Disease Progression Patterns of Brain Morphology in Schizophrenia: More Progressed Stages in Treatment Resistance

Abstract

Background and Hypothesis

Given the heterogeneity and possible disease progression in schizophrenia, identifying the neurobiological subtypes and progression patterns in each patient may lead to novel biomarkers. Here, we adopted data-driven machine-learning techniques to identify the progression patterns of brain morphological changes in schizophrenia and investigate the association with treatment resistance.

Study Design

In this cross-sectional multicenter study, we included 177 patients with schizophrenia, characterized by treatment response or resistance, with 3D T1-weighted magnetic resonance imaging. Cortical thickness and subcortical volumes calculated by FreeSurfer were converted into z scores using 73 healthy controls data. The Subtype and Stage Inference (SuStaIn) algorithm was used for unsupervised machine-learning analysis.

Study Results

SuStaIn identified 3 different subtypes: (1) subcortical volume reduction (SC) type (73 patients), in which volume reduction of subcortical structures occurs first and moderate cortical thinning follows, (2) globus pallidus hypertrophy and cortical thinning (GP-CX) type (42 patients), in which globus pallidus hypertrophy initially occurs followed by progressive cortical thinning, and (3) cortical thinning (pure CX) type (39 patients), in which thinning of the insular and lateral temporal lobe cortices primarily happens. The remaining 23 patients were assigned to baseline stage of progression (no change). SuStaIn also found 84 stages of progression, and treatment-resistant schizophrenia showed significantly more progressed stages than treatment-responsive cases (P = .001). The GP-CX type presented earlier stages than the pure CX type (P = .009).

Conclusions

The brain morphological progressions in schizophrenia can be classified into 3 subtypes, and treatment resistance was associated with more progressed stages, which may suggest a novel biomarker.

Introduction

Schizophrenia is a common psychiatric disorder presenting with psychotic symptoms as well as negative and cognitive symptoms.¹ Despite longstanding and continuous efforts, we have not identified any distinct pathophysiology or established objective biomarkers in schizophrenia. While the diagnosis of schizophrenia is still based on psychiatric symptoms, patients with schizophrenia often show heterogeneous symptoms and treatment response,^2,3 calling into question whether it represents a single disease, particularly in terms of neurotransmitter systems.⁴ In addition to symptom heterogeneity, treatment response is also diverse. For example, treatment-resistant schizophrenia (TRS) defines a distinct subpopulation showing poor response to conventional pharmacological treatment⁵ and, as a result, a form of the illness associated with serious social and economic burden.⁶ The neurobiological basis of TRS remains to be elucidated, despite numerous strategies including neuroimaging studies.^7–9

To address this disease heterogeneity, studies have proposed schizophrenia subtypes based on symptoms^10,11 as well as brain structures.¹² In the latter, eg, each individual’s brain structural abnormality is categorized into 2 distinct subtypes by machine learning; however, to date such brain morphological subtypes have shown little relationship with clinical symptoms.¹² In employing such a strategy, it is important to acknowledge that such brain morphological abnormalities may be progressive and involve cortical thinning in the temporal or frontal lobes.^13,14 TRS has been associated with longer duration of untreated psychosis,^15,16 raising the possibility that TRS may be caused by more disease progression.

In light of the above, categorization that incorporates staging may prove valuable in our understanding of treatment resistance in schizophrenia. In this regard, machine-learning analysis has been increasingly applied to uncovering patterns in clinical parameters that may translate to personalized, more reliable biomarkers.^17,18 Subtype and Stage Inference (SuStaIn) is an unsupervised machine-learning algorithm to uncover data-driven disease phenotypes with temporal progression patterns, and it has been widely utilized to identify disease subtypes and stages.^19–22 SuStaIn identifies distinct temporal progression trajectories and subtypes from cross-sectional data,¹⁹ based on the event-based model, which constructs discrete pictures of progression,²³ and reformulation of the events.¹⁹ Staging is the degree of disease progression as inferred by machine-learning models and is only model based. While, in some diseases, progression is related to duration of disease,²⁴ disease progression does not always simply correlate with time-course. In addition, unsupervised learning is well suited for detecting hidden patterns and structures in unlabeled data. Psychiatry has used symptom-based diagnosis and assessment, but such symptom-based approaches have so far failed to find useful objective biomarkers. In this regard, we hypothesized that patterns derived by unsupervised learning from biological data alone, rather than supervised learning with the risk of bias from symptom-based labels, might yield subtyping and staging that better reflects the biological aspects of disease.

