Neurobiological divergence of the positive and negative schizophrenia subtypes identified upon a new factor-structure of psychopathology using non-negative factorization: An international machine-learning study

Abstract

Objective:

Disentangling psychopathological heterogeneity in schizophrenia is challenging and previous results remain inconclusive. We employed advanced machine-learning to identify a stable and generalizable factorization of the “Positive and Negative Syndrome Scale (PANSS)”, and used it to identify psychopathological subtypes as well as their neurobiological differentiations.

Methods:

PANSS data from the Pharmacotherapy Monitoring and Outcome Survey cohort (1545 patients, 586 followed up after 1.35±0.70 years) were used for learning the factor-structure by an orthonormal projective non-negative factorization. An international sample, pooled from nine medical centers across Europe, USA, and Asia (490 patients), was used for validation. Patients were clustered into psychopathological subtypes based on the identified factor-structure, and the neurobiological divergence between the subtypes was assessed by classification analysis on functional MRI connectivity patterns.

Results:

A four-factor structure representing negative, positive, affective, and cognitive symptoms was identified as the most stable and generalizable representation of psychopathology. It showed higher internal consistency than the original PANSS subscales and previously proposed factor-models. Based on this representation, the positive-negative dichotomy was confirmed as the (only) robust psychopathological subtypes, and these subtypes were longitudinally stable in about 80% of the repeatedly assessed patients. Finally, the individual subtype could be predicted with good accuracy from functional connectivity profiles of the ventro-medial frontal cortex, temporoparietal junction, and precuneus.

Conclusions:

Machine-learning applied to multi-site data with cross-validation yielded a factorization generalizable across populations and medical systems. Together with subtyping and the demonstrated ability to predict subtype membership from neuroimaging data, this work further disentangles the heterogeneity in schizophrenia.

INTRODUCTION:

Schizophrenia is a heterogeneous disorder with marked inter-individual variability of psychopathology, which is related to treatment response and long-term outcomes (1,2). Earlier clinical subtypes (e.g., hebephrenic or paranoid) were eliminated in recent nosological classifications due to poor diagnostic stability, validity, and utility (3). Considerable efforts have been devoted to better understand and categorize schizophrenia phenomenology by factorizing symptoms into cardinal dimensions or clustering patients into psychopathological subtypes based on scales such as the Positive and Negative Syndrome Scale (PANSS), a well-established assessment of schizophrenia psychopathology (4).

The three PANSS subscales (negative, positive, and general psychopathology [GPS]) are generally suggested to not optimally and adequately capture the latent organization of schizophrenia symptomatology; items within a subscale show modest internal consistency (5), while those across subscales are strongly correlated (6–8). Previous factorizations of the PANSS have been inconsistent, advocating solutions between four and seven factors (6–11). Although a five-factor structure was most frequently proposed (9–11,14), it has continuously failed to be confirmed in independent samples (12–15). Interpretations of previous factor models are, furthermore, complicated by lack of sparsity (all items contribute to any factor) (16) and co-existence of positive and negative weights (17). Finally, most previous studies investigated rather small and geographically restricted samples, raising doubts over generalization to different populations and medical systems as systematic cross-validation to assess stability and generalizability have rarely been performed. Previous work on psychopathological subtyping is likewise inconclusive (18–20) with added concerns related to longitudinal stability and neurobiological differentiability. These aspects are particularly relevant in the emerging context of precision psychiatry, and raise the following questions: Do psychopathological subtypes represent stable patient characteristics, and do they relate to divergent neurobiological substrates that are identifiable from brain imaging data? Functional MRI (fMRI) parameters may serve as an endophenotype, underpinning the symptomatic heterogeneity (21), which have added ample valuable insights into the neural pathophysiology of schizophrenia and its relation to clinical presentations (22,23). However, whether, and to what extent, the brain functional connectivity (FC) could discriminate psychopathological subtypes remains unknown. A successful classification using endophenotypical characteristics would support distinctiveness of symptomatically derived schizophrenia subtypes expressed along the cardinal axes of psychopathology.

In the present study, we addressed the aforementioned questions as follows: 1) a robust, cross-validated, and interpretable factor-structure of schizophrenia psychopathology was identified based on PANSS scores of over 2000 patients using an unsupervised machine-learning approach (orthonormal projective non-negative matrix factorization [OPNMF]) (24–27); 2) core schizophrenia subtypes were derived by applying soft-clustering to the identified factor-structure, whose longitudinal stability was evaluated in repeatedly assessed patients; and 3) neurobiological differentiation of those subtypes based on resting-state FC (rsFC) patterns was investigated by cross-validated classification analysis, serving as a biological validation of a clinical (multivariate) construct.

