ARTICLE OPEN Combining MRI and clinical data to detect high relapse risk after the first episode of psychosis Aleix Solanes 1,2,3, Gisela Mezquida1,4,5,6, Joost Janssen5,7, Silvia Amoretti 1,4,5,8, Antonio Lobo 5,9,10, Ana González-Pinto5,11,12,13, Celso Arango 5,7, Eduard Vieta1,5,8,14, Josefina Castro-Fornieles 1,5,8,15, Daniel Bergé3,5,16, Auria Albacete2, Eloi Giné 17, Mara Parellada5,7, Miguel Bernardo 1,4,5,8, PEPs group (collaborators)*, Edith Pomarol-Clotet2,5,36 ✉ and Joaquim Radua1,2,5,18,19,36 ✉ Detecting patients at high relapse risk after the first episode of psychosis (HRR-FEP) could help the clinician adjust the preventive treatment. To develop a tool to detect patients at HRR using their baseline clinical and structural MRI, we followed 227 patients with FEP for 18–24 months and applied MRIPredict. We previously optimized the MRI-based machine-learning parameters (combining unmodulated and modulated gray and white matter and using voxel-based ensemble) in two independent datasets. Patients estimated to be at HRR-FEP showed a substantially increased risk of relapse (hazard ratio = 4.58, P < 0.05). Accuracy was poorer when we only used clinical or MRI data. We thus show the potential of combining clinical and MRI data to detect which individuals are more likely to relapse, who may benefit from increased frequency of visits, and which are unlikely, who may be currently receiving unnecessary prophylactic treatments. We also provide an updated version of the MRIPredict software. Schizophrenia (2022)8:100 ; https://doi.org/10.1038/s41537-022-00309-w 1234567890():,; INTRODUCTION HRR is crucial since relapse puts their psychosocial recovery at risk, The discovery of associations between magnetic resonance raises the chance of treatment resistance, and has been linked to imaging (MRI) measures and mental disorders1 led to an initial higher direct and indirect social and economic costs10. A few enthusiasm about finding MRI-based biomarkers, but we have studies have created models to estimate this risk based on clinical failed so far. However, new machine-learning methods have data11,12, using variables such as the presence of manic and reopened the possibility of creating MRI-based tools that, while far negative symptoms13–15, the diagnosis12,15, or cannabis use11,16,17. from perfect biomarkers, could still help the clinicians2. These Fewer studies have created models to estimate the risk of tools could help the clinicians diagnose, predict the response to outcomes other than relapse (e.g., the severity of future treatment, or estimate the risk of a bad outcome, adjusting the symptoms) based on brain MRI data18,19, using volumetric brain overall intervention accordingly. changes during the first year20 or voxel/surface-based data18. And Up to the moment, most MRI-based machine-learning studies to our knowledge, no studies have attempted to create MRI-based have aimed to classify the individuals (e.g., patient vs. control, or relapse risk-estimation models. between two diagnoses), and some other research has been This lack of research is unfortunate, given that a structural MRI- devoted to creating models that estimate the risk of a bad based tool able to detect FEP-HRR would be clinically valuable and outcome. For instance, many studies have investigated whether it feasible. It would be valuable because even if the accuracy of the is possible to use clinical data3, MRI data4, or their combination5 to HRR-FEP detection was modest, it could help the clinician adjust the detect healthy individuals at high risk for psychosis. These studies follow-up and treatment of the patients as deemed beneficial21. It have reported higher transition rates to psychosis in individuals would be feasible since individuals with a FEP may undergo an MRI that are males, have brief limited intermittent psychotic symp- to discard organic brain pathology, so that the structural MRI toms, or show reduced cortical gray matter6,7. Conversely, very little research has focused on detecting those required for this tool would serve both. This better clinical patients with first episode of psychosis (FEP) at high relapse risk management would reduce the number of relapse-related hospita- (HRR). This lack of research is striking because FEP represents one lizations in patients at HRR-FEP and exclude patients at low relapse of the main challenges for mental health8. Without an appropriate risk from therapies unnecessary for them. Therefore, it would differential diagnosis and early intervention, clinical development improve the quality of life of individuals with a FEP and reduce the after FEP can lead to a chronic condition9. Detecting subjects at burden on National Health System expenditure. 1 Institut d’Investigacions Biomèdiques August Pi i Sunyer (IDIBAPS), Barcelona, Spain. 2FIDMAG Germanes Hospitalàries Research Foundation, Barcelona, Spain. 3Department of Psychiatry and Forensic Medicine, Universitat Autònoma de Barcelona, Barcelona, Spain. 4Barcelona Clínic Schizophrenia Unit, Neuroscience Institute. Hospital Clínic de Barcelona, Barcelona, Spain. 5 Centro de Investigación Biomédica en Red en Salud Mental (CIBERSAM), Instituto de Salud Carlos III, Madrid, Spain. 6Department of Clinical Foundations, Pharmacology Unit, University of Barcelona, Barcelona, Spain. 7Department of Child and Adolescent Psychiatry, Institute of Psychiatry Mental Health, Hospital General Universitario Gregorio Marañón, School of Medicine, Universidad Complutense, IiSGM, Madrid, Spain. 8Department of Medicine, University of Barcelona, Barcelona, Spain. 9Department of Medicine and Psychiatry, Universidad de Zaragoza, Zaragoza, Spain. 10Instituto de Investigación Sanitaria Aragón (IIS Aragón), Zaragoza, Spain. 11Universidad del País Vasco / EHU, Leioa, Bizkaia, Spain. 12Instituto de Investigación Sanitaria Bioaraba, Vitoria-Gasteiz, Alava, Spain. 13Psychiatric Department, Hospital Universitario de Alava, Vitoria-Gasteiz, Alava, Spain. 14Barcelona Clínic Bipolar and Depressive Disorders Unit, Institute of Neurosciences, Hospital Clínic de Barcelona, Barcelona, Spain. 15Department of Child and Adolescent Psychiatry and Psychology, 2017SGR881. Institute of Neuroscience, Hospital Clínic, Barcelona, Spain. 16Hospital del Mar Medical Research Institute, Barcelona, Spain. 17Hospital de Mataró, Barcelona, Spain. 18Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, London, United Kingdom. 19Centre for Psychiatric Research and Education, Department of Clinical Neuroscience, Karolinska Institutet, Stockholm, Sweden. 