Introduction

Psychotic disorders can be characterized along the dimensions of positive, negative, and cognitive symptoms [1, 2]. Positive symptoms describe abnormal perceptions, thoughts, and behaviors (e.g., hallucinations) while negative symptoms describe a significant reduction of typical behaviors and motivation (i.e., asociality) [1]. Cognitive symptoms tend to arise before the onset of positive and negative symptoms [3, 4] and include neurocognitive deficits, particularly in episodic memory [2]. Due to its broad range of symptoms, psychotic disorders are among the most disabling mental disorders and a leading cause of disability worldwide [5]. The most prominent measure of disability caused by psychotic disorders is functional outcome, which assesses the influence of disease on social and occupational functioning [6, 7]. Within these two domains, affected individuals have fewer stable friendships, difficulties with social interactions, higher unemployment rates, and increased occupational stress [7]. Beyond the considerable impact on the affected individual and their social surroundings, psychosis also carries an immense economic burden caused by the often-persisting difficulties of maintaining employment following a first episode of psychosis (FEP) [8, 9]. Thus, identifying disease trajectories up to and including poor functional outcome of psychosis is crucial.

Multiscale neuroscience is a framework within which poor functional outcome in psychosis may be understood through its neurocognitive mechanisms, for example, from neurobiological changes (meso scale) that result in poor functional outcomes (social scale) through cognitive and clinical symptoms (manifest or behavior scale). Within this framework, psychiatric disorders arise as the result of aberrant interactions within and across those scales [10, 11], implying that dysfunction in one scale influences and reflects functionality of higher- and lower-order scales [10]. On the meso scale, graph-theoretical measures of connectivity identify the hippocampus as a key convergence zone for cortical input [12,13,14]. By grouping hippocampal regions into a hippocampal module, Makowski et al. [15] show that longitudinal structural covariance-based connectivity between the hippocampus and large-scale brain networks [16] is significantly reduced in FEP compared to controls. These findings further provide a link between the hippocampus and the manifest- or behavior-scale, as reductions in hippocampal connectivity are associated with heightened negative symptoms, as mediated by impaired episodic memory [15]. Using longitudinal data, Makowski et al. [15] were, thus, the first to characterize a multiscale disease progression from hippocampal-cortical dysconnectivity to episodic memory and negative symptoms. Notably, the association between episodic memory and psychosis is robustly established cross-sectionally and longitudinally across the psychosis spectrum, spanning individuals at clinical high risk [17], FEP [18], and chronic psychosis [2]. Within the manifest-scale, the relationship between episodic memory and negative symptoms is mediated through social cognition (i.e., emotion recognition and theory of mind) [19], linking cognitive impairments to negative symptoms. On the social-scale, sex differences in episodic memory impairments [20], and impaired social cognitive abilities [21, 22] predict functional outcome as mediated by negative symptoms, while negative symptoms are also directly associated with impaired functioning [23, 24]. We thus propose a multiscale model in which hippocampal-cortical dysconnectivity leads to impaired episodic memory and social cognition, resulting in higher negative symptoms and, ultimately, poor functional outcome.

To assess this model, we implemented an unsupervised machine-learning algorithm called Subtype and Stage Inference (SuStaIn) [25]. SuStaIn uniquely combines disease progression modeling and clustering to infer longitudinal disease progressions from cross-sectional data. By clustering patient groups with shared common disease progression patterns, SuStaIn is particularly well-suited to evaluate a multiscale model of psychosis, as heterogeneous impairments are seen across patients [26,27,28,29,30,31], indicating that our hypothesized disease progression pattern may apply to only a sub-population of patients. While SuStaIn has primarily been implemented in neurodegenerative disorders [25, 32,33,34,35,36,37], two studies to date have applied SuStaIn to neuroimaging markers of psychosis [38, 39]. Both identified distinct subtypes of volume atrophy in patients, with one subtype beginning to deteriorate in the hippocampus. Jiang et al. [38] further related these subtypes to distinct treatment responses to antipsychotic medication. Building on this important work, our study will, for the first time, implement SuStaIn in a multiscale psychiatric model.