METHODS AND MATERIALS:

Sample

We used two large datasets collectively providing individual-item PANSS scores for 2035 schizophrenia patients: 1) a subset of 1545 patients (586 followed up after 1.35±0.70 years) with complete individual-item PANSS scores and a diagnosis of schizophrenia (DSM-IV criteria) retrieved from the Pharmacotherapy Monitoring and Outcome Survey (PHAMOUS) database (28,29). This dataset was recruited from four institutions located in the Netherlands and assessed with a uniform protocol; 2) a deliberately heterogeneous sample pooled from nine centers located in Europe, the USA, and Asia (490 patients; Table 1; Table S1). This international dataset covers a broad range of clinical states, settings, and medical systems, making it ideal to evaluate the generalization of our factor-model to new and diverse populations. Diagnoses in the international sample were established based on the DSM-IV, DSM-IV-TR, or DSM-V criteria (Supplement). At all sites, data was acquired in accordance with the declaration of Helsinki and after obtaining informed consent from the patients. Approval for the pooled re-analysis was obtained from the ethics committee of the HHU Düsseldorf.

Table1.

Demographic and clinical characteristics of the patients with schizophrenia

Characteristics	PHAMOUS sample (N=1545)	International dataset from 9 centers (N=490)	International dataset with imaging (N=147)	Statistics	p-value
Demographic
Age (years)a	44.15 (11.42)	33.82 (10.28)	34.89 (11.67)	183.51	<.001
Gender (male/female)	1108/437	333/157	102/45	2.45	.292
Illness during (years)b	18.22 (10.54)	9.13 (8.98)	11.37 (10.36)	134.71	<.001
PANSS
Positivec	12.48 (4.91)	14.24 (5.76)	15.36 (5.50)	37	<.001
Negative	14.60 (6.20)	14.67 (7.21)	15.07 (6.06)	0.375	.687
Generald	26.70 (8.16)	29.10 (11.34)	30.93 (10.97)	23.67	<.001
Illness severity (Total PANSS)e	53.78 (16.35)	58.01 (21.87)	61.36 (19.57)	19.48	<.001
P3 item (hallucinations)f	2.30 (1.47)	2.66 (1.83)	3.22 (1.91)	28.18	<.001
Medicationg
Atypical antipsychotics	NA	167 (34.1%)	110 (74.8%)
Typical antipsychotics	NA	26 (5.3%)	8 (5.4%)
Both A & T	NA	16 (3.3%)	9 (6.1%)
None or unknown	NA	281 (57.3%)	20 (25.9%)
Current antipsychotic medicationh	NA	19.64 (14.15)	19.30 (12.57)

Data are mean (SD), or n (%). p-values in bold indicate a significance of p < .05. Except for gender, which was based on χ² test, other statistics were all based on one-way ANOVA analyses. Of note, since the detailed medication information was missing for several patients in different proportions for those with or without imaging data in the international dataset, statistical comparisons were not conducted.

Post-hoc analysis after one-way ANOVAs showing significant pair-wise differences among the three datasets:

^aPHAMOUS > International sample = International sample with imaging at p < .05, Bonferroni corrected.

^bPHAMOUS > International = International with imaging at p < .05, Bonferroni corrected.

Information of illness duration was available for 1326 patients in the PHAMOUS sample, 393 patients in the international sample

^cPHAMOUS < International sample = International sample with imaging at p< .05, Bonferroni corrected.

^dPHAMOUS < International sample = Internationalsample with imaging at p< .05, Bonferroni corrected.

^ePHAMOUS < International sample = Internationalsample with imaging at p< .05, Bonferroni corrected.

^fPHAMOUS < International sample < Internationalsample with imaging at p< .05, Bonferroni corrected.

^g211 patients with medication information in the whole international sample. 149 patients with also illness duration information were included in the ANOVA analysis.

^hdemonstrated in olanzapine-equivalent dosage (mg/day).

Abbreviations: PANSS, the Positive and Negative Syndrome Scale; A, atypical antipsychotics; T, typical antipsychotics; NA, not available.

Factorization of PANSS using OPNMF

OPNMF (25,27) decomposes given data (PANSS) into two non-negative matrices; a basis matrix (dictionary) with factors as columns which can be readily generalized to new data due to the projective constraint, and a factor-loading matrix representing symptomatology of individual patients along these factors. The orthonormality constraint promotes a sparse, and hence, interpretable representation. For choosing the number of factors, a set of sophisticated evaluation strategies were implemented (Supplement, Figures S1–S2).