36These authors jointly supervised this work: Edith Pomarol-Clotet, Joaquim Radua. *A list of authors and their affiliations appears at the end of the paper. ✉email: epomarol-clotet@fidmag.org; radua@recerca.clinic.cat Published in partnership with the Schizophrenia International Research Society A. Solanes et al. 2 Predicon of age Baseline data acquision MRI pre-processing Test of MRI-based ML Predicon models parameters •Structural MRI •Visual inspecon •Addion of features •Remove effects sex •N=120 healthy individuals •Gray & white maer •Varying smoothing kernels •Gaussian LASSO regression segmentaon •Varying subsampling •Outcome: Age •MNI normalizaon •Ensemble methods •Independent variables: MRI voxel values Predicon of schizophrenia Baseline data acquision MRI pre-processing Test of MRI-based ML Predicon models parameters •Structural MRI •Visual inspecon •Addion of features •Remove effects sex and age •N=128 paents with •Gray & white maer •Varying smoothing kernels •Logisc LASSO regression schizophrenia segmentaon •Varying subsampling •Outcome: Schizophrenia •N=127 healthy individuals •MNI normalizaon •Ensemble methods diagnosis •Independent variables: MRI voxel values Detecon of HRR-FEP Baseline data acquision MRI pre-processing Follow-up Predicon models •Structural MRI •Visual inspecon •Followed unl relapse or for 18- •Remove effects of site, sex and •Clinical variables (PANSS, CGI-S, •Gray & white maer 24 months age etc.) segmentaon •Cox LASSO regression followed by •N=227 paents with FEP •MNI normalizaon group assignaon (HR>1 vs HR<1) •7 different sites •Addion of features •Outcome: Risk group (HRR-FEP 1234567890():,; •Subsampling and smoothing vs. low relapse risk) •Voxel-level ensemble •Independent variables: MRI voxel values + Clinical data Fig. 1 Main study steps. Overall steps followed in this study. The current study investigated whether structural MRI might help detect patients at HRR-FEP. To this end, we created an HRR- Table 1. Description of the cohort (n = 227). FEP detection tool. Additionally, we report how we previously optimized the MRI-based machine-learning parameters, using two Age 24.2 (7.4) independent datasets to avoid data leakage or over-complexity Sex: female 78 (34.4%) (see clarification later). We also freely provide the updated MRI- Familiar psychiatric history 110 (58.2%) based machine-learning software to allow other groups to Affective* 54 (28.6%) develop their own detection models and a website (see “Available Suicide* 3 (1.4%) resources”) that estimates HRR-FEP quickly to help other groups Affective psychosis 48 (21.1%) independently replicate our model’s accuracy assessment. Baseline diagnosis Schizophrenia 78 (34.4%) METHODS Bipolar disorder 41 (18.1%) See Fig. 1 for a view of the overall steps of the study. This study Schizoaffective disorder 12 (5.3%) complies with the Transparent Reporting of a multivariable Substance-induced psychosis 10 (4.4%) prediction model for Individual Prognosis or Diagnosis (TRIPOD, see checklist in the Supplement). Major depressive disorder 5 (2.2%) Other** 81 (35.7%) Participants Positive and Negative Syndrome Scale (PANSS) 71.2 (SD 24.4) The cohort included 227 adolescents/adults with a FEP from 7 Positive scale 17.6 (SD 7.9) different hospitals in Spain, including a previous multicenter Negative scale 18.2 (SD 8.5) study22,23, prospectively followed for two years. We invited all General psychopathology scale 35.9 (SD 12.7) patients who met the inclusion criteria during the recruitment Global Assessment of Functioning scale (GAF) 50.5 (SD 19.7) periods to join the study. We estimated the sample size based on a Clinical Global Impression scale (CGI-S) 4.4 (SD 1.1) previous meta-analysis24, in which the relapse rate at two years was around 37%. With this estimation, the overall sample size to detect a Young Mania Rating Scale (YMRS) 7.9 (SD 10.2) hazard ratio (HR) = 2 between patients at HRR-FEP and patients at Montgomery Asberg Depression rating scale (MADRS) 0.2 (SD 9.9) low relapse risk had to be 190 according to R package powerSurvEpi Long-acting injectable antipsychotic 15 (6.6%) (https://CRAN.R-project.org/package=powerSurvEpi). We included * Familiar affective history included diagnoses such as bipolar disorder or 20% more to compensate for potential early drop-outs. The mean MDD. Familiar suicide history included consummated attempts. age was 24.2 years (SD 7.4), and there were 78 females (34.4%) (Table ** Other baseline diagnoses included brief psychotic disorder, schizophreni- 1). The sample included both young adolescents (12–14 years, n = 6) form disorder, delusional disorder, and psychotic disorder not otherwise and late adolescents/adults (15–59, n = 221); as detailed later, to specified. ensure that the estimation of the model accuracy is not confounded Data are presented as mean (SD) or number (%). by mixing young adolescents with old adolescents/adults, we Schizophrenia (2022) 100 Published in partnership with the Schizophrenia International Research Society A. Solanes et al. 3 repeated the validation of the model after excluding young schizophrenia diagnosis or not). See the Supplement for details. adolescents. The creation of machine-learning models was analog to the one We defined relapses as exacerbations of symptoms during at described later. least one week with at least one of eight PANSS items (P1, P2, P3, We defined the default settings as unmodulated gray matter N1, N4, N6, G5, and G9) scoring above 3 (mild)25. On the contrary, images, smoothed with a kernel of σ = 4 mm (corresponding to remission was defined as scoring <3 in all eight PANSS items. We FWHM = 9.5 mm) and a voxel size of 3 × 3 × 3 mm3. We tested only considered relapse after at least 6 months of remission. whether the accuracy of MRI-based machine-learning models We detail the inclusion/exclusion criteria and a more detailed depended on: the addition of features (gray and white matter description of the cohort in the Supplement. The ethical images, modulated and unmodulated images as they convey committees of all hospitals had approved the study, conducted complementary volumetric information30, global gray matter according to the Declaration of Helsinki. Furthermore, all volume and global brain volume, and the midline abnormalities participants and parents/legal guardians for adolescents under cavum septum pellucidum and absence of adhesion interthala- 16 had given written informed consent. mica, previously reported as good predictors in FEP31,32), the size of the smoothing kernels (from σ = 2 to 6 mm, corresponding to Collection and processing of baseline structural MRI data FWHM ≈ 5.3–15.8 mm, i.e., encompassing the usual widths of standard neuroimaging software) since the previous literature We acquired a high-resolution structural image from each differs in the optimal kernel size18,26, or the use of ensemble participant with a T1-weighted gradient-echo sequence with methods. Ensemble learning methods seek better prediction different devices (see Supplement for details). We used a voxel- performance and robustness by combining the predictions of based morphometry (VBM) pre-processing pipeline because we different models. We used two ensemble methods: (a) we have previously found higher accuracy using VBM data26 (see resampled the subjects with replacement 18 times and repeated Supplement for details). the creation of the risk-estimation model with each of the 18 resampled datasets, and (b) we selected half of the brain 18 times Removal of the effects of the site (i.e., dividing the brain in different angles) and repeated the The effects of the site (e.g., differences in MRI data due to using creation of the risk-estimation model with each of the 18 half different devices) might increase noise and confound the analyses. brains. Any of these two ensemble methods resulted in 18 risk- To remove them, we used a recently developed method to control estimation models, which we applied to the test set, resulting in 18 for batch effects named ComBat, as several studies have shown its risk estimations per patient. Finally, we calculated the