Further, SuStaIn’s staging approach provides a translational component to our data-driven model by directly allowing for patient stratification and disease prediction. This feature is of particular importance in relation to clinical staging models, such as McGorry’s clinical stages [40, 41]. McGorry et al. [40] hypothesize a unidirectional clinical development of severe mental disorders, in which new stages represent a significant change in clinical status [40, 41]. In this work, we sampled across clinical stages by combining data on first- and multi-episode psychosis (MEP). Based on the qualitative differences between clinical stages regarding symptoms, cognition, and functioning [40, 41], it was of further interest to explore how SuStaIn’s data-driven disease stages would stratify individuals from distinct clinical stages. To this end, we employed z-score SuStaIn [25], which assesses monotonous disease progression through z-score deviations from the mean, resembling the assumed unidirectional progression of McGorry’s clinical stages.

Applying z-score SuStaIn thus allowed us to investigate our novel multiscale model across subtypes of psychosis while addressing whether our data-driven disease stages would map onto clinical staging models [40]. Taking into consideration the heterogeneous nature of psychotic disorders, we hypothesize (i) that there may exist at least two patient subtypes, one subtype showing normal-range performance on the components of the model and one subtype showing impairments. In the impaired subtype, we hypothesize (ii) a disease progression pattern from hippocampal-cortical dysconnectivity to impaired episodic memory and social cognition, higher negative symptoms, and poorer functional outcome. This work is the first to outline a relationship between hippocampal-cortical connectivity and functional outcomes through cognition, social cognition, and symptoms. Given this multiscale nature, our data-driven findings can thus contribute to precision diagnostic and treatment efforts across scales.

Methods

Participants

We sampled patient and control data from two independent datasets. Study 1 collected data in a FEP cohort (recruitment: 2015–2021) and Study 2 from a MEP cohort (recruitment: 2011–2016; [42]); (see Supplementary F1 for pre-established inclusion and exclusion criteria of both studies). Patient samples were recruited from the Douglas Research Centre, Montréal, Canada. FEP patients were recruited after being admitted to the prevention and early intervention program for psychosis (PEPP-Montréal) [43], while MEP patients were patients of the outpatient and inpatient units. Both control samples were recruited from the same catchment area. After excluding participants (see Supplementary F2 for the exclusion process), Study 1 consisted of 55 patients and 53 controls, and Study 2 consisted of 108 patients and 66 controls, rendering our total sample size 163 patients and 119 non-clinical controls. Ethical approval was granted by the Douglas Institute Research Ethics Board. Both studies were conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained prior to the study. Participants were compensated with monetary rewards. See Supplementary S1 for a power calculation of the sample size.

Data collection

In both studies, data including sex, age, duration of illness in years (first onset of frank psychotic symptoms), and medication were assessed through an interview with trained research staff. Diagnoses were established through the Structured Clinical Interview for DSM-IV and V diagnoses [44, 45]. Episodic memory and social cognition were measured via the Cogstate Schizophrenia Battery [46], namely, the International Shopping List and Social-Emotional Cognition Tests. Depressive symptoms were assessed through the Calgary Depression Scale [47]. Negative and positive symptoms were assessed using the Scale for the Assessment of Negative/Positive Symptoms (SANS/SAPS) [48, 49], respectively. Functional outcome was measured through the Social and Occupational Functional Assessment Scale (SOFAS, [6]), which assesses social and occupational functioning on a scale from 0 (grossly impaired functioning) to 100 (excellent functioning). The SOFAS was not measured in MEP patients, and we thus reconstructed SOFAS scores on the basis of the Personal and Social Performance Scale (PSP, [50]). The PSP assesses occupational functioning, interpersonal functioning, self-care, and aggressive behavior and is scored on a scale from 1 (no autonomy) to 100 (excellent functioning in all four domains). To score the PSP, we consulted the SANS for occupational functioning (item 15), interpersonal functioning (items 20 and 21), self-care (item 14), and the SAPS for aggressive behavior (item 23). PSP scores were then transferred onto the SOFAS. The scoring of both scales was completed by two independent raters, and a consensus was reached. Validity of this procedure was justified through the scoring of an independent dataset with pre-existing SOFAS ratings. See Supplementary S2 for scale and scoring descriptions, and Supplementary S3 and F6 for details on the reconstruction procedure.