We first applied OPNMF to the PANSS scores from the 1545 PHAMOUS patients with the number of factors ranging from two to eleven. The optimal number of factors was identified by using cross-validation in 10,000 split-half analyses. The PHAMOUS sample was split into two halves, and on each split-sample, OPNMF was performed to derive the dictionary. The congruency between item-to-factor assignments, based on its largest coefficient, was assessed using the adjusted Rand-index (30) and variation of information (31), along with the concordance index (32) between the dictionaries. We also quantified out-of-sample reconstruction error by projecting one split-sample data on the dictionary from the other split-sample. A lower increase in out-of-sample error compared to within-sample reconstruction error indicates better generalizability. This split-sample analysis was repeated on the international dataset. Additional bootstrap and 10-fold cross-validation analyses were conducted on each of the two samples independently.

Most critically, we assessed stability and generalizability between the factorizations of the PHAMOUS sample (good for learning a structure due to size) and the international sample (good for validation due to heterogeneity). We performed OPNMF independently on the bootstrapped samples from each dataset. The resulting factorizations were then compared using the approaches described above. That is, for each number of factors, we assessed stability by comparing the dictionaries obtained from factorization of bootstrapped samples (PHAMOUS vs. international). Most importantly, we also evaluated generalization to new data by measuring increase in reconstruction error for the international data following projection on the PHAMOUS dictionary. This cross-sample evaluation was repeated after accounting for between-dataset differences in sample size, age, and illness duration. Leave-one-site-out validation, on both the PHAMOUS and international samples, was performed to check for site bias. We also repeated all analyses after removing outliers or including the repeated PANSS measurements (Supplement). Factorizations of the pooled (PHAMOUS+international) sample, as well as the stability and accuracy of PHAMOUS generated dictionaries in estimating out-of-sample loadings/item-scores were additionally assessed (Supplement).

After identifying the optimal PANSS factorization, the international sample was projected onto this PHAMOUS-derived dictionary to obtain factor-loadings for subsequent analyses (except for longitudinal analysis as reassessments were only available in the PHAMOUS sample), in order to avoid “double dipping”/“leakage” that would occur if scores were analyzed in the same dataset used to derive the dictionary.

Internal consistency and relationship among variables

Internal consistency of the optimal OPNMF model, as well as the PANSS subscales (as reference), was assessed using Cronbach’s alpha, where higher values indicate more closely related items within a set. Relationships between the OPNMF factor-loadings were assessed using linear and partial correlations (controlling for symptom severity, i.e., total PANSS score), including bootstrap stability analyses. The OPNMF factor-loadings were correlated with the three PANSS subscales. Correlations between individual items were also computed. For comparison, we performed an exploratory factor analysis (EFA) on the PHAMOUS sample and a confirmatory factor analysis (CFA) on the international sample, as well as a principal component analysis (PCA) on both samples. Effects of gender, age, illness duration, and symptom severity on the OPNMF factor-loadings were analyzed in the international sample (N=393, with complete information). Following a MANOVA to assess effects on all the loadings, individual 4-way ANOVAs were performed on each loading to identify their association with the demographic and clinical features (corroborated by bootstrap and leave-one-site-out analyses). Current drug dosages of anti-psychotic medication, available for 149 patients, were olanzapine-equivalent transformed (33), and included in the 4-way ANOVA models for a supplementary analysis.

Psychopathological subtypes

After adjusting for age, gender, illness duration, and symptom severity, factor-loadings were used as features for clustering patients into psychopathological subtypes. After confirming the data clusterability (34), we applied fuzzy c-means clustering (35) which provided cluster membership likelihoods for each patient. The optimal cluster number was determined based on the fuzzy silhouette index (36), the Xie and Beni index (37), and partition entropy (35). Stability was tested by leave-one-site-out replication, subsampling, and bootstrap resampling (Supplement). Given the heterogeneous nature of schizophrenia and the observation of multiple patients with ambiguous memberships, a cutoff over the membership likelihoods was adopted to remove cluster ambiguous patients. For this, additional evidence from Gaussian mixture modeling (GMM) was considered. Specifically, patients were clustered again using GMM, and the optimal cluster number was determined by Bayesian information criterion (38). After assigning patients to the clusters, we took the intersection of the c-means and GMM results. A cutoff was chosen, based on the c-means membership likelihoods that well discriminated the patients inside the intersection from those outside, while also retaining a decent sample size for classification. This filtering step is critical, as ambiguous patients might obscure otherwise classifiable rsFC patterns for the identified subtypes. Afterwards, differences between subtypes regarding factor-loadings and demographic and clinical features were ascertained by permutation tests (39). To assess longitudinal stability, the same c-means clustering was applied to the repeated assessments of the PHAMOUS sample. The optimal dictionary, identified on the PHAMOUS 1545 patients without repeatedly assessed PANSS scores, was used for projection to yield the factor-loadings. After excluding ambiguous cases, patients assigned to the same clusters in both initial and follow-up stages were regarded as longitudinally stable (Supplement). For comparison, the same clustering was additionally performed on the factor-loadings without any covariates or symptom severity adjustment, as well as on the PANSS subscales/items both with and without covariates adjustment (Supplement).