mean of the superiority to simply adding “site” as a covariate in the linear 18 risk estimations to obtain a single risk-estimation per patient. models27,28. We found the ComBat parameters (i.e., the MRI On another note, we tested two approaches to reduce the differences between sites) using the processed images from computational cost: applying additional subsampling exclusively the training set (i.e., we did not use the test set to find (6 × 6 × 6 mm3 or 12 × 12 × 12 mm3, instead of 3 × 3 × 3 mm3) the parameters). We then removed the effects of the site from the and limiting the analyses to statistically significant voxels processed images of both the training and the test sets using (P < 0.05 uncorrected at the univariate analysis). these parameters. We must highlight again that the effects of the We defined the accuracy of the age predictions as the mean site were estimated only using individuals from the training set absolute error (MAE) between the predicted and the actual age (i.e., not a single piece of information from the test set), thus and the accuracy of the diagnostic predictions as the proportion preventing any information leak. We have previously modified the of correct predictions. Finally, we assessed whether differences in ComBat functions to allow this separate estimation and applica- accuracy between the analysis using a given parameter and the tion of the ComBat parameters28. reference analysis (unmodulated gray matter smoothed with We also controlled the effects of the site when estimating the σ = 4 mm) were statistically significant by conducting a paired- model’s accuracy (see details later), which is important because sample Wilcoxon test of the absolute errors of the two analyses. the effects of the site might bias the accuracy even when researchers attempted to remove them during the creation of the Creation and validation of the HRR-FEP detection tool machine-learning model29. We used a cross-validation scheme to create the tool using a set of patients and validate it using a new set of patients. Specifically, we Optimization of MRI-based machine-learning parameters randomly divided the overall cohort into ten groups or “folds” trying We optimized the MRI-based machine-learning parameters using to preserve a similar number of relapses in each fold. First, we two independent datasets. Our main reason for using indepen- created the model using data from individuals from folds 2 to 10 (the dent datasets was to avoid any data leakage. We reasoned that if “training set”), and we estimated the relapse risk of individuals from we used the same cohort to optimize the machine-learning fold 1 (the “test set”) (Fig. 2). We then created the model using data parameters and create the risk-estimation model, we could end up from individuals from folds 1 and 3–10, and we estimated the relapse validating this model in patients we had previously used to risk of individuals from fold 2. And so on. Therefore, we could optimize the parameters (based on the best relapse risk estimate the relapse risk of all individuals, but we never used the estimations). We acknowledge that one strategy to prevent such same individuals for training and validating the model. data leakage would be optimizing the parameters separately for The creation of the HRR-FEP models in the training set consisted each fold via within-fold cross-validation using the training sets of fitting a multiple regression. The dependent variable was the exclusively. However, such a strategy could result in different MRI time to relapse. The independent variables were the clinical data parameters for the different folds, creating over-complexity in the (including the items from the symptom scales PANSS, GAF, model. Rather, we looked for general MRI settings that would be MADRS, YMRS, the diagnosis, and whether the patient was taking stable not only for the different folds but for different predictions long-acting injectable antipsychotic treatment) and the voxel or studies. values of the pre-processed MRI. Before conducting the regres- One dataset included 120 healthy individuals30, and we used sion, we removed the effects of age and sex from the training MRI their MRI data to predict a continuous variable (their age). The data with standard linear models. We must highlight once more other dataset included 255 individuals, half of them with a that the effects of age and sex were estimated only using schizophrenia diagnosis26,30,31, and we used their MRI data to individuals from the training set (i.e., not a single information from predict a binary variable (whether they had received the the test set), thus preventing any information leak. We also scaled Published in partnership with the Schizophrenia International Research Society Schizophrenia (2022) 100 A. Solanes et al. 4 INPUT CALCULATE TRAINING TESTING DATA ACCURACY Distribute subjects into train Clinical Data MRI Data Clinical Data MRI Data and test Create and Apply mask apply mask Create brain Create brain Create and Create and Missing Ensemble parons (to Missing Ensemble parons (to YES apply imputed YES YES apply imputed YES informaon? learning? be analyzed informaon? learning? be analyzed data data separately) separately) NO NO NO NO Fit and apply Apply Different sites? YES Different sites? YES COMBAT COMBAT NO NO Confounding variables Confounding variables Predictor variables Predictor variables Center and Calculate and Center and Remove effects scale predictor remove effects scale predictor of covariates variables of covariates variables Calculate Fit Model Apply Model accuracy Fig. 2 MRIPredict flowchart. Creation of the high relapse risk after the first episode of psychosis (HRR-FEP) detection tool. the clinical variables to the [0-1] range to have a distribution like To rule out whether the model’s accuracy could mainly depend the MRI voxels. To avoid overfitting, we used a “lasso” regression, on MRI data or clinical data, we also created HRR-FEP detection which automatically selects a few regressors by penalizing the tools exclusively based on MRI data or clinical data. Also, for sum of the absolute value of the coefficients and has been proven descriptive purposes, we mapped the brain regions univariately to be able to deal with high-dimensional data and still achieve associated with increased or decreased relapse risk after the FEP high-performance models33. A regularization parameter defines using standard survival analyses (see Supplement). the amount of penalization, ranging from null (no penalization, as We conducted the analyses with our freely available graphical in a standard regression) to infinity (maximum penalization). This software MRIPredict (which we have updated for this work), based regularization parameter is automatedly selected by the algorithm on the “glmnet” package for R (https://glmnet.stanford.edu/). via internal cross-validation within the training set. We chose the lasso regression algorithm for its good performance26, simplicity, Available resources and adequacy for survival analyses. All these previous steps Groups interested in conducting similar analyses can download estimated using the training set were applied later to the test set our free graphical-user-interface MRIPredict software at https:// to validate the performance of the model. www.mripredict.com/. In other