Here, we applied the SuStaIn algorithm to classify disease progression patterns and staging of brain morphology in schizophrenia, with the goal of identifying distinct biological subtypes in the context of illness progression and associations with clinical measures. We hypothesized that TRS may be associated with more progressed disease staging; in addition, we investigated the consistency of anatomical subtype categorizations with previously published data¹² as well as relationship with other clinical characteristics.

Methods

Participants

We analyzed international multicenter cross-sectional neuroimaging data comprising 177 patients with schizophrenia and 73 healthy controls (HCs): 54 patients with schizophrenia (24 TRS, 30 non-TRS) and 28 HCs from Komagino hospital,²⁵ Tokyo, Japan, 70 patients with schizophrenia (49 TRS, 21 non-TRS) and 21 HCs from the Centre for Addiction and Mental Health (CAMH),²⁶ Toronto, Canada, and 53 patients with schizophrenia (25 TRS, 28 non-TRS) and 24 HCs from Shimofusa Psychiatric Medical Center, Chiba, Japan (table 1). During the FreeSurfer processing, we excluded some participants due to poor segmentation quality, which resulted in slightly reducing the number of subjects from our past studies.^25,26 In each cohort, there were no significant differences in age and sex between the schizophrenia and HC groups.

Table 1.

Demographics and Subtype/Staging of Participants From the Three Institutes

	Cohort 1 (Komagino)	Cohort 2 (Toronto)	Cohort 3 (Shimofusa)
HC	N = 28	N = 21	N = 24
Age (y)	46.0 (18)a	36.0 (23)a	41.5 (18)a
Sex (M:F)	12:16b	15:6b	14:10b
Schizophrenia	N = 54	N = 70	N = 53
Age (y)	43.5 (17)a	46.5 (22)a	39.0 (17)a
Sex (M:F)	24:30b	53:17b	29:24b
TRS (N)	24	49	25
Onset age (y)	25.0 (11)	23.0 (9)	20.0 (10)
Illness duration (y)	14.0 (15.5)	20.0 (20.3)c	15.0 (17.5)
Education (y)	12.0 (3)	12.5 (2)c	12.0 (3)
Chlorpromazine-equivalent dose of antipsychotics (CP)	600 (500)	493.75 (300)c	450 (570)
Treatment with CLZ (N)	0	49	6
PANSS-P	14.5 (17)	13.0 (10)c	17.0 (10)
PANSS-N	22.0 (17)	17.5 (5)c	18.0 (8)
PANSS-G	30.5 (29)	32.5 (11)c	36.0 (16)
PANSS-T	67.5 (65)	65.0 (24)c	70.0 (26)
Subtypes	SC = 15, GP-CX = 19, pure CX = 8, stage 0 = 12	SC = 31, GP-CX = 13, pure CX = 20, stage 0 = 6	SC = 27, GP-CX = 10, pure CX = 11, stage 0 = 5
Staging	5.0 (9)	7.0 (11)	6.0 (10)

Continuous variables are shown as median (IQR). Note: CLZ, clozapine; GP-CX type, globus pallidus hypertrophy and cortical thinning type; PANSS, Positive and Negative Syndrome Scale (P: positive, N: negative, G: general, T: total); pure CX type, cortical thinning type; SC type, subcortical volume reduction type; TRS, treatment-resistant schizophrenia.

^aNo significant differences between HC and Schizophrenia in each cohort (P = .395, .072, and .656, respectively, Mann-Whitney U tests).

^bNo significant differences between HC and Schizophrenia in each cohort (P = .891, .692, and .767, respectively, χ² tests).

^cMissing in 4 patients.