Classifying psychopathological subtypes from rsFC

Multivariate classification analysis was conducted on patients from the international sample for whom imaging data was available after excluding those with ambiguous subtype assignment, low image quality, and excessive head motion (N=84; Figure S24). After standard preprocessing (Supplement), regional time-series were extracted based on a parcellation scheme with 600 cortical (40) and 36 subcortical parcels (41), adjusted for confounders (42), and used to compute the functional connectome. We tested each parcel for whether its pattern of rsFC to all other parcels allowed to classify subtype membership in novel subjects. Resulting parcel-wise accuracies yielded a whole-brain map indicating the classifiable power of each parcel’s connectivity profile. The radial basis function kernel support vector machine classifier, which can deal with the potentially nonlinear relationship between the psychopathological and the neural spaces, was employed. A stratified 10-fold cross-validation was implemented to assess the out-of-sample classification performance (Figure S25). Effects of age, gender, site, illness duration, symptom severity, and head-motion parameters were adjusted using a linear-regression model fitted only in the training sample (43). Significance of the parcel-wise accuracy was estimated by permutation tests, followed by false discovery rate (FDR) correction for multiple comparisons (Supplement). Parcels surviving FDR were functionally characterized (http://brainmap.org/)(44) (Supplement).

RESULTS:

Dimensions of psychopathology

The most robust and generalizable model consisted of three factors for the PHAMOUS sample (Figure 1A), which effectively combined the positive and affective symptoms compared to the optimal four-factor model for the international dataset (Figure 1B). The additional factor in the international dataset may relate to the higher prevalence of psychotic symptoms, particularly auditory hallucinations, compared to the PHAMOUS sample, which contains more chronic patients (Table 1). Consequently, a four-factor model was identified as the most stable and, importantly, generalizable model of psychopathology in the cross-sample evaluation (Figure 1C). The first dimension mainly represents negative symptoms, such as blunted affect and apathy (Figure 1D). The second represents positive symptoms, e.g., delusions and hallucinations. The third factor comprises symptoms, such as depression, anxiety, and tension, reflecting an affective dimension, while the last factor represents cognitive impairments. Notably, only a few items contributed to multiple dimensions, e.g., active social avoidance contributed to both negative and affective factors.

Figure 1. Split-half cross-validation (10,000 repetitions) of stability and generalizability of the factor-solutions derived by OPNMF^a.

^aThe three indices; aRI, VI and CI demonstrate the factor stability, while out-of-sample increased RE reflects the performance of generalizability. Box-plots show stability and generalizability results of the factor solutions. Higher values for aRI and CI (upper row) indicate higher stability. Lower values for VI and out-of-sample increase in RE (bottom row) indicate better stability and generalizability, respectively. For the box-plots, the red line depicts the median, the green diamond depicts the mean, and the whiskers represent the 5th and 95th percentiles. For the factor-models, the weight of an item in assigning to a specific psychopathological factor (columns of the matrix) is color coded according to the coefficients by a heat color map, from grey (minimum) to dark red (maximum). aRI, adjusted Rand index; VI, variation of information; CI, concordance index; RE, reconstruction error. Panel A) shows the best factor solution derived from the PHAMOUS data (1545 patients). According to the four aforementioned evaluation indices, a three-factor model was indicated as the best since both of the mean and median values for VI and out-of-sample increase in RE achieve the lowest, and the aRI and CI reach the highest at that point. Panel B) shows the best factor solution derived from the international sample (490 patients). As shown, four factors is the optimal solution, as mean and median values of VI and out-of-sample increase in RE achieve the local minimum, while the aRI reaches maxima and the CI reaches a local maximum. Panel C) shows the best factor solution identified by the bootstrap comparison of the two datasets (PHAMOUS vs. international). As shown, a four-factor solution is optimal, as the mean and median values of adjusted RI and CI reach the maximum, while the mean and median values of VI and median value of out-of-sample increase in RE achieve the minimum. Panel D) shows the most stable and generalizable four-factor structure derived from the PHAMOUS sample, serving as the best basis for future studies. This four-factor model consists of a negative (factor 1), a positive (factor 2), an affective (factor 3), and a cognitive (factor 4) factor which were named based on the items they contained.

All findings were fully confirmed by (i) bootstrap and 10-fold cross-validation, (ii) removing outliers (18 patients), (iii) adding PANSS data from follow-up examination in the PHAMOUS sample, (iv) leave-one-site-out validation, (v) accounting for between-dataset differences in sample size, age, and illness duration, (vi) pooling the two datasets with cross-validation and out-of-sample generalization assessments, and (vii) loading/item-score predictions across factor-solutions, bootstrapped samples, and sites (Figures S3–S11).