words, we found a risk-estimation model using the We encourage independent groups to replicate our model’s patients of the training set exclusively, and afterward, we applied accuracy assessment. To help them, we provide a website-based the model to the patients of the test set to estimate their risk of version of the tool (https://www.mripredict.com/hrr-fep/) that relapse. To estimate a patient’s risk, we multiplied each coefficient quickly estimates the HRR-FEP of an individual. For the website, of the lasso model (see Table 3) by the value of the variable in the we fitted a model using the whole cohort and selected the patient and added the results. If the sum was >0 (corresponding coefficients with an absolute value ≥0.05 (see Supplement); its risk to a HR > 1), we considered that the patient was at HRR-FEP. estimations seem perfect (all relapses are in HRR-FEP individuals). Conversely, if the sum was ≤0 (corresponding to HR ≤ 1), we However, this accuracy is inflated because it uses the same considered the patient at low relapse risk. individuals for training and testing; we obtained a more reliable To test whether individuals estimated to be at HRR-FEP had accuracy estimation with cross-validation (see next). In addition, statistically more relapses than individuals at low relapse risk, we we only offer this tool to support replication by other researchers; used the “multisite.accuracy” package29, which considers the site’s the tool estimations must be considered experimental. residual effects when estimating the accuracy. Specifically, we conducted a mixed-effects Cox proportional hazards regression (https://CRAN.R-project.org/package=coxme). The dependent RESULTS variable was the time to relapse. The independent variable was Cohort description the estimated risk group (HRR-FEP vs. low relapse risk), and the There were 16 relapses, representing a 9.4% relapse rate at site was a random factor of no interest. 24 months. Note that while the number of relapses was limited, Schizophrenia (2022) 100 Published in partnership with the Schizophrenia International Research Society A. Solanes et al. 5 Table 2. Optimization of MRI-based machine-learning parameters. Adjustment Age predictions Diagnostic predictions MAE Absolute Relative Accuracy Absolute Relative P value(a) P value(b) P value(a) P value(b) Addition of features No (reference) 7.4 years <0.001 – 70.8% <0.001 – + white matter and modulated images 6.2 years <0.001 0.006 68.6% <0.001 0.039 + global volumes 7.4 years <0.001 n.s. 70.8% <0.001 n.s. + midline abnormalities – – – 70.7% <0.001 n.s. <0.001 Varying smoothing σ = 2 mm 7.8 years <0.001 n.s. 68.9% <0.001 n.s. kernel width σ = 3 mm 7.2 years <0.001 n.s. 69.4% <0.001 n.s. σ = 4 mm (reference) 7.4 years <0.001 – 70.8% <0.001 – σ = 5 mm 7.5 years <0.001 n.s. 71.2% <0.001 n.s. σ = 6 mm 7.5 years <0.001 n.s. 70.8% <0.001 n.s. <0.001 Subsampling Single (reference) 7.4 years <0.001 – 70.8% <0.001 – Double subsampling 7.5 years <0.001 n.s. 70.6% <0.001 n.s. Triple subsampling 7.3 years <0.001 n.s. 70.3% <0.001 n.s. Only statistically significant 7.4 years <0.001 n.s. 71.0% <0.001 n.s. <0.001 Ensemble No (reference) 7.4 years <0.001 – 70.8% <0.001 – Subjects 8.5 years <0.001 <0.001 64.4% <0.001 0.001 Voxels (half brains) 7.1 years <0.001 0.002 73.2% <0.001 0.034 <0.001 Optimal parameters + white matter and modulated images, 6.3 years <0.001 <0.001 74.4% <0.001 0.043 triple subsampling, the ensemble of voxels MAE mean absolute error. (a) Wilcoxon test comparing the predictions obtained with these settings with the predictions obtained with a null model (i.e., predicting that all individuals have the average age of the sample for age predictions, or to flipping a coin for diagnostic predictions). (b) Wilcoxon test comparing the predictions obtained with these setting with the predictions obtained with the reference settings (unmodulated gray matter smoothed with σ = 4 mm and no subsampling or ensemble). it still yielded enough statistical power to detect meaningful differences in relapse risk between groups (e.g., using the R package powerSurvEpi, we estimated that we had 70%/80%/ 90% power to detect a HR = 4.3/5.7/9.5). The median time from scan to relapse (in patients who had a relapse during the follow- up) was 7.4 months, and the median time from scan to the last follow-up visit (in patients with no relapse during the follow-up) was 23.7 months. We detected no statistically significant differences in relapse risk between affective and non-affective psychosis or between diagnoses, except increased risk in patients with a schizoaffective disorder diagnosis (HR = 3.6, P = 0.046). Optimal MRI-based machine-learning parameters When optimizing the MRI-based machine-learning parameters, we found that adding gray and white matter images, unmodulated modulated images, and the use of a voxel-level ensemble improved the accuracy (Table 2). Conversely, using subject-level ensemble worsened the accuracy. The other varying parameters did not influence accuracy. We thus selected the addition of gray and white matter images, unmodulated and modulated images, and the use of a voxel- Fig. 3 Observed relapses depending on estimated risk group. level ensemble for the HRR-FEP analyses. We also chose triple Kaplan–Meier curves of the observed relapse in patients estimated subsampling because it makes all calculations substantially less to be at high relapse risk after the first episode of psychosis (HRR- computationally expensive. FEP) vs. patients at low relapse risk. Published in partnership with the Schizophrenia International Research Society Schizophrenia (2022) 100 A. Solanes et al. 6 Table 3. Descriptive univariate analysis and machine-learning estimators of high relapse risk after the first episode of psychosis (HRR-FEP). Descriptive univariate analysis Machine learning Clinical variables Schizoaffective disorder HR = 3.6, P = 0.046 β = +0.24 ↓ Poor rapport (PANSS N3) – β = −0.01 ↓ Difficulty in abstract thinking (PANSS N5) HR = 0.6, P = 0.044 β = −0.074 ↓ Conceptual disorganization (PANSS P2) – β = −0.01 ↓ Poor attention (PANSS G11) – β = −0.04 ↑ Age in years HR = 1.1, P = 0.008 – ↑ Long-acting injectable antipsychotic – β = 0.01 Gray matter increase ↑ R Postcentral – Unm, [54, −6, 24], β = +0.93 Gray matter decrease ↓ R middle temporal – Unm, [66, −6, −12], β = −0.43 ↓ R inferior frontal/precentral – Unm, [30, 6, 36], β = −0.21 Mod, [42, 6, 36], β = −0.18 ↓ R middle frontal – Unm, [30, 42, 36], β = −0.20 ↓ R/L rectus Unm, [−6, 30, −36], z = −2.6 Mod, [6, 30, −24], β = −0.17 Unm, [6, 30, −24], β = −0.15 ↓ L superior frontal Unm, [−18, 66, −24], z = −2.8 – ↓ R medial frontal Unm, [6, 78, −12], z = −2.6 – ↓ R Angular – Unm, [30, −54, 36], β = −0.05 White matter increase ↑ R precentral – Unm, [42, 6, 36], β = +0.54 ↑ L Middle frontal – Unm, [−42, 6, 36], β = +0.10 White matter decrease ↓ R middle frontal – Unm, [30, 30, 36], β = -0.86 ↓ L inferior frontal – Mod, [−42, 18, 12], β = −0.73 Unm, [−42, 18, 12], β = −0.57 ↓ R Cuneus – Mod, [18, −90, 12], β = −0.18 ↓ R superior frontal Unm, [6, 54, 24], z = 2.9 – ↓ L corpus callosum Unm, [−18, 18, 24], z = 2.7 Unm, [−18, −30, 24], β = −0.05 ↓ R corpus callosum – Mod, [6, 30, 0], β = −0.05 ↓ L Middle frontal – Mod, [−30, 42, 12], β = −0.06 ↓ R postcentral – Mod, [54, −6, 24], β = −0.09 L left, Mod