Participants partly overlapped with previous studies in which the same inclusion/exclusion criteria and clinical evaluations were used.^9,25–29 Patients were diagnosed with schizophrenia based on the Diagnostic and Statistical Manual of Mental Disorders 4th Ed (DSM-IV).³⁰ The Positive and Negative Syndrome Scale (PANSS)³¹ and the Clinical Global Impression Severity Scale (CGI-S)³² were used for assessment of clinical symptoms. TRS was determined by the modified Treatment Response and Resistance in Psychosis (TRRIP) Working Group Consensus criteria.³³ Treatment response was defined by (1) CGI-S score ≤3, (2) PANSS positive symptom item scores ≤3, and (3) no symptomatic relapse in the previous 3 months. In contrast, inadequate treatment response was defined by (1) CGI-S score ≥4, and (2) ≥4 on at least 2 PANSS positive symptom items after adequate antipsychotic trials. Response to past antipsychotic trials was determined based on medical records. We also confirmed no history of psychiatric illness in HCs by using the Mini-International Neuropsychiatric Interview (MINI).³⁴ The following exclusion criteria were applied to all participants: (1) substance abuse or dependence within the past 6 months; (2) positive urine drug screen at inclusion or before the MRI scan; (3) history of head trauma resulting in unconsciousness for >30 minutes; or (4) an unstable physical illness or neurological disorder.

All participants provided written informed consent, and the study protocol was approved by the Ethics committees at each institute.

MRI Acquisition and Preprocessing

Participants underwent 3D T1-weighted structural MRI scans on the following protocols: (1) at the Komagino Hospital, 3 T Signa HDxt scanner (GE Healthcare) with an 8-channel head coil (BRAVO, echo time [TE] = 2.8 ms, repetition time [TR] = 6.4 ms, inversion time [TI] = 650 ms, flip angle = 8°, field of view [FOV] = 230 mm, matrix size = 256 × 256, slice thickness = 0.9 mm), (2) at the Centre for Addiction and Mental Health, a 3 T GE Discovery R750 scanner (GE Healthcare) with an 8-channel head coil (BRAVO, TE = 3 ms, TR = 6.74 ms, TI = 650 ms, flip angle = 8°, FOV = 230 mm, matrix size = 256 × 256, slice thickness = 0.9 mm), (3) at the Shimofusa Psychiatric Medical Center, a 1.5 T Signa Explorer (GE Healthcare) with a 12-channel head coil (FSPGR, TE = 5.1 ms, TR = 12.2 ms, TI = 913 ms, flip angle = 25°, FOV = 256 mm, matrix size = 256 × 256, slice thickness = 1.0 mm).

We used FreeSurfer software (v.6.0, https://surfer.nmr.mgh.harvard.edu) to calculate cortical thickness (CT) and subcortical gray matter (GM) volumes of the whole cerebrum as well as the intracranial volumes (ICV) based on the 3D T1-weighted images of all the participants. Image processing included the removal of non-brain tissues with a hybrid watershed/surface deformation procedure, automated Talairach transformation, and segmentation of the subcortical structure and cortex based on the Desikan-Killiany Atlas. We confirmed segmentation accuracy in all subjects with visual inspection.

SuStaIn Analysis

Firstly, the subcortical GM volumes were corrected for individual’s ICV, and then all CT and subcortical GM volumes were corrected for age and sex. As SuStaIn uses z scores for the machine-learning analysis,¹⁹ we calculated z scores for each cohort. In other words, z score calculations were performed using each HC cohort with the same scanner and protocol at each institute.

It is also necessary to select the relevant regions of interest (ROIs) for obtaining reliable results by machine learning; we chose all ROIs with significant changes in the multicenter mega analysis by ENIGMA consortium,^35,36 one of the most reliable strategies in evaluating brain morphological alteration in schizophrenia. The detailed list of 28 identified ROIs is shown in Supplementary File S1. Since globus pallidus (GP) may show increased volumes,³⁵ we converted the z score of GP by multiplying (−1) to reflect hypertrophy, while the z score of the other ROIs represented cortical thinning or GM volume loss.

Finally, the z scores of the 28 ROIs for the 177 patients with schizophrenia were entered into the SuStaIn algorithm (https://github.com/ucl-pond/SuStaInMatlab). As SuStaIn represents an unsupervised machine-learning strategy, any other information than the z scores, eg, the anatomy of each ROI or clinical data, were not taken into account. The linear z score model and mathematical model underlying the SuStaIn algorithm are described in the previous study¹⁹; steps included model-fitting, convergence, uncertainty estimation, 10-fold cross-validation, and similarity between subtypes. As described previously,^19,21,22 SuStaIn categorized individuals into subtypes and estimated the most likely sequence in which selected ROIs reach different progression stages over time.