Hippocampal-cortical connectivity

All MRI data were acquired with the same 3T Siemens Magnetom Trio scanner located at the Cerebral Imaging Centre of the Douglas Research Centre. See Supplementary S4 for the scan parameters of both studies. We used the high-performance computing resources available via the Digital Research Alliance of Canada to extract cortical thickness using CIVET [51], and hippocampal volumes with MAGeT [52, 53]. See Fig. 1 and Supplementary S5 for a detailed explanation of the processing steps. We then computed structural covariance between the 62 DKT and 18 hippocampal regions by performing full-sample correlations for the patient and the control samples separately [13] and calculated subject-specific structural covariance matrices through the jackknife bias estimation procedure [54, 55]. Hippocampal-cortical connectivity for each hippocampal region was derived through the graph-theoretical participation coefficient [56], with higher values indicating more diverse nodal connectivity to large-scale cortical networks [16] rather than hippocampal regions. See Supplementary S5 for a detailed explanation of these calculations. The participation coefficients of the hippocampal regions were averaged for each hemisphere respectively.

Fig. 1: Methods workflow.
figure 1

This figure displays the workflow of data processing, calculation of hippocampal-cortical connectivity, and the application of SuStaIn.

Full size image

SuStaIn analyses

We first regressed out the effects of sex, age, and total brain volume on the six input markers and then z-scored the data in comparison to control values. As no control values were sampled for the SANS and SOFAS, we z-scored these values in relation to previously reported control means [57, 58] and applied z-score SuStaIn [25] to the patient data only. Z-score SuStaIn places subjects at specific disease stages depending on their z-score deviation from a non-clinical control sample. Through maximum likelihood estimation, SuStaIn assesses the temporal sequence of marker deviations and employs purely cross-sectional data to infer common patterns of disease progression via a stage progression model. Under the assumption of monotonous marker deterioration, z-score SuStaIn thus places markers with the highest likelihood of impairment into early disease stages, while markers with less likelihood are placed into later disease stages. This type of modeling implies that patients in early disease stages share impairments on a few markers, whereas the number of markers that show impairments grows over the course of the disease progression. SuStaIn’s disease stages thus provide an overview of shared patterns of marker impairment at a given stage while also allowing for the inference of disease progression following this stage. In addition, distinct subtypes with common impairments are identified through a clustering component, the split-cluster expectation maximization, with likelihood testing determining the optimal number of clusters.

Z-score SuStaIn was performed in Python (pySuStaIn [59]). We set the threshold of z-scores to one, two, and three deviations from the control mean for each biomarker while reaching a maximum threshold of five z-scores at the end of the disease progression, as in previous work [25]. Considering the absence of negative symptom and functional outcome data in controls, we ran SuStaIn on the patient data only. We then validated the resulting subtypes using 10-fold Markov Chain Monte Carlo iteration cross validation. Log-likelihood tests and the Cross-Validation Information Criterion (CVIC, [25]) were used to obtain an indication of the optimal number of subtypes. Following SuStaIn, follow-up analyses were performed and corrected for multiple comparisons. See Supplementary S6 for a detailed explanation of these analyses.

Results

Demographic data

A summary of demographic characteristics can be seen in Table 1. We further examined the role of diagnosis and depressive symptoms in our sample. Study 2 had significantly more individuals with non-affective psychosis than Study 1 (X2(1, N = 163) = 9.865; p = 0.002), but none of the input markers of our model differed significantly between diagnoses (ps > 0.582). When looking at depressive symptoms, no significant difference between FEP and MEP was found (p > 0.244). We did, however, find a significant correlation between negative symptoms and depressive symptoms in FEP (r(52) = 0.372, p = 0.006), but not MEP. See Supplementary F3 for further visualization of age and illness duration in both datasets.

Table 1 Demographic characteristics.
Full size table

Subtype and stage inference

SuStaIn considers individuals who are sorted into disease stage 0 as normal-range performing on all markers and assigns them to a separate Subtype 0. In our case, Subtype 0 thus consists of patients with no impairment on the markers. To differentiate Subtype 0 from other subtypes with impairment, we therefore refer to all subtypes except Subtype 0 as disease subtypes. We ran SuStaIn on a maximum of two disease subtypes, which resulted in a CVIC of 3329.44 (log-likelihood of—168.32) for two disease subtypes and a CVIC of 3326.73 (log-likelihood of—168.08) for one disease subtype. Lower CVIC scores suggest a better model fit [25], yet differences <6 are not considered to be meaningful [60]. As our results, therefore, suggest a suitable fit of both models, we report the results for a model with two disease subtypes below (see Supplementary S7 for the model with one disease subtype). In this model, 58 participants (25 FEP; 33 MEP) were sorted into Subtype 0, 83 participants (24 FEP; 59 MEP) were sorted into Subtype 1, and 22 participants were sorted into Subtype 2 (6 FEP; 16 MEP).