modulate, PANSS Positive and Negative Syndrome Scale, R right, Unm unmodulated. In the descriptive univariate analysis, we only report the peaks of MRI clusters with voxel uncorrected P value <0.005 and the clinical variables with uncorrected P value <0.05. For the sake of simplicity, we only report here the machine-learning coefficients with an absolute value ≥0.01 for clinical variables and ≥0.05 for MRI voxels; see the entire model in the Supplement. HRR-FEP detection variables detected in the descriptive univariate analysis and the The Cox regression of the time to relapse comparing patients machine-learning model. We report the entire machine-learning estimated to be at HRR-FEP vs. low relapse risk was clinically model in the Supplement. relevant (HR = 4.58, i.e., HRR-FEP patients had five times more risk The HRR-FEP detection tools exclusively using MRI data or solely to relapse) and had a (borderline) statistical significance (HR 95% based on clinical variables failed to detect patients at HRR-FEP. confidence interval = 1.01–20.74, Z = 1.98, P = 0.048, Fig. 3). The results were identical when we excluded young adolescents (i.e., 12–14-years-old). In the 114 individuals estimated to be at HRR- DISCUSSION FEP, there were 13 relapses, representing a 14.8% relapse rate at In this work, we created an MRI-based machine-learning tool to 24 months. Conversely, there were only three relapses in the 113 detect those patients at HRR-FEP using a cohort of 227 individuals individuals estimated to not be at HRR-FEP, representing a 2.9% with a FEP. The model showed to detect HRR-FEP successfully. The relapse rate at 24 months. Using the R package powerSurvEpi, we hazard of relapse was 4.5 larger in individuals estimated to be at estimated that the power to detect a HR = 4.58 with 16 relapses is HRR-FEP than in low relapse risk individuals (14.8% vs. 2.9% 72%. The variables automatedly selected by the lasso regression relapse rate at 2 years), and we estimated the power to detect to create the HRR-FEP detection tool were the diagnosis of such a hazard ratio of 4.5 with 16 relapses is 72%. schizoaffective disorder, the lack of difficulty in abstract thinking The study thus achieved the aim of creating a tool that may and poor impulse control, and the increase or decrease of provide valuable information to the mental health professional. unmodulated and modulated gray and white matter in several Ideally, the clinician could input the tool with a few MRI and clinical brain regions. Table 3 details the specific brain regions and clinical data to know if the patient is estimated to be at HRR-FEP or not, Schizophrenia (2022) 100 Published in partnership with the Schizophrenia International Research Society A. Solanes et al. 7 and thus adjust the prophylactic treatment. Knowing this One key variable selected by the lasso regression was the information early is important because currently, clinicians can diagnosis of schizoaffective disorder; this partly agrees with previous only know which patients are at HRR-FEP after several relapses. studies reporting associations of diagnosis or manic symptoms with And before that, patients at HRR-FEP may experience repeated increased relapse rate12,13,15. In addition, we think that in the current relapses if the prevention is too weak, while patients at low relapse debate about the validity of DSM/ICD diagnoses, it is worth noting risk may experience increased adverse events if the prevention is that diagnostic labels more than clinical scales helped predict future too strong. That said, any adjustment of the prophylactic treatment relapses. That said, this debate is entirely out of the scope of this should follow the “first, do no harm” principle because in our paper. On another note, the protective effects of the difficulty in cohort, most (85%) individuals estimated to be at HRR-FEP did abstract thinking and poor impulse control are intriguing. We indeed not relapse. Not less important, the clinician could also speculate that these symptoms may be related to latent disorder consider removing or reducing the prophylactic treatment in subtypes that might be clearer in subsequent phases of the illness. individuals estimated to be unlikely to relapse. These patients Finally, we must acknowledge that the variables showing statistical currently may be receiving treatments that, if the patient is truly significance in the descriptive univariate analysis (see Supplement) unlikely to relapse, may be little useful while harmful. were primarily different from the variables selected by the lasso However, in any case, we want to highlight the need to validate regression. However, this disagreement is expectable because the the HRR-FEP detection tool before recommending it. We have noted latter only aims to predict and thus discards brain regions that do not previously that independent studies often fail to replicate the add much to the prediction accuracy, even if they are statistically accuracy reported in mental health machine-learning publications34, significant when considered alone. and our study may not be an exception. We cannot share participant Before creating the HRR-FEP detection tool, we used two data for privacy reasons. However, we provide the trained classifiers independent datasets to find the optimal parameters for VBM- online so that independent researchers can still try to replicate our based machine learning. Finding the optimal parameters in two study results. This approach has been stated to be one of the most different datasets keeps the main study data unseen until we convincing forms of replication35. However, without intending to create the model for the HRR -FEP detection tool. We acknowl- create hype, we also think that our work shows the potential clinical edge that the accuracy of the age predictions was lower than that utility of MRI-based machine-learning when understood as a source reported elsewhere41. This lower accuracy was probably related to of additional information for the psychiatrist. the limited sample size of the age prediction dataset. However, we We also want to highlight that this tool could be complemented only aimed to compare the accuracy depending on different by other tools that update the relapse risk during the follow-up. parameters. We found that the optimal parameters were the For example, we have reported for other disorders that the relapse addition of gray and white matter images, the addition of risk at 12 months substantially decreases in patients who have unmodulated and modulated images, and the use of voxel-level been relapse-free for at least one year24. Thus, some patients ensemble. We encourage future studies to use these parameters. initially at HRR-FEP may later be at low relapse risk. Similarly, Also, we found that even triple subsampling did not affect the information about changes in the first months could also likely accuracy while substantially reducing computational costs. offer valuable information for updating the risk estimation20. In We want to comment that, while previous work has searched this context, we would like to note that, as far as relapses also for gold biomarkers with little success, this work shows the depend on events that will happen during the follow-up, it is potential clinical use of MRI-based machine learning in risk