Statistical Analysis

Statistical analyses were performed by SPSS (IBM Corp., Version 25.0). Parametric or nonparametric distributions of variables were examined by Shapiro-Wilk test, and the null hypothesis of normal distribution was rejected in all the clinical variables in this study. On the other hand, the corrected CT and subcortical GM volumes in HCs were normally distributed, which should justify the conversion process to z score for SuStaIn.

As a primary analysis, we investigated the relationships of TRS with disease subtypes or staging derived from the SuStaIn analysis. The categorical relationship, ie, TRS/non-TRS vs disease subtypes, was analyzed by χ² test, and the estimated stages between TRS and non-TRS were compared by Mann-Whitney U test. For more exploratory analyses, we examined the association of the subtypes and staging with other clinical characteristics, including onset age, disease duration, medication dose, or PANSS scores. Among the subtypes, continuous variables were compared by Kruskal-Wallis tests, while χ² tests were used for categorical variables. Regarding the staging, Spearman’s rank correlation tests were used to reveal relationships with other variables. In addition, since clozapine (CLZ) has been reported to be involved in brain atrophy progression,^37–39 we conducted an additional analysis of the relationship of CLZ use with staging or subtypes, using Kruskal-Wallis tests with post hoc Bonferroni’s correction and χ² tests. A P < .05 was considered as statistically significant.

Validation for Reproducibility

To confirm the reproducibility of the subtype and staging categorization, we repeated the SuStaIn analysis in each cohort separately. The subtypes and staging results from each additional analysis were compared with the main original results from all the patients, using Cohen’s Kappa coefficient and Kendall’s Tau.

Results

Estimated Subtypes, Stages, and Treatment Resistance

SuStaIn identified 3 different subtypes of brain morphological changes in schizophrenia (figure 1A), ie, (1) subcortical volume reduction (SC) type (73 patients), (2) globus pallidus hypertrophy and cortical thinning (GP-CX) type (42 patients), and (3) cortical thinning (pure CX) type (39 patients). In the SC type, subcortical volume loss, particularly the hippocampi and thalami, initially occurs and cortical thinning follows (left in figure 1A). In the GP-CX type, the GP hypertrophy initially happens, followed by cortical thinning with no severe atrophy of other subcortical structures (middle in figure 1A). In the pure CX type, cortical thinning, particularly in the lateral temporal and insular cortices, mainly occurs and subcortical volumes are not severely affected (right in figure 1A). The forest plots of regional morphological differences among the 3 subtypes are shown in Supplementary File S2, which were generally consistent with the concepts and characteristics of each subtype described above. The remaining 23 patients were assigned to baseline stage of progression (no change) and not categorized into any subtypes.

Fig. 1.

(A) Estimated 3 patterns of brain morphological disease progression in schizophrenia, namely (1) subcortical volume reduction (SC) type, in which volume reduction of subcortical structures occurs first and moderate cortical thinning follows, (2) globus pallidus hypertrophy and cortical thinning (GP-CX) type, in which globus pallidus hypertrophy initially occurs followed by progressive cortical thinning, and (3) cortical thinning (pure CX) type, in which thinning of the insular and lateral temporal lobe cortices primarily happens. (B) Histogram of stage progressions between non-TRS and TRS. (C) Relationship of clozapine use with stage progression. Note: TRS-CLZ, TRS with clozapine use; TRS-non-CLZ, TRS without clozapine use.

SuStaIn also found 84 stages of progression (figure 1A). The histograms of disease stages of each participant in the TRS and non-TRS groups are presented in figure 1B. The TRS group showed significantly more progressed disease stages than non-TRS (P = .001, Mann-Whitney U test). With regard to subtype results, GP-CX type showed significantly less progressed stages than pure CX type (P = .009), and a similar trend was found in comparison to SC type (table 2) although there was no direct association of subtypes with TRS.

Table 2.