Internal consistency and relationship among variables

Items within a factor showed higher and more homogeneous positive correlations (while fewer anti-correlations) for the OPNMF factors than the PANSS subscales (Figure 2A–2B, Figure S12). Internal consistency of our OPNMF four-factor structure (positive: Cronbach’s alpha=0.75; negative: 0.92; affective: 0.85; cognitive: 0.83) was on average higher than that for the PANSS subscales (positive: 0.72; negative: 0.87; GPS: 0.87), previously reported factor models (ranging from 0.6 [excited] to 0.9 [negative])(7–9,45), and the EFA models derived from the current sample (0.49-0.91, Table S2). All PHAMOUS-derived 4-7 factor EFA-models could not be confirmed in the international sample, i.e., inadequate fit (Table S2). Compared to PCA, OPNMF showed better generalizability (Figure S13). The positive and negative factors were highly correlated with the positive and negative PANSS subscales, respectively, both before (r=.92 and r=.97) and after (r=.85 and r=.89) controlling for symptom severity (Figure S14). Interestingly, after adjusting for symptom severity, the cognitive factor did not correlate with GPS (r=.02).

Figure 2. Inter-item correlations, relationship between factors, socio-demographic and clinical information^a.

^aThe four-factor structure, derived from the PHAMOUS sample with initial measure of PANSS scores, was adopted as the reference on which the international sample was projected to derive the factor-loadings. Heat maps A-B) show inter-item correlations for (A) the original PANSS subscales, and (B) the current OPNMF four-factor representation of psychopathology after controlling for symptom severity (total PANSS score). Correlation strength is color-coded (light yellow to red: positive correlations; cyan to blue: negative correlations). Box-plot C) shows the bootstrap results (repeated 10,000 times) for the Pearson correlations between the four factor-loadings. Bootstrap samples were drawn with replacement from the original international sample, and then the correlation analysis was done on them. The red line depicts the median, the green diamond depicts the mean, and the whiskers represent the 5th and 95th percentiles. Graphs D-E) show effects of socio-demographic and clinical features on the four factor-loadings. D). Scatter plots show 4-way ANOVA results of the significant negative association between age (adjusted for gender, illness duration, and total PANSS score) and the cognitive loading (p = .033), as well as the significant positive associations between the symptom severity (total PANSS score) and the four factor-loadings (negative: p = 4.98E-105; positive: p = 2.52E-51; affective: p = 1.75E-133; cognitive: p = 1.58E-106) after adjusting for age, gender, and illness duration. Regression lines are depicted with a 95% confidence interval on the fitted values. E) shows bootstrap results for the 4-way ANOVA analysis. Bootstrap samples were drawn with replacement from the original international sample and then the ANOVAs were done on them. Boxes refer to the beta values. The red line depicts the median, the green diamond depicts the mean, and the whiskers represent the 5th and 95th percentiles.* (Mdn, p < .05), ** (M and Mdn, p < .05).

Over individual patients, the loadings on our four factors were significantly inter-correlated, with negative and positive factors showing lowest (r=.32, averaged over 10,000 bootstraps), negative and affective factors highest (r=.70) correlations (Figure 2C). After controlling for symptom severity, positive and negative factors became anti-correlated (r=−.59, Figure S15).

MANOVA revealed a significant influence of symptom severity on the joint factor-loadings (p<.001). Follow-up 4-way ANOVAs showed that symptom severity had a significant effect on each factor (all p<.001, all 6ϐ0.07, Figure 2D). The cognitive factor showed a trend towards a positive relationship with illness duration (p=.081, ϐ=0.014) and a significant negative relationship with age (p=.033, ϐ=−0.015, Figure 2D), though both covariates were collinear (r=.65). In contrast, loadings on the negative factor were higher for older individuals (p=.11, ϐ=0.016) and lower for those with longer illness duration (p=.18, ϐ=−0.015). Gender-differences were not observed in any factor. Bootstrap (Figure 2E) and leave-one-site-out analyses corroborated the aforementioned ANOVA findings (Supplement). Adding olanzapine-equivalent dosage to the 4-way ANOVA did not reveal any significant association with medication.

