean difference in age, the age distributions were overlapping (supplementary figure S1). No differences in medication dose or type at baseline were evident (supplementary tables S2–S4 and S7). Clinical machine learning correctly predicted subgroup membership with a balanced accuracy of 80.8% (table 3). The feature selection procedure identified 14 variables that were selected by >50% of the SVM models (figure 2). The most reliably selected variables were age, illness duration, and sex. Differential symptom patterns were also found with subgroup 1 classification including proportionately more negative symptoms compared with subgroup 2, which also showed hyperthymic and conceptual disorganization symptoms. Further reducing the variable set by only using the top 10 reliably selected variables produced an average accuracy of 80.4% in the external validation samples (table 3).

Table 3.

Performance of Clinical Classifiers to Identify Individuals Within Each Subgroup in the Discovery (n = 71) and Validation (n = 158) Samples

	TP	FP	TN	FN	Sens (%)	Spec (%)	BAC (%)	FPR (%)	PPV (%)	NPV (%)	DOR
All measures
Discovery sample	35	8	23	5	87.5	74.2	80.8	25.8	81.4	82.1	11.5
Top 10 features
Discovery sample	35	5	26	5	87.5	83.9	85.7	16.1	87.5	83.9	29.4
Validation 1	27	11	29	4	87.1	72.5	79.8	27.5	71.1	87.9	10.0
Validation 2	38	23	89	8	82.6	79.5	81.0	20.5	62.3	91.8	16.2

Note: WG, whole group.

Fig. 2.

Feature selection probability showing the most reliable clinical contributions to the separation of subgroups. Features selected more than 50% of the time during multivariate training using demographic details and symptoms from the positive and negative symptom scale. Features with a positive weight indicating relative increases in the first subgroup are indicated in yellow. Features with a negative weight indicating increases in the second subgroup are indicated in blue.

Supplementary Analyses.

Differential cognitive and illness outcome patterns were suggested in supplementary analyses (supplementary materials). For cognition, a multivariate pattern (81.6% balanced accuracy) was found consisting of age, sex, and differential contributions of attention versus reasoning/problem solving for each subgroup. For illness outcomes, a differential multivariate pattern of symptom changes (75.9% balanced accuracy) was found that mainly highlighted contributions of age, sex, and illness duration, but also included differential contributions of conceptual disorientation and lack of conversational flow in the first subgroup, versus changes in guilt feelings and delusions in the second.

Discussion

Results suggested that brain subtyping can enhance the neuroanatomical discrimination of schizophrenia, as evidenced from the primary analysis, supplementary analyses, and when the subtype models were externally validated in an independent sample. The findings provide proof-of-concept evidence that reducing brain heterogeneity enhances neurodiagnostic accuracy,⁶ which is an important methodological contribution to initiatives aiming to create clinical tools (e.g., PRONIA, PSYSCAN, NAPLS)⁴⁵ where even modest accuracy increases (eg, <10%) can maximize public health gains.⁴⁶

The results add to previous research that has found performance gains when reducing heterogeneity through clustering with cognitive variables using similar comparisons.²⁰ However, neuroanatomical clustering research of Alzheimer’s disease patients has not reported similar gains.²⁹ This discrepancy may be due to the use of different methodological approaches, or differences in the clinical heterogeneity of the samples. For example, the processes leading to meso- and macroscopic brain abnormalities are largely unknown for schizophrenia and the condition may be more heterogeneous. Thus, the results may specifically highlight the need to resolve neuroanatomical heterogeneity for applications in psychiatry in order to provide applications that exceed the performance metrics currently reported in the field.^2,47

The subgroups only partially agree with previous symptom clustering research showing a negative syndrome subgroup with widespread brain abnormalities.^15,25,48 Specifically, we found that illness duration strongly drove the results as the first subgroup had longer illness duration and a mixed clinical pattern coupled with both cortical and subcortical brain volume reductions. The second subgroup had reduced illness duration coupled with comparatively less negative symptoms, conceptual disorganization, hallucinations, and hyperthymic symptoms that were connected to a predominantly cortical brain pattern. Notably, the specific clinical symptoms of the first subtype group including grandiosity, delusions, and social/emotional withdrawal have been linked to the limbic, paralimbic, and superior temporal functional abnormalities as found in this study,^49–51 and the symptoms found in the second subtype have been related to prefrontal abnormalities that were also found here.^52–55 When illness duration was included in the predictive model, differences between cognitive profiles and illness courses were also tentatively suggested by further multivariate analyses.