unlikely that a machine-learning model that only uses baseline assessment. We speculate that such risk assessment will very data scan achieves high risk-estimation accuracy. likely be far from perfect, i.e., we will not be able to know for A particularity of our study is that, instead of focusing on sure which patients will have a relapse and which will not, or detecting those healthy individuals at high risk for FEP, it focuses the date of the relapse. Indeed, such predictions may seem on detecting those FEP patients at HRR. Many studies have already unrealistic considering that relapses also depend on life events been published regarding predicting transition to psychosis, with and stressors after the assessment 42. However, the estimation varied results3,4. Conversely, no studies have been conducted to will be clinically valuable as far as we can estimate risk with estimate HRR-FEP from MRI data to our knowledge. This lack of enough accuracy to help the physician, i.e., so that the research is striking because assessing the relapse risk is essential information translates into in an effective improvement of the to properly adjusting the preventive antipsychotic dose. care. Our study does not provide this level of accuracy yet, but Our tool requires an MRI, but patients with a FEP may indeed we hope to have made a step for future studies. already undergo an MRI to discard organic brain pathology, so that This work has some limitations. First, this sample does not the structural MRI required for our tool would serve both. This fact include the patients who did not meet the inclusion criteria or increases the feasibility of the HRR-FEP detection tool, given that for refused to participate in the study, who may differ from those many patients, it would only involve minor calculations on any included. It is a common limitation in many other studies. Second, computer. The context is different, for instance, for the detection of even if we included 227 patients and followed them for individuals with a higher risk of psychosis in the general population, 18–24 months, representing one of the largest brain imaging where screening detection tools should only require inputting a FEP cohorts worldwide, there were only 16 relapses. This relapse small amount of available information. An example of such a rate is lower than those reported in some previous cohorts24,43,44. screening detection tool is the Psychosis Polyrisc Score (PSS)36, which To check whether the difference in relapse rate was due to our only asks about the presence of a few risk factors37 and has shown relapse criteria being only based on PANSS while others also feasible in a real-world digital implementation38. considered hospitalizations, we retrieved hospitalizations, and the Interestingly, the accuracy of HRR-FEP detection tools was updated relapse rate (37%) was more in agreement with previous poorer when we created machine-learning models that used only cohorts. However, we could not successfully repeat the analyses clinical data or only MRI data. Ad hoc, it may seem evident that the with hospitalizations because this information was unavailable on more information, the better the detection. However, many some sites. Third, the statistical significance was weak, probably previous studies only used MRI to find biomarkers that should due to our cohort’s limited number of relapses. In any case, the surpass clinical judgment. These may include serum component power to detect a hazard of relapse of 4.5 with the sample size protein 4 (C4)39, polygenic related Risk Score (PRS)40, neuroana- and the number of relapses in this study was 72%, very close to tomical variables18,20. Thus, poetically, we have found that, in the the conventional 80% required in sample size calculations. Fourth, fight between clinical-based and biomarker-based psychiatry, more complex machine-learning algorithms, such as neural joining efforts predicts better. networks, might detect more patterns than the relatively simple Published in partnership with the Schizophrenia International Research Society Schizophrenia (2022) 100 A. Solanes et al. 8 algorithms used here. However, these algorithms usually require 9. Bernardo, M. et al. The prevention of relapses in first episodes of schizophrenia: substantially larger cohorts, which may be challenging to achieve. the 2EPs Project, background, rationale and study design. Revista de psiquiatria y Fifth, for simplicity, we considered patients estimated to have a salud mental 14, 164–176 (2021). HR > 1 at HRR-FEP. However, the optimal division between groups 10. Ascher-Svanum, H. et al. The cost of relapse and the predictors of relapse in the treatment of schizophrenia. BMC Psychiatry 10, 2 (2010). could be at another HR threshold. Future studies evaluating the 11. Bhattacharyya, S. et al. Individualized prediction of 2-year risk of relapse as benefits and costs of the interventions at different HR levels may indexed by psychiatric hospitalization following psychosis onset: model devel- provide more insights into this question. Sixth, we could not opment in two first episode samples. Schizophrenia Res. 228, 483–492 (2021). evaluate medication adherence, DUP, and premorbid functioning 12. Puntis, S., Whiting, D., Pappa, S. & Lennox, B. Development and external valida- because data was missing in some sites. Due to its established role tion of an admission risk prediction model after treatment from early interven- in relapse, the use of these variables could improve model tion in psychosis services. Transl. Psychiatry 11, 35 (2021). accuracy. Finally, we could not report statistics such as sensitivity 13. Arrasate, M. et al. Prognostic value of affective symptoms in first-admission psychotic and specificity. We could not estimate such statistics because our patients. Int. J. Mol. Sci. 17, https://doi.org/10.3390/ijms17071039 (2016). data was not binary (relapse vs. not relapse). Note that 38% of 14. Wunderink, L. et al. Negative symptoms predict high relapse rates and both predict less favorable functional outcome in first episode psychosis, independent patients did not complete the follow-up, and thus we could not of treatment strategy. Schizophrenia Res. 216, 192–199 (2020). classify them as relapse or not relapse - we knew that they had not 15. Hui, C. L. et al. Predicting first-episode psychosis patients who will never relapse relapsed until the last visit, but we did not know if they had over 10 years. Psychological Med. 49, 2206–2214 (2019). relapsed afterward. However, even if there were no follow-up 16. Berge, D. et al. Predictors of relapse and functioning in first-episode psychosis: a losses, we would still report the Cox regression as the primary two-year follow-up study. Psychiatric Services 67, 227–233 (2016). validation statistic because it considers whether relapses occurred 17. Schoeler, T. et al. Poor medication adherence and risk of relapse associated with earlier or later. In contrast, binary statistics do not. continued cannabis use in patients with first-episode psychosis: a prospective To conclude, this study might represent a step towards a analysis. Lancet. Psychiatry 4, 627–633 (2017). translational application of neuroimaging to mental health. Up to 18. Nieuwenhuis, M. et al. Multi-center MRI prediction models: predicting sex and illness course in first episode psychosis patients. NeuroImage 145, 246–253 (2017). now, brain imaging prediction models have mainly aimed to imitate 19. Dazzan, P. et al. Clinical utility of MRI scanning in first episode psychosis. Schi- clinical judgment, for example, by training a support vector machine zophrenia Bull. 44, S50–S51 (2018). to differentiate between patients and controls based on their brain 20. Cahn, W. et al. Brain volume changes in the first year of illness and 5-year images45. Conversely, we combined clinical and MRI data to improve outcome of schizophrenia. Br. J. Psychiatry 189, 381–382 (2006). the accuracy of a tool that, instead of finding reliable biomarkers, 21. Alvarez-Jimenez, M., Parker, A. G., Hetrick, S. E., McGorry, P. D. & Gleeson, J. F. aims to help the clinician, ultimately paving the way toward more Preventing the second episode: a systematic review and meta-analysis of psy- personalized medicine in mental disorders. chosocial and pharmacological trials in first-episode psychosis. Schizophrenia Bull. 37, 619–630 (2011). 