Psychopathological subtypes

Fuzzy c-means clustering on the adjusted loadings revealed an optimal two-cluster solution (Figure 3A–3B, Figures S16–S17). Although GMM demonstrated an optimal three-cluster solution, one of the clusters was diffusely distributed in space containing patients from both the c-means clusters. This GMM cluster was excluded as it would not represent any specific subtype (Figure 3C). Patients inside the c-means GMM intersection had higher c-means membership likelihoods (roughly>0.7) to belong to their own cluster than those outside the intersection (p<.01, Wilcoxon rank-sum test; Figure 3D). We chose the “cluster cores” using the likelihoods of c-means with a cutoff of 0.7. Resultantly, two core subtypes were defined after filtering out 50 ambiguous patients from each of the two c-means clusters (Figure 3E). The first subtype showed a psychopathological profile dominated by negative and affective symptomatology (subtype A). The other featured prominent positive symptoms (subtype B; all p<.001 in permutation tests) (Figure 3F, Figure S18). Importantly, subtypes did not differ in gender distribution, age, or illness duration (all p>.05), but subtype B showed higher symptom severity (p=.008). The same two-cluster solution was replicated on the PHAMOUS patients with complete demographic and clinical information (N=1326; 56% of the 603 ambiguous patients in subtype B when hard-clustered), and those with repeatedly assessed PANSS scores (N=527; 45% ambiguous). Almost 80% of the reassessed patients retained their subtype, with subtype A being more stable (85%) (Supplement). The additional clustering analyses supported our four-factor model with covariates adjustment for a clinically meaningful subtyping (Figures S19–S23).

Figure 3. Fuzzy c-means clustering results of patient subgroups based on the loadings of the generalizable four-factor structure^a.

^aPanel A) shows the internal validity indices used for determining the optimal cluster number. Higher values of FSI (in triangle) and lower values of XB and PE (in inverted triangle) indicate a better clustering quality. The maximum for FSI and the minimums for XB and PE all suggested a two-cluster solution. FSI and XB reflect the compactness and separation of the generated clusters, while PE reflects the fuzziness of the cluster partition, i.e., the uncertainty of the patients to be assigned to a certain cluster. Box-plot B) shows results of the assessment of clustering stability based on the subsampling technique. The cluster number two reaches the highest aRI. aRI reflects the convergent assignment of the patient-pairs to the clusters between the sub-samples and the original sample. C) Four-dimensional visualization of the optimal three GMM clusters determined by the Bayesian information criterion (a higher value indicates a better clustering solution). Magnitude of the cognitive loading was color-coded differently for the three clusters (cluster 1, corresponds to the cluster I (i.e., subtype A) in fuzzy c-means, yellow to Modena; cluster 2, corresponds to the cluster II (i.e., subtype B) in fuzzy c-means, blue to shallow flaxe; cluster 3, i.e., the excluded diffused cluster which would not present any specific subtype, black to light grey). Boxplot D) shows the fuzzy c-means membership likelihoods of the patients inside and outside the intersection of the c-means and GMM clustering results. The black line indicates a heuristic cutoff of 0.7. Panel E) shows a four-dimensional visualization of the optimal fuzzy c-means two-cluster solution. Ambiguous assignments were defined by membership likelihoods < 0.7, which was selected by interacting with GMM. Those subtype ambiguous patients are shown in small dots, X represents the centroid. Magnitude of the cognitive loading is color-coded differently for the two clusters (cluster I, yellow to Modena; cluster II, blue to shallow flaxe). Grouped box-plots F) show the between-subtype (without subtype ambiguous patients) comparison results of the four factor-loadings, age, illness duration, and total PANSS score. Cluster I is dominated by negative and affective symptoms (i.e., subtype A), cluster II is significantly prominent in positive symptom expressions (i.e., subtype B). The black dashed line depicts the median, the yellow diamond depicts the mean, and the whiskers represent the 5th and 95th percentiles. *p < .01; **p < .001. FSI, fuzzy silhouette index; XB, Xie and Beni index; PE, partition entropy; GMM, Gaussian mixture modeling.

Classifying psychopathological subtypes from rsFC

The rsFC profile of the parcel located in right ventro-medial prefrontal cortex (vmPFC) yielded the highest out-of-sample classification accuracy (70% of patients not used for training were assigned to the correct psychopathological subtypes), followed by parcels in the right temporoparietal junction (TPJ), the bilateral precuneus, and the posterior cingulate cortex (PCC). Permutation tests showed that the top 104 classifiable parcels were significant (p<.05) against chance (i.e., randomized subtype labels), and 53 parcels survived FDR correction (q<.05). Of note, parcels are labeled by their microanatomical location, with their functional implications (Table S4). Classification with additional global mean signal removal or with a RS-subcortical parcellation replacement (7 parcels [46]; as a control analysis instead of the finer Brainnetome subcortical parcellation), replicated these results (Supplement; Figure S26).

DISCUSSION

By factorizing the PANSS scores from a large sample using OPNMF and cross-validating the results in a heterogeneous multi-site dataset, we revealed a robust, replicable, and generalizable four-factor structure comprising negative, positive, affective, and cognitive dimensions across populations, settings, and medical systems. Based on this four-factor structure, two core psychopathological subtypes were obtained which showed good longitudinal stability and could be discriminated by regional rsFC patterns, with right vmPFC showing the highest (70%) classification accuracy.