Clinical Features Among the Three Subtypes Derived From SuStaIn Analysis

	SC Type (N = 73)	GP-CX Type (N = 42)	Pure CX Type (N = 39)	P
Age (y)	44.0 (22)	40.0 (20)	44.0 (14)	.870c
Sex (M:F)	48:25	24:18	21:18	.415d
TRS (N)	41	23	24	.805d
Onset age (y)	21.0 (13)a	24.0 (8)	23.0 (8)	.950c
Illness duration (y)	15.0 (20)a	19.0 (15.5)b	19.0 (18)	.673c
Education (y)	12.0 (2)a	12.0 (4)b	12.0 (2)	.480c
Chlorpromazine-equivalent dose of antipsychotics (CP)	474 (430.5)a	581.25 (467)b	450 (375)	.419c
PANSS-P	15.0 (11)a	14.0 (16)b	14.0 (14)	.862c
PANSS-N	18.5 (9)a	19.0 (13)b	18.0 (12)	.852c
PANSS-G	35.0 (15)a	32.0 (20)b	32.0 (18)	.662c
PANSS-T	69.0 (28)a	67.0 (48)b	68.0 (43)	.755c
Staging	8.0 (14)	6.0 (7)e	10.0 (11)	.010c

Continuous variables are shown as median (IQR). Note: GP-CX type, globus pallidus hypertrophy and cortical thinning type; PANSS, Positive and Negative Syndrome Scale (P: positive, N: negative, G: general, T: total); pure CX type, cortical thinning type; SC type, subcortical volume reduction type; SuStaIn, Subtype and Stage Inference; TRS, treatment-resistant schizophrenia.

^aMissing in 2 patients.

^bMissing in 1 patient.

^cKruskal-Wallis test.

^dχ² test.

^eSignificantly lower than pure CX type (P = .009, post hoc Dunn test with Bonferroni correction) and a trend toward lower than SC type (P = .076). No significance between pure CX and SC types (P = .766).

Associations With Other Clinical Characteristics

As shown in table 2, there were no significant relationships between the 3 subtypes and other clinical characteristics. The proportion of TRS also did not significantly differ across the 3 subtypes (table 2). Of 23 patients in the baseline stage, 10 subjects (43%) were TRS. In addition, the estimated stages were not correlated with most of other clinical variables except for the PANSS positive and total scores (uncorrected P < .05, table 3). Regarding the use of CLZ, there was no significant difference in staging and subtyping when directly comparing TRS with CLZ and TRS without CLZ (figure 1C and Supplementary File S3). On the other hand, when compared with non-TRS, TRS with CLZ use showed significantly more progressed stage, while TRS without CLZ use only presented a trend-level progression (figure 1C).

Table 3.

Correlation Analysis With the Estimated Stage of Disease Progression

	Spearman’s rs	P
Age	0.010	.900
Sex (M = 1, F = 2)	−0.024	.749
Onset age	0.084	.270
Illness duration	0.009	.905
Education years	−0.056	.455
Chlorpromazine-equivalent dose of antipsychotics	0.116	.124
PANSS-P	0.175	.021
PANSS-N	0.105	.171
PANSS-G	0.129	.091
PANSS-T	0.154	.043

Bold font denotes uncorrected P < .05. Note: PANSS, Positive and Negative Syndrome Scale (P: positive, N: negative, G: general, T: total).

Reproducibility Analysis

The results of the reproducibility analysis are shown in figure 2. Although the subtype patterns were generally consistent with the main analysis, SuStaIn did not identify a pure CX type in cohort 2 (figure 2A). Therefore, 18 of 20 patients in cohort 2 with pure CX type in the main analysis were classified into GP-CX type (figure 2B). Otherwise, patients were categorized into the same subtype groups (Cohen’s κ = 0.609, P < .001). The stage of disease progression was well reproduced (Kendall’s Tau-b = 0.934, P < .001, figure 2C).

Fig. 2.

Reproducibility analysis by analyzing each cohort separately. (A) Subtype and staging results of each cohort. Cohorts 1 and 3 (left and right) reproduced 3 groups of the SC, GP-CX, and pure CX types, respectively. In cohort 2, SuStaIn classified 2 groups of the SC type and GP-CX type (middle). (B) Subtype classification between main original analysis and reproducibility analysis. It shows which group the patients in each of the groups classified in the main original analysis were classified in the reproducibility analysis. (C) Correlation of stage progressions between main original analysis and reproducibility analysis.