22. Pina-Camacho, L. et al. Age at first episode modulates diagnosis-related structural DATA AVAILABILITY brain abnormalities in psychosis. Schizophrenia Bull. 42, 344–357 (2016). 23. Berge, D. et al. Elevated extracellular free-water in a multicentric first-episode Data are available upon request to the Research Ethics Committees of Benito Menni psychosis sample, decrease during the first 2 years of illness. Schizophrenia Bull. CASM, Hospital General de Granollers, Hospital de Mataró, Hospital Sant Rafael, Hospital https://doi.org/10.1093/schbul/sbz132 (2020). de Bellvitge, Hospital Clínic de Barcelona, Hospital Universitario 12 de Octubre, Hospital 24. Radua, J., Grunze, H. & Amann, B. L. Meta-analysis of the risk of subsequent mood Clínic de València, Hospital del Mar, Instituto de Investigación Sanitaria Aragón, Hospital episodes in bipolar disorder. Psychother. Psychosomatics 86, 90–98 (2017). General Universitario Gregorio Marañón, Hospital Sant Joan de Déu Barcelona, Hospital 25. Andreasen, N. C. et al. Remission in schizophrenia: proposed criteria and rationale Santiago Apóstol de Vitoria-Gasteiz, and Servicio de Salud del Principado de Asturias. for consensus. Am J Psychiatry 162, 441–449 (2005). 26. Salvador, R. et al. Evaluation of machine learning algorithms and structural fea- tures for optimal MRI-based diagnostic prediction in psychosis. PLoS ONE 12, CODE AVAILABILITY e0175683 (2017). Groups interested in conducting similar analyses can download and check the R code 27. Fortin, J. P. et al. Harmonization of cortical thickness measurements across of the software used to conduct the analysis of this study, and use our free graphical- scanners and sites. NeuroImage 167, 104–120 (2018). user-interface MRIPredict software at https://www.mripredict.com/. 28. Radua, J. et al. Increased power by harmonizing structural MRI site differences with the ComBat batch adjustment method in ENIGMA. NeuroImage 218, 116956 (2020). Received: 10 June 2022; Accepted: 28 October 2022; 29. Solanes, A. et al. Biased accuracy in multisite machine-learning studies due to incomplete removal of the effects of the site. Psychiatry Res. Neuroimaging 314, 111313 (2021). 30. Radua, J. et al. Anisotropic kernels for coordinate-based meta-analyses of neu- roimaging studies. Front. Psychiatry 5, 13 (2014). REFERENCES 31. Landin-Romero, R. et al. Midline brain abnormalities across psychotic and mood 1. DeLisi, L. E. et al. Cerebral ventricular enlargement as a possible genetic marker disorders. Schizophrenia Bull. 42, 229–238 (2016). for schizophrenia. Psychopharmacol. Bull. 21, 365–367 (1985). 32. Kasai, K. et al. Cavum septi pellucidi in first-episode schizophrenia and first- 2. Radua, J. & Carvalho, A. F. Route map for machine learning in psychiatry: absence of episode affective psychosis: an MRI study. Schizophrenia Res. 71, 65–76 (2004). bias, reproducibility, and utility. Eur. Neuropsychopharmacol. 50, 115–117 (2021). 33. Greenshtein, E. & Ritov, Y. A. Persistence in high-dimensional linear predictor 3. Rosen, M. et al. Towards clinical application of prediction models for transition to selection and the virtue of overparametrization. Bernoulli 10, 971–988, 918 psychosis: a systematic review and external validation study in the PRONIA (2004). sample. Neurosci. Biobehav. Rev. 125, 478–492 (2021). 34. Radua, J. What is the actual accuracy of clinical prediction models? The case of 4. Smieskova, R. et al. Neuroimaging predictors of transition to psychosis-a sys- transition to psychosis. Neurosci. Biobehavioral Rev. 127, 502–503 (2021). tematic review and meta-analysis. Neurosci. Biobehav. Rev. 34, 1207–1222 (2010). 35. Young, J., Kempton, M. J. & McGuire, P. Using machine learning to predict out- 5. Schmidt, A. et al. Improving prognostic accuracy in subjects at clinical high risk comes in psychosis. Lancet Psychiatry 3, 908–909 (2016). for psychosis: systematic review of predictive models and meta-analytical 36. Oliver, D., Radua, J., Reichenberg, A., Uher, R. & Fusar-Poli, P. Psychosis polyrisk sequential testing simulation. Schizophrenia Bull. 43, 375–388 (2017). score (PPS) for the detection of individuals at-risk and the prediction of their 6. Salazar de Pablo, G. et al. Probability of transition to psychosis in individuals at outcomes. Front. Psychiatry 10, 174 (2019). clinical high risk: an updated meta-analysis. JAMA Psychiatry https://doi.org/ 37. Radua, J. et al. What causes psychosis? An umbrella review of risk and protective 10.1001/jamapsychiatry.2021.0830 (2021). factors. World Psychiatry 17, 49–66 (2018). 7. Fortea, A. et al. Cortical gray matter reduction precedes transition to psychosis in 38. Oliver, D. et al. Real-world digital implementation of the psychosis polyrisk score individuals at clinical high-risk for psychosis: a voxel-based meta-analysis. Schi- (PPS): a pilot feasibility study. Schizophrenia Res. 226, 176–183 (2020). zophrenia Res. 232, 98–106 (2021). 39. Mondelli, V. et al. Baseline high levels of complement component 4 predict worse 8. Harrison, G. et al. Recovery from psychotic illness: a 15- and 25-year international clinical outcome at 1-year follow-up in first-episode psychosis. Brain Behav. follow-up study. Br. J. Psychiatry 178, 506–517 (2001). Immunity 88, 913–915 (2020). Schizophrenia (2022) 100 Published in partnership with the Schizophrenia International Research Society A. Solanes et al. 9 40. Harrisberger, F. et al. Impact of polygenic schizophrenia-related risk and hippo- Recordati. CDC has received honoraria from Sanofi and Exeltis. Dr. Parellada has campal volumes on the onset of psychosis. Transl. Psychiatry 6, e868–e868 (2016). received educational honoraria from Otsuka, research grants from Instituto de Salud 41. Baecker, L. et al. Brain age prediction: a comparison between machine learning Carlos III (ISCIII), Ministry of Health, Madrid, Spain, has received grant support from models using region- and voxel-based morphometric data. Human Brain Mapping ISCIII, Horizon2020 of the European Union, CIBERSAM, Fundación Alicia Koplowitz, 42, 2332–2346 (2021). and Mutua Madrileña and travel grants from Otsuka, Exeltis and Janssen; she has 42. Simhandl, C., Radua, J., Konig, B. & Amann, B. L. The prevalence and effect of life served as a consultant for