The Kruskal–Wallis test showed a significant main effect of Subtype on left (H(2) = 52.203, p < .001, η2H = 0.31) and right hippocampal-cortical connectivity (H(2) = 35.592, p < .001, η2H = 0.21). Post-hoc Mann–Whitney U tests revealed that Subtype 2 exhibited significantly lower bilateral hippocampal-cortical connectivity than Subtype 0 and Subtype 1, while these two did not differ significantly. Thus, only patients in Subtype 2 were characterized by impaired hippocampal-cortical connectivity when collapsing across stages. The significant main effect of Subtype on episodic memory (H(2) = 87.867, p < 0.001, η2H = 0.54) and follow-up analyses showed that Subtype 1 showed significantly lower episodic memory than Subtype 0 and Subtype 2, while Subtype 2 also showed lower episodic memory than Subtype 0. A different pattern was observed for social cognition (H(2) = 21.816, p < 0.001, η2H = 0.12) and negative symptoms (H(2) = 26.319, p < 0.001, η2H = 0.15), where follow-up analyses revealed significantly poorer social cognition and higher negative symptoms in Subtype 1 in comparison to Subtype 0 and Subtype 2, while Subtype 0 and Subtype 2 did not differ significantly. Regarding functional outcome (H(2) = 10.860, p = 0.004, η2H = 0.06), Subtype 1 was significantly more impaired than Subtype 0, yet did not differ from Subtype 2, indicating that Subtype 1 scored lowest on functional outcome. See Fig. 2a for a visualization of the follow-up Mann–Whitney U tests and Table 2 for demographic characteristics and statistical comparisons of the subtypes. With regard to subtype differences on diagnosis and depressive symptoms, no significant effects were observed after correcting for multiple comparisons (ps > 0.233).

Fig. 2: Subtype differences and disease progression.
figure 2

a Shows the mean comparisons between subtypes on the six markers of our model. **p < 0.001, *p < 0.05. b, c Shows the disease progression of the markers in both disease subtypes, respectively. The color coding indicates a deviation from normality by one z-score in blue, by two z-scores in magenta, and by three z-scores or more (up to five) in red. The stronger the color, the higher the probability of individuals at that stage to deviate by one z-score, two z-scores, three z-scores. Subtype 1 progresses from episodic memory to social cognition, negative symptoms, functional outcome, left hippocampal-cortical connectivity, and right hippocampal-cortical connectivity. Subtype 2 progresses from left hippocampal-cortical connectivity to right hippocampal-cortical connectivity, episodic memory and social cognition, functional outcome, and negative symptoms.

Full size image
Table 2 Demographic characteristics of subtypes.
Full size table

As Subtype 0 per definition implies no disease progression, SuStaIn exclusively infers disease progressions for the disease subtypes. For Subtypes 1 and 2, data are provided up to disease stages 9 and 7, respectively, with SuStaIn’s longitudinal inference inferring biomarker progression until disease stage 18. As seen in Fig. 2b, c, disease progression is shown in z-score deviations from the control mean, with blue squares indicating one z-score deviation, magenta indicating two z-score deviations, and red indicating three or more (up to five) z-score deviations. In case no change occurs at a specific stage, the square is left white. The chosen model only allows for up to five z-score deviations from the mean, meaning that i.e., episodic memory and social cognition reach the highest possible deviation at disease stage 6. Higher z-score deviations might still occur in later disease stages yet will not be modeled due to the chosen cut-off score of five in line with prior implementations of SuStaIn [25].