Due to the strong contribution of illness duration coupled with differential clinical and neuroanatomical predictive markers, the results of this study may suggest that neuroanatomical subtyping in schizophrenia identifies a previously undetected separation boundary during illness progression; that is, from a process driven by cortical volume decreases (with a notable focus in the prefrontal lobes) toward greater subcortical decreases with concomitant differences in symptoms. Alternatively, the results could reflect the identification of a more severe subtype of schizophrenia that inadvertently appears in older adults of cross-sectional samples because remitted patients are not help seeking.^10,11,14 Further longitudinal studies are required to disentangle these hypotheses and potentially identify neuroanatomical transition and inflexion points during the course of the disorder. Achieving this goal would be an important step toward the development of biologically informed clinical staging models.^8,56,57

The potential for clinical translation was highlighted by analyses showing that individuals in each subgroup could be identified in the external validation samples on the basis of just 10 questionnaire items at up to 81% accuracy. These findings were driven by collinear age and illness duration effects, but the method still demonstrates that once distinct biological subtypes of illness are found they can be approximated with selective and precise batteries of easily accessible patient-reportable items.^8,57 These results also add to existing brain subtyping research across disorders^22,25,28 by suggesting that clinical assessments may ultimately still be the most cost-effective and least burdensome proxy to biologically defined illness categories for diagnoses or staging models.

A number of limitations need to be kept in mind. The first limitation was that we were not able to investigate whether the subgroups reported here were truly reflective of neurobiological differences between individuals, for example, at transcriptomic or genetic levels. Analyses were restricted to 2 clusters and these were not further validated with other biological measures, treatment response variables, or comprehensive functional outcome data. In similarity to early clinical clustering studies^10,12,13 and recent studies in depression,²⁸ we have putatively investigated the first major subdivision of a hierarchy of psychosis-related subgroups. However, future studies need to investigate the lower levels of such a hierarchy, determine if they are transdiagnostic, assess whether a continuum model may be more appropriate, and to search for unique biological differences that may inform clinical treatment. An intriguing question is whether these further results will follow the nuanced evolution of theoretical models mentioned in the introduction that now balance categorical definitions with continuum models of schizophrenia (eg, wherein dichotomies such as type I and type II schizophrenia are extreme ends of a continuum).^9,10

A further major limitation to interpretations arising from this study regarding underlying biological subgroups of schizophrenia is that the potential cumulative effects of antipsychotics⁵⁸ and medication class⁵⁹ may have influenced subgroup classification. This limitation is especially relevant given subgroup associations with illness duration. If a future aim is to identify fundamental biological subdivisions in individuals diagnosed with schizophrenia, or within a psychosis spectrum, then future longitudinal studies are needed in unmedicated individuals who are at familial psychosis risk, in those at clinical high risk, and in individuals experiencing a first psychotic episode.

Despite the limitations, this study provides initial evidence that subgroup identification enhances neurodiagnostic performance for schizophrenia. In doing so, it complements initiatives aiming to redefine diagnoses and staging models based on biological data¹⁹ and highlights the important role of illness duration within this context. Future studies are required to determine whether this approach can be extended to enhance the accuracy of differential diagnoses, prognoses, and treatment selection.

Supplementary Material

Supplementary material is available at Schizophrenia Bulletin online.

Funding

Dominic B. Dwyer was supported by the Deutsche Forschungsgemeinschaft (DFG) within the framework of the projects www.kfo241.de and www.psycourse.de (SCHU 1603/4-1, 5-1, 7-1; FA241/16-1). Lana Kambeitz-Ilankovic was supported by the EU-FP7 project PRONIA (“Personalised Prognostic Tools for Early Psychosis Management”; Grant Agreement No. 602152). Prof C. Pantelis by NHMRC Senior Principal Research Fellowships (IDs: 628386 and 1105825). The data were collected under the National Institutes of Health grant (#NIH P20GM103472) to Vince Calhoun.

Conflict of interest: The authors have declared that there are no conflicts of interest in relation to the subject of this study.