Discussion

The current study applied the unsupervised machine-learning model to data of brain morphology in patients with schizophrenia and identified 3 subtypes with the following progression patterns: SC type, in which subcortical volume loss is more dominant; GP-CX type, in which GP increase initially occurs and cortical thinning follows; and, pure CX type, in which cortical thinning mainly occurs. Furthermore, more patients with TRS were at the progressed disease stages compared with those with non-TRS. Additionally, GP-CX type was associated with less progressive stages. These brain morphological subtypes and staging may, in turn, lead to the development of clinically useful individualized biomarkers. Schizophrenia is a syndrome with heterogeneity, and to date, no clinically available biomarkers have been established from the symptom-based approaches. Thus, individual-level classification and staging based solely on the biological features obtained in this study may lead to more biologically accurate classification, which would allow us to clarify pathophysiology and select personalized treatment. We clarified the relationship between stage progression and TRS, suggesting the possibility of individual-level disease monitoring for treatment resistance in clinical practice.

There have been various attempts to identify subtypes and staging in neurodegenerative diseases by SuStaIn. In frontotemporal dementia, 4 distinct subtypes with progression patterns were detected using structural MRI, which were consistent with genetic variance.¹⁹ Another study applied SuStaIn to tau-PET images of Alzheimer’s disease and identified 4 distinct spatiotemporal trajectories of progression.²¹ In the present study, we applied SuStaIn to schizophrenia and found 3 subtypes with distinct features in terms of anatomical patterns. SuStaIn represents an unsupervised algorithm, and the neuroanatomical information of each ROI and other clinical data were not incorporated into the analysis. Nevertheless, the 3 subtypes were anatomically consistent; cortical and subcortical patterns could be separated and the left/right sides generally changed simultaneously (figure 1A). Such anatomical consistency would support the validity of the SuStaIn classification.

In a previous study by machine learning and structural MRI in schizophrenia, 2 distinct patterns were reported: ie, (1) widespread GM volume loss in the frontotemporal and insular cortices, thalamus, and nucleas accumbens, and (2) increased subcortical GM volumes with no distinct cortical GM loss.¹² These findings may partly align with our results in terms of classification into cortical and subcortical patterns. However, given the possible progression of schizophrenia, incorporating disease staging may be desirable. Our results also advance the field incorporating subpopulation data (ie, TRS and non-TRS) that seems linked to different neurobiological dysfunction.^4,40 Among the identified 3 subtypes, the SC type showed initial subcortical GM loss and subsequent cortical thinning, whereas both the GP-CX and the pure CX types presented earlier and with more distinct cortical thinning and less evident subcortical GM decrease than the SC type (figure 1A). The clear difference between the GP-CX and pure CX types was the initial GP volume increase. GP increase in patients with schizophrenia has been consistently reported by multicenter studies at a group-level comparison with HCs.^35,41 Although no direct associations with TRS were found, the GP-CX type showed less disease progression than the pure CX type. On the other hand, in the reproducibility analysis, it was difficult to distinguish between GP-CX and pure CX in a smaller cohort (cohort 2, figure 2); accordingly, more careful interpretation may be needed. At any rate, the pathological meaning of GP increase in schizophrenia is not well understood, and further investigation is needed to interpret our current model.

Furthermore, a latest study also applied SuStaIn to brain morphological data of schizophrenia.⁴² In the study, 2 distinct patterns were identified, ie, (1) beginning with cortical thinning initially in the Broca’s area and insula, and (2) beginning with subcortical atrophy particularly in the hippocampus.⁴² Although there are some differences in the ROI selection, these 2 patterns are very similar to the typologies we found. While our study lacks external validation cohorts, we believe that this study supports the reproducibility of our analysis. On the other hand, the difference is that this study modeled atrophy as the only pathological change and did not take into account GP hypertrophy.⁴² We incorporated GP hypertrophy into the model and found that the cortical thinning type can be further classified into those with and without initial GP hypertrophy (the GP-CX and pure CX types, figure 1). In addition, regarding the treatment resistance, we believe that our study may have provided more rigorous evidence of the pathophysiology of TRS by applying a more rigorous definition of TRS.³³