Relationship to previous factor models of the PANSS

The three PANSS subscales do not reflect the latent structure of this inventory well (5–7). In turn, our model represents a stable, generalizable and well interpretable description of schizophrenia psychopathology suited for representing the full range of acute and chronic symptoms. Resonating with this view, a pyramidical model proposed by the PANSS developers comprised four components (6). Three of these (negative, positive, and affective dimensions) showed good agreement with our model. The fourth component, however, isolated only excitement, while cognitive disturbances were distributed across all dimensions or discarded. Such a representation is obviously at odds with the importance of cognitive dysfunction, and has prompted the proposal of more complex models (7–11), e.g., a recent five-factor model reflecting negative, positive, depressed, excited, and cognitive dimensions (45). However, replicability, external validity, and generalization remain a concern for these models (47). As a striking example, White and colleagues (12) found that, none of 20 tested models fit their data adequately, and put forward a new pentagonal model, which later (together with 24 others) also could not be confirmed (13). In the same study (14), the authors developed an improved five-factor model using 10-fold cross-validation. However, it still failed to be confirmed, along with 31 other five-factor models, in a later study involving two large Chinese samples (15). In our sample, inadequate fit for EFA models with 4-7 factors was also manifested, and the fifth OPNMF factor, compared to the four factors, showed the poorest out-of-sample loading predictions (Figure S10B). These facts, as a whole, point to a fundamental instability of five-factor models (9–15,44). Addressing these concerns, the current work was not only based on a large sample for model identification, but, importantly, focused strongly on cross-validated stability and out-of-sample generalization. Critically, the external validation was based on a heterogeneous, international sample, and the optimal model suggested a single factor to combine both the cognitive and excited symptoms. This view is corroborated by observations that cognitive and excited symptomatology is highly correlated (45) and share similar neurobiological substrates (48,49).

Internal consistency and relationship among variables

Although we identified the optimal representation by its robustness and ability to generalize to new populations, the positive and negative dimensions of our model also showed better internal consistency than the PANSS subscales while differentiating the broad “general psychopathology”. Moreover, our affective and cognitive factors showed higher internal consistency than those reported in previous factor-models (7,8,10,45). Finally, correlations between individual items within OPNMF factors were higher and more homogeneous compared to the PANSS subscales.

Matching previous reports (4,8,45), negative and positive factors from our model were least related before, and showed a strong anti-correlation after, controlling for symptom severity. The inversed age vs. illness duration effect on negative symptoms implies that this effect may be more related to age than illness duration. This is intriguing from the perspective of early aging/degeneration, but needs to be taken with caution, as age and illness duration are highly correlated. Co-linear variables in a single linear regression model make it difficult to disentangle their respective effects on the negative factor, as well as on the cognitive factor (50).

Psychopathological subtypes, longitudinal stability, and neurobiological differentiability

Our results revealed two distinct schizophrenia subtypes featuring predominantly positive and negative symptoms, respectively. The subtypes were longitudinally stable and could be classified from neuroimaging data. Such a positive-negative dichotomy has been widely supported (51,52). Finer distinctions have been proposed, but show poor replicability (18–20). The inconsistency of a finer subtyping may relate to idiosyncrasies in small samples from a single geographical region and to the lack of explicit analyses of stability and replicability. Moreover, longitudinal stability of our new subtypes was higher than that reported for traditional clinical subtypes or for a positive/negative/mixed topology (53–55). Interestingly, we found subtype A to be particularly stable. Previous studies indicated that both mixed and negative symptom states increase over time, whereas psychotic expressions usually diminish outside acute episodes and over time (54,55). Future studies with a longer follow-up duration are desired; the mean of 1.35 years’ follow-up assessed in the present study is not a long period in schizophrenia. Additionally, the employed soft-clustering method better accommodates ambiguous patients compared to previous hard-clustering methods, and patients with ambiguous memberships can be furthermore filtered out with appropriate cutoffs to improve the ability of detecting neurobiological distinctions between subtypes. Nonetheless, the cutoff value chosen in the present study should be noted as heuristic. The cluster ambiguous patients might represent a transient group lying in-between the two more differentiated subtypes.

The current top classifiable brain regions are all implicated in schizophrenia pathophysiology and processes relevant to the psychopathological distinction (22,56–60). Previous findings in the literature relating fMRI parameters to differential symptoms, so far, exclusively relied on group-level analyses, while our approach bridged an important gap between neurobiological divergence and distinct symptomatic patterns at the individual level. To our knowledge, it is the first study to successfully classify psychopathological subtypes in schizophrenia. Of note, the current classification accuracy was similar to those previously reported for classifications of schizophrenia patients vs. healthy participants (61–63). The demonstrated neurobiological differentiability corroborates the currently identified schizophrenia subtypes expressed along the four OPNMF dimensions.