Based on the z-score means per stage (Fig. 3), Subtype 1 shows an early deterioration of episodic memory, which deviates one z-score from the mean at disease stage 1, followed by social cognition, reaching one z-score deviation from the mean at disease stage 4. Negative symptoms and functional outcome then both reach one z-score deviation from the mean at disease stage 8. Left hippocampal-cortical connectivity reaches one z-score deviation from the mean at disease stage 9, while right hippocampal-cortical connectivity never substantially deviates from the mean. The longitudinal inference of SuStaIn shows that it most likely starts deteriorating from disease stage 12 onwards, which is beyond the stages represented in our sample. Subtype 2 shows a deterioration of left hippocampal-cortical connectivity by two z-score deviations from the mean at disease stage 1, while right hippocampal-cortical connectivity follows with one z-score deviation from the mean at disease stage 2. Social cognition reaches one z-score deviation from the mean at disease stage 3, while episodic memory reaches two z-score deviations from the mean at disease stage 4. Given that social cognition and episodic memory, therefore, start deteriorating at around the same stage, yet episodic memory continues to deteriorate while social cognition scores improve again (Fig. 3), we mention episodic memory before social cognition in our disease progression. Negative symptoms and functional outcome do not deviate substantially from the mean in our sample, yet the longitudinal inference of SuStaIn (Fig. 2c) shows that functional outcome deviates slightly earlier. We, therefore, identify a disease progression from episodic memory to social cognition, negative symptoms, functional outcome, left and right hippocampal-cortical connectivity in Subtype 1, and a disease progression from left and right hippocampal-cortical connectivity to episodic memory, social cognition, functional outcome and negative symptoms in Subtype 2 (Fig. 2).

Fig. 3: Subtype and stage inference.
figure 3

This figure shows how patients are first subtyped into Subtype 0 (white), disease Subtype 1 (yellow), or disease Subtype 2 (blue) and how the disease progression of the respective disease subtypes is then inferred by placing the patients onto a disease stage of this subtype. The numbers on top of the bars reflect the raw number of (FEP/MEP) patients in each stage. The table on the right indicates the percentage of affective and non-affective FEP and MEP patients in each stage.

Full size image

First- and multi-episode psychosis

FEP and MEP patients are distributed equally across stages (Fig. 3). To identify further similarities, we split both samples into low versus high negative symptoms based on the respective mean value for negative symptoms. Results revealed a significant main effect of negative symptom group (FEPlow (n = 28), FEPhigh (n = 27), MEPlow (n = 53), MEPhigh (n = 55)) on negative symptoms (H(3) = 121.174, p < 0.001, η2H = 0.74) and functional outcome (H(3) = 20.119, p < 0.001, η2H = 0.11). Follow-up results for those variables are visualized in Fig. 4a. The findings of functional outcome suggest that MEP with high negative symptoms are particularly impaired, yet FEP and MEP overall score comparably on most variables, irrespective of negative symptoms. Figure 4b, c further visualize that individuals with low negative symptoms tend to be grouped into earlier disease stages while individuals with high negative symptoms are distributed across all stages and dominate the later disease stages.

Fig. 4: FEP and MEP along low and high negative symptoms.
figure 4

This figure shows the comparison between FEP and MEP who score low and high on negative symptoms on the markers of our model, **p  < 0.001, *p < 0.05, (a) and their distribution across the stages of our disease subtype 1 (b) and 2 (c).

Full size image

Difference to one-subtype model

In a model with one disease subtype, Subtype 1 is characterized by significantly lower episodic memory, social cognition, higher negative symptoms, and poorer functional outcome than Subtype 0, and progresses from episodic memory to social cognition, negative symptoms, functional outcome, left and right hippocampal-cortical connectivity (see Supplementary S7 for the subtype description, T1 for subtype demographics, F4 for a circle diagram showing the patient distribution in a model with one vs. two disease subtypes, and F5 for figures of SuStaIn). While both models show a statistically similar fit for our data, the two-subtypes model seems to uncover the importance of hippocampal-cortical connectivity in leading the disease progression in part of the patient sample.

Discussion

We proposed a multiscale model of psychosis to advance our understanding of functional outcomes of psychosis and explored heterogeneous disease trajectories across patient subtypes through machine-learning. This first implementation of SuStaIn in a multiscale model of psychiatry allowed us to identify three distinct patient subtypes with heterogeneous disease progression patterns. Subtype 0 included patients without impaired connectivity, (social) cognition, symptoms, or functioning. Subtype 1 and Subtype 2 mostly supported our proposed progression model in terms of episodic memory, social cognition, negative symptoms, and functional outcome. Notably, hippocampal-cortical dysconnectivity preceded these markers in Subtype 2, while Subtype 1 showed that hippocampal-cortical dysconnectivity emerged following cognitive and functional impairments and higher symptoms. This suggests that hippocampal-cortical dysconnectivity can but does not necessarily precede impairments at the manifest- and social-scale. These results demonstrate how multiscale frameworks can offer important insights into the complex temporal nature of cross-scale interactions while underlining the need to differentiate between patient subtypes in disorders as heterogeneous as psychosis.