Servier, Exeltis, Fundación Alicia Koplowitz, and ISCIII. LPC events in 222 bipolar I and II patients: a prospective, naturalistic 4 year follow-up has received honoraria or grants unrelated to the present work from Rubió, Rovi, and study. J. Affective Disorders 170, 166–171 (2015). Janssen. Dr. R. Rodriguez-Jimenez has been a consultant for, spoken in activities of, or 43. Robinson, D. et al. Predictors of relapse following response from a first episode of received grants from Instituto de Salud Carlos III, Fondo de Investigación Sanitaria schizophrenia or schizoaffective disorder. Archives General Psychiatry 56, 241–247 (FIS), Centro de Investigación Biomédica en Red de Salud Mental (CIBERSAM), Madrid (1999). Regional Government (S2010/BMD-2422 AGES; S2017/BMD-3740), Janssen-Cilag, 44. Tiihonen, J. et al. A nationwide cohort study of oral and depot antipsychotics Lundbeck, Otsuka, Pfizer, Ferrer, Juste, Takeda, Exeltis, Casen-Recordati, Angelini. Dr. after first hospitalization for schizophrenia. Am. J. Psychiatry 168, 603–609 (2011). Arango has been a consultant to or has received honoraria or grants from Acadia, 45. Nieuwenhuis, M. et al. Classification of schizophrenia patients and healthy con- Angelini, Biogen, Boehringer, Gedeon Richter, Janssen Cilag, Lundbeck, Medscape, trols from structural MRI scans in two large independent samples. NeuroImage Menarini, Minerva, Otsuka, Pfizer, Roche, Sage, Servier, Shire, Schering Plough, 61, 606–612 (2012). Sumitomo Dainippon Pharma, Sunovion and Takeda. Dr. Vieta has received grants and served as a consultant, advisor, or CME speaker for the following entities (work unrelated to the topic of this manuscript): AB-Biotics, Abbott, Allergan, Angelini, ACKNOWLEDGEMENTS Dainippon Sumitomo Pharma, Galenica, Janssen, Lundbeck, Novartis, Otsuka, Sage, We are grateful to all participants. We would also like to thank the Instituto de Salud Sanofi-Aventis, and Takeda. The remaining authors report no financial relationships Carlos III, the Spanish Ministry of Science, Innovation, and Universities, the European with commercial interests. Regional Development Fund (ERDF/FEDER), European Social Fund, “Investing in your future”, “A way of making Europe” (projects: PI08/0208, PI11/00325, PI14/00292, PI14/ 00612, PI14/01148, PI14/01151, PI17/00481, PI17/01997, PI18/01055, PI19/00394, ADDITIONAL INFORMATION PI19/00766, PI20/00721, PI20/01342, and PI21/00713; contracts: CP14/00041, JR19/ Supplementary information The online version contains supplementary material 00024, CPII19/00009, and CD20/00177); CERCA Program; Catalan Government, the available at https://doi.org/10.1038/s41537-022-00309-w. Secretariat of Universities and Research of the Department of Enterprise and Knowledge (2017SGR01271, 2017SGR1355, and 2017SGR1365); Institut de Neuro- Correspondence and requests for materials should be addressed to Edith Pomarol- ciències, Universitat de Barcelona; Madrid Regional Government (B2017/BMD-3740 Clotet or Joaquim Radua. AGES-CM-2); the University of the Basque Country (2019 321218ELCY GIC18/107); the Basque Government (2017111104); European Union Structural Funds, European Reprints and permission information is available at http://www.nature.com/ Union Seventh Framework Program, European Union H2020 Program under the reprints Innovative Medicines Initiative 2 Joint Undertaking (grant agreement No 115916, Project PRISM, and grant agreement No 777394, Project AIMS-2-TRIALS), Fundación Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims Familia Alonso, Fundación Alicia Koplowitz and Fundación Mutua Madrileña. The in published maps and institutional affiliations. funding organizations played no role in the study design, data collection, analysis, or manuscript approval. Open Access This article is licensed under a Creative Commons AUTHOR CONTRIBUTIONS Attribution 4.0 International License, which permits use, sharing, J.R. and A.S. contributed to the data analysis and writing of the paper. J.R. and E.P. adaptation, distribution and reproduction in any medium or format, as long as you give designed the study. E.P., J.R., G.M., J.J., S.A., A.L., A.G., C.A., E.V., J.C., D.B., A.A., E.G., M.P., appropriate credit to the original author(s) and the source, provide a link to the Creative M.B., and PEPs group (collaborators) contributed in the data collection. J.R. Commons license, and indicate if changes were made. The images or other third party contributed acquiring funding. All authors critically revised the manuscript and material in this article are included in the article’s Creative Commons license, unless approved the completed version. All authors are accountable for this work. indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http:// COMPETING INTERESTS creativecommons.org/licenses/by/4.0/. Dr. Bernardo has been a consultant for, received grant/research support and honoraria from, and been on the speakers/advisory board of AB-Biotics, Adamed, Angelini, Casen-Recordati, Janssen-Cilag, Menarini, Rovi, and Takeda. Dr. C. De-la- © The Author(s) 2022 Camara received financial support to attend scientific meetings from Janssen, Almirall, Lilly, Lundbeck, Rovi, Esteve, Novartis, AstraZeneca, Pfizer, and Casen- PEPs GROUP (COLLABORATORS) Miquel Bioque1,4,5,8, Constanza Morén20,21, Laura Pina-Camacho5,7, Covadonga M. Díaz-Caneja5,7, Iñaki Zorrilla5,11,12,13, Edurne Garcia Corres5,11,12,13, Concepción De-la-Camara5,9,10,22, Fe Barcones5,9,10,23, María José Escarti5,24,25, Eduardo Jesus Aguilar5,24,25,26, Teresa Legido16, Marta Martin16, Norma Verdolini1,5,8,14, Anabel Martinez-Aran1,5,8,14, Immaculada Baeza5,15, Elena de la Serna1,5,8,15, Fernando Contreras5,27, Julio Bobes5,28,29,30,31, María Paz García-Portilla5,28,29,30,31, Luis Sanchez-Pastor32, Roberto Rodriguez-Jimenez5,32,33, Judith Usall34, Anna Butjosa5,35, Pilar Salgado-Pineda2,5 and Raymond Salvador2,5 20 Centro de Investigación Biomédica en Red (CIBER) de Enfermedades Raras (CIBERER), Madrid, Spain. 21Cellex, IDIBAPS, University of Barcelona-Hospital Clínic of Barcelona, Barcelona, Spain. 22Hospital Clínico Universitario, Zaragoza, Spain. 23Servicio Aragonés de la Salud, Centro de Salud de Tarazona, Zaragoza, Spain. 24Department of Psychiatry, Hospital Clínico Universitario de Valencia, Valencia, Spain. 25Biomedical Research Institute INCLIVA, Valencia, Spain. 26Department Psychiatry, Faculty of Medicine, University of Valencia, Valencia, Spain. 27Psychiatry Unit. Bellvitge University Hospital. IDIBELL, Barcelona, Spain. 28Department of Psychiatry, Universidad de Oviedo, Oviedo, Spain. 29Servicio de Salud del Principado de Asturias (SESPA), Oviedo, Spain. 30Instituto de Investigación Sanitaria del Principado de Asturias (ISPA), Oviedo, Spain. 31Instituto Universitario de Neurociencias del Principado de Asturias (INEUROPA), Oviedo, Spain. 32Instituto de Investigación Sanitaria Hospital 12 de Octubre (imas12), Madrid, Spain. 33 CogPsy Group, Universidad Complutense de Madrid (UCM), Madrid, Spain. 34Parc Sanitari Sant Joan de Déu, Teaching, Research & Innovation Unit, Institut de Recerca Sant Joan de Déu, Barcelona, Spain. 35Hospital Infanto-juvenil Sant Joan de Déu, Institut de Recerca Sant Joan de Déu, Esplugues de Llobregat, Barcelona, Spain. Published in partnership with the Schizophrenia International Research Society Schizophrenia (2022) 100