Limitations and Considerations

We assessed factor-structure, subtypes, and their neurobiological differentiations with a particular emphasis on robustness and generalization. This conservative approach seems necessary given current concerns of non-replicability in biomedical research, but may have contributed to the fact that we corroborated the clinically well-established positive-negative distinction rather than identifying more differentiated subtypes. We note, though, that a recent imaging-based clustering also provided evidence for two subtypes (64), and stress that the current analysis of a large, heterogeneous sample did not reveal any evidence for a more fine-grained differentiation among the patients. It, thus, remains to be seen whether an additional differentiation between these two core subtypes may be robustly revealed by analyzing substantially larger samples, or whether previously proposed additional subtypes represent distinctions that could be found in a particular dataset but are not universally present. We also acknowledge that patients were on their regular medication as prescribed by the attending psychiatrists, and the current results might thus be confounded by direct and indirect mediation effects thereof. However, it stands to reason that a multi-site study, pooling patients from different psychiatrists with differential medication strategies, will render medication largely as a source of random variation in our data. Such noise would effectively make it harder to identify generalizable psychopathological factors, robust subtypes, and, in particular, train models that work well for out-of-sample classification of subtype membership. We would thus argue that the current results should not be driven by medication effects, but rather represent general structures of psychopathology and schizophrenia subtypes. In addition, rs-fMRI has its own limitations, such as variability across scanning sessions and the issue of confounding factors (65–67). We focused on rs-fMRI because it could temporally better map the likewise state-dependent psychopathology compared to structural MRI.

Using advanced machine-learning with cross-sample validation, the present study suggested a stable and generalizable four-factor model of PANSS. This representation allowed for the definition of a reliable positive-negative subtype differentiation that showed good longitudinal stability and a neurobiological divergence in rsFC. Overall, the present work further disentangled the heterogeneity of schizophrenia, possibly allowing for the design of more specifically targeted treatments.

Figure 4. Classifying psychopathological subtypes from resting-state functional connectivity³.

^aPanel A) illustrates the brain parcellation scheme (600 cortical parcels plus 36 subcortical parcels), and the resting-state functional connectivity matrix that was constructed based on this parcellation system. In parcel-wise classification analysis, one column of the connectivity matrix was taken to represent the functional connectivity pattern for a single parcel. B) Cortical surface rendering and subcortical axial slices show parcel-wise classification results for those parcels which survived FDR correction (q < .05), demonstrating a neurobiological divergence between the two identified psychopathological subtypes of schizophrenia. C) Cortical surface rendering and subcortical axial slices show parcel-wise classification results for the whole brain. The balanced classification accuracy is color-coded from light grey to dark red.

KEY RESOURCES TABLE

Resource Type	Specific Reagent or Resource	Source or Reference	Identifiers	Additional Information
Add additional rows as needed for each resource type	Include species and sex when applicable.	Include name of manufacturer, company, repository, individual, or research lab. Include PMID or DOI for references; use “this paper” if new.	Include catalog numbers, stock numbers, database IDs or accession numbers, and/or RRIDs. RRIDs are highly encouraged; search for RRIDs at https://scicrunch.org/resources.	Include any additional information or notes if necessary.
Deposited Data; Public Database	COBRE	http://fcon_1000.projects.nitrc.org/indi/retro/cobre.html	RRID:SCR_010482
Software; Algorithm	a Dimensions and Clutsering Tool for assessing schizophrenia Symptomatology (DCTS)	this paper; http://webtools.inm7.de/sczDCTS	N/A
Software; Algorithm	Matlab R2016a	Mathworks	RRID:SCR_001622
Software; Algorithm	R version 3.4.2	http://www.r-project.org/	RRID:SCR_001905
Software; Algorithm	mclust package (version 5.3) in R version 3.4.2	http://CRAN.R-project.org/package=mclust	N/A
Software; Algorithm	LIBSVM	https://www.csie.ntu.edu.tw/~cjlin/libsvm/	RRID:SCR_010243
Software; Algorithm	SPM12	Wellcome Centre for Human Neuroimaging; https://www.fil.ion.ucl.ac.uk/spm/software/spm12/	RRID:SCR_007037
Software; Algorithm	CAT12	Structural Brain Mapping Group at the Departments of Psychiatry and Neurology, University of Jena; http://www.neuro.uni-jena.de/cat/	N/A
Other	Shaefer local-global cortical parcellation	https://github.com/ThomasYeoLab/CBIG/tree/master/stable_projects/brain_parcellation/Schaefer2018_LocalGlobal	N/A
Other	Brainnetome atlas	http://atlas.brainnetome.org	RRID:SCR_014091