We identified two distinct disease trajectories between our subtypes, with Subtype 1 progressing from episodic memory to social cognition, negative symptoms, functional outcome, and hippocampal centrality, while hippocampal centrality led the disease progression in Subtype 2. These disease progression patterns do not seem to be driven by differences in demographic (i.e., age) or clinical (i.e., diagnoses, depressive symptoms) variables, pointing towards distinct methodological and neurobiological mechanisms underlying these trajectories. Notably, we performed cluster analyses, which could explain why the impairments in hippocampal centrality, memory, and symptoms, as identified by Makowski et al. [15], were spread across our two distinct subtypes. More precisely, we identified Subtype 1 with impaired episodic memory and high negative symptoms and Subtype 2 with impaired hippocampal centrality and episodic memory. While the findings by Makowski et al. [15] might thus indicate a pooling of impairments across the sample, their use of longitudinal data may also point toward impairments at later time points, not captured in the cross-sectional implementation of SuStaIn. This would be in line with the longitudinal inference feature of SuStaIn for Subtype 2, showing an increase in negative symptoms at disease stage 9. These methodological aspects are of particular importance when considering the partial overlap of Makowski’s sample and the sample of Study 1.

That hippocampal dysconnectivity does not lead the disease progression in Subtype 1 might further be explained by other neural markers. Makowski et al. [15] used a quantitative MRI, which is an index of intracortical myelination. Considering other MRI modalities, such as quantitative, diffusion-weighted, or functional MRI, might thus provide a more comprehensive overview of the hippocampal dysfunction in our neurocognitive model and might underlie the disease progression in Subtype 1. In addition, neural correlates of social cognition might drive this progression, such as the amygdala [61], for which dorsomedial shape development has been shown to predict conversion to psychosis [62] while further being associated with poorer emotion recognition [62]. Beyond the association with social cognition, the amygdala and hippocampus are distinctly associated with negative symptoms [63]. Future work aiming to disentangle multiscale contributions to impaired social cognition and episodic memory may wish to consider distinct MRI modalities while also looking at the independent contributions of the amygdala and hippocampus.

We further identified an almost identical temporal development of negative symptoms and functional outcome in both disease subtypes. One explanation might lie in the pooling of SOFAS social and occupational functioning scores [6], which are used to estimate overall functional outcome. With regard to the social functioning component, premorbid social functioning, in particular, has been shown to be predictive of negative symptom severity in later disease stages [64, 65]. While premorbid social functioning and social outcomes of psychosis are two distinct constructs, considering premorbid functioning as a potential confounder may help to disentangle the relationship between symptoms and functioning. We would further like to point out that our key objective was to propose a multiscale perspective on functional outcomes in psychosis, providing a link from the brain to functional outcomes through socio-cognitive deficits and symptoms. We did not intend to provide an exhaustive account of factors influencing functional outcomes of psychosis and recommend that future research considers these aspects further.

Our findings further need to be discussed in the light of clinical staging models. McGorry et al. [40] hypothesize a unidirectional disease progression in which new clinical stages represent a significant change in clinical status [40, 41]. By including patients from two distinct clinical stages and applying a machine-learning algorithm which models unidirectional disease progression, we were in the unique position to address whether McGorry’s clinical stages could be mapped onto the disease stages identified by SuStaIn. As McGorry suggests, MEP patients showed lower episodic memory and functioning and higher negative symptoms, yet FEP and MEP participants were spread equally across SuStaIn subtypes and, most importantly, disease stages. While this seems contradictory to McGorry’s clinical stages at first sight, it is important to differentiate between clinical stages and disease states. Clinical stages carry qualitative value about patient history and inform treatment decisions. Disease states, however, are the patient’s scores on symptom severity, functioning, and cognition at a specific point in time. SuStaIn, therefore, infers longitudinal disease stages from patient states and places individuals with good state performance (i.e., low negative symptoms, high functioning at assessment) in early disease stages and individuals with poor state performance (high symptoms, low functioning) in later disease stages. The assumption underlying this modeling is the one of unidirectionality, implying that individuals who currently have good state performance will later develop poorer state performance, with the degree of impairment and the order of deterioration differing between disease subtypes. By showing that MEP and FEP are placed in the same SuStaIn stages, we show that states across MEP and FEP are comparable, indicating that states can vary within clinical stages. In this sense, our results can be considered a natural extension of McGorry’s clinical stages, contributing to a refining of stage-based frameworks by allowing for a fusion between clinical staging models and biomarker-informed disease trajectories and potentially explaining clinical variation seen within clinical stages.