Several studies have investigated the neurobiological mechanism of TRS, including involvement of glutamate and GABA systems^{25,26,29,43,44} as well as cortical abnormality patterns.^9,27 Our previous cross-sectional studies on TRS have reported elevated glutamate and glutamine levels in the anterior cingulate cortex,²⁵ abnormal subcortical volume and CT,^9,27 and changes in white matter connections and glutamate in the frontostriatal circuit.²⁸ The current study took into account the temporal progression of disease using SuStaIn and identified that stage progression may be associated with TRS, despite the use of cross-sectional data. Thus, the previous neuroanatomical or metabolic findings on TRS may occur with disease progression of schizophrenia. Indeed, our classification study of CT suggested that TRS and non-TRS share similar characteristics and TRS may be a more advanced form of non-TRS.⁹ Additionally, regarding the subcortical volumes, our past study revealed larger GP volumes in non-TRS,²⁷ which would be consistent with the current findings that the GP-CX type showed less progression stages. On the other hand, in the current study, stage progression was associated with TRS but not with disease duration (table 3). Thus, it does not appear that the disease simply progresses as time progresses. To this point, longer duration of untreated psychosis and treatment nonadherence have been identified as associated factors with TRS.¹⁶ The point has also been made that TRS is associated with central oxidative stress and increased variability of glutathione.⁴⁵ Together with our findings, such unfavorable factors may advance the disease stages and pathological changes in TRS. These different influences may, in fact, contribute to the variable outcomes associated with TRS, eg, CLZ response.

Although the significance was low (uncorrected P < .05), we also found the relationship between stage progression and PANSS scores (table 3). More progressed stages were correlated with PANSS positive and total scores, with correlation coefficients of 0.154–0.175. Since PANSS is a symptom-based measure while treatment resistance has a more biological aspect, it may be reasonable that PANSS had a lower correlation with parameters derived solely from biological data by SuStaIn. However, the latest study with larger longitudinal samples suggested an association with treatment-induced changes in PANSS scores,⁴² and thus, we speculate that biomarkers from SuStaIn can be useful in the future for monitoring and predicting prognosis, even on a symptom basis.

The strengths of this study include a novel data-driven approach for individualized subtyping and staging, findings of a significant association between progressed staging and treatment resistance, and the reproducibility of multisite data. On the other hand, this study has several limitations. First, it is difficult to demonstrate the validity of current results in the absence of a well-established gold standard. One possible solution is to longitudinally follow these cohorts to confirm whether established subtypes follow specific trajectories. At least, it should be noted that the present results are based solely on cross-sectional data and theoretical models and will need to be further proven by appropriate methodologies. Selection of ROIs may also raise some controversy although we adopted the most reliable evidence from the international mega-analyses.^35,36 The z score normalization is required for SuStaIn and we adopted it with no additional statistical correction, like other multicenter SuStaIn studies.^19,20,42 However, given the concerns on multisite neuroimaging studies,⁴⁶ an appropriate correction might have been desirable. In addition, we did not find any relationships between subtypes and clinical characteristics; however, this issue is consistent with the previous study.¹² We speculate that this reflects the limitations of diagnoses based solely on symptoms and the assessment of disease status only by scoring. Much more knowledge related to neurobiological mechanisms may serve to address this limitation. Other limitations include sample size, cross-sectional design, and potential effects of medications and previous unreported use of other drugs. The impact of antipsychotic medications is important, especially for the possible progression by the CLZ use.^37–39 Although we found no significant difference in staging between TRS with CLZ and TRS without CLZ (figure 1C), the CLZ use group showed slightly more progressed stages. Given the statistical insignificance, it would not be possible to conclude an association or causal relationship between CLZ and stage progression. There might also be other unknown or unevaluated confounders. Thus, further investigations would be needed to address these limitations.

In conclusion, we identified 3 distinct subtypes based on progression patterns of brain morphology. More progressed disease stages were found in TRS. The GP-CX type reflected an earlier stage of disease, but otherwise the subtypes did not show any relationship with clinical characteristics. Current findings provide new knowledge that may be relevant to the neural basis of TRS and, in so doing, lead to clinically useful personalized biomarkers.

Funding

This study was supported by Japan Society for the Promotion of Science (18H02755, 22H03002), Takeda Science Foundation, Watanabe Foundation, Uehara Memorial Foundation, Inokashira Hospital Research Foundation (S.N.), Ontario Mental Health Foundation (OMHF) Type A grant (A.G.-G.) and by Canadian Institutes of Health Research (CIHR) Grant Nos. MOP-142493 (A.G.-G., P.G., and G.R.) and MOP-141968 (A.G.-G).