An additional point to consider is the cyclical nature of psychotic symptomatology. Z-score SuStaIn is well-suited to mimic the unidirectionality of clinical staging models; however, progression patterns might deviate across the scales of our model. When looking at markers such as negative symptoms, symptom severity might not increase but fluctuate over the course of the disease progression [66]. In line with this hypothesis, low and high negative symptoms are associated with distinct degrees of functional impairment, suggesting that the markers might fluctuate depending on symptomatic vs. remitted illness. SuStaIn thus likely sorted remitted MEP patients into early disease stages, as these patient scores had similar mean scores to patients in early clinical stages (FEP). To better capture intra-individual fluctuations within clinical stages, future work should evaluate other disease progression models which do not assume unidirectional disease progression but rather model bidirectional marker fluctuations. Such an approach might allow for an even more reliable modeling of McGorry’s clinical stages while acknowledging the variety of disease states within each clinical stage.

Our findings have important implications for multiscale perspectives on psychosis. To our knowledge, our work addresses a multiscale model of psychiatry using SuStaIn for the first time. Even within multiscale frameworks, the components of clustering in combination with disease progression modeling are essential to capture the complex nature of psychosis. By implementing a mechanistic approach towards disease progression that integrates multiple levels of measurement, we provide a data-driven approach which could complement traditional symptom-based diagnostics [1]. As z-score SuStaIn assigns individuals to distinct disease subtypes and stages, our results also provide clinical utility and translational opportunities for patient stratification and diagnostics based on machine-learning. A good example of clinical utility beyond the implications for McGorry’s clinical staging model [40] is the application of SuStaIn in psychosis by Jiang et al. [38], who found that one of their disease subtypes was related to higher antipsychotic medication efficacy. While we did not find significant differences in medication, future studies in larger samples are warranted. Our findings might further allow for transdiagnostic applications predicting disease progression, considering that hippocampal volume and episodic memory are also highly valuable in major depressive disorder [67, 68]. Lastly, it should be noted that negative and depressive symptoms correlated significantly in FEP, underlining the importance of mood symptoms in early clinical stage. While this might point to potential confounding of negative symptoms and depressive symptoms in FEP, there is also a rich literature on the presence of depressive symptoms early on in psychosis, which diminish following the initiation of treatment [69]. With regard to the role of persistent negative symptoms (PNS) and secondary persistent symptoms (including mood symptoms) in predicting clinical and functional outcomes [70], we encourage future research to investigate the potential of SuStaIn in identifying distinct trajectories of PNS.

Our study has several limitations that need to be considered. We used a novel procedure to reconstruct the SOFAS. While we have statistical support for this method, we nevertheless suggest the replication of this methodology in a larger sample. Further, the SOFAS [6] and SANS [48] were not sampled for the control groups of both studies, and the SOFAS and the SANS patient data had to be z-scored relative to previously reported control data [57, 58]. Our SuStaIn analysis thus had to include patient data only, and our reference for follow-up analyses (Subtype 0) also consisted of patients. Through the future inclusion of a control group, which would most likely be almost entirely subtyped as Subtype 0 [25], it would be possible to compare the deterioration of markers to a more reliable Subtype 0. Additionally, we need to acknowledge the relatively small sample of our study, which consists of 163 patients. According to SuStaIn power estimates, our results are, therefore, slightly underpowered, and we suggest a replication of our findings in a larger sample.

To conclude, as the first investigation of multiscale disease progression in psychosis using SuStaIn, we identified two data-driven disease subtypes with distinct longitudinal disease trajectories. These findings are of particular importance for elucidating the complex and heterogeneous nature of psychosis even within multiscale models and further elucidate the role of hippocampal connectivity and episodic memory in predicting poor functional outcomes. In addition, these are the first data-driven findings to complement traditional clinical staging models by inferring longitudinal disease trajectories across clinical stages. Further research should target broader measures of hippocampal-cortical connectivity through combining distinct neuroimaging modalities and should consider implementing machine-learning models which allow for bidirectional modeling.