Final Report Summary - PSYSCAN (Translating neuroimaging findings from research into clinical practice)
Executive Summary:
PSYSCAN is a multi-centre research programme, involving 22 academic and industry partners across Europe, Australia, Asia, and the Americas (see Figure 1). Its aim is to build tools that will help predict clinical outcomes in patients in the early stages of psychosis by integrating clinical, cognitive, and neuroimaging data. We anticipate that these tools will allow academics and industry to stratify patient samples according to clinical outcomes.
PSYSCAN comprises 6 highly integrated work packages (WPs, see Figure 2). WP1 relates to management and scientific coordination. The main aims of the remaining WPs were to:
• WP2: Collate legacy datasets comprising neuroimaging and non-neuroimaging measures in subjects at clinical high risk (CHR) for psychosis, patients with first-episode psychosis (FEP) and healthy controls (HC), then use these datasets to evaluate novel methods of analysis that may be employed in the new tools.
• WP3: Develop new software that applies machine learning to multi-modal data to stratify patients according to clinical outcomes.
• WP4: Develop patient stratification tools that employ the technology developed in WP3.
• WP5: Conduct a large multi-centre naturalistic study, collecting longitudinal neuroimaging and non-neuroimaging data from CHR, FEP and HC groups, to provide data for developing and testing the PSYSCAN tool.
• WP6: Interact with representatives from key stakeholder groups and networks within the field of psychosis and disseminate the results of the PSYSCAN project.
The PSYSCAN project ran from February 2014 to July 2021. Despite substantial disruptions due to the Covid pandemic, it has been highly successful:
• WP2 collated 16 legacy datasets, comprising over 3000 subjects. Analyses of these have facilitated the development of machine learning algorithms used in WP3 and generated six publications.
• WP3 created state-of-the-art pre-processing pipelines to build structural and functional brain networks which have been applied to data acquired in WP5 to investigate how brain connectivity is altered in psychotic disorders.
• WP4 developed a novel iPad-based electronic data capture system that was used to collect clinical, demographic, and cognitive data in WP5 participants, and an online database (implementing automated quality control checks) for integrating legacy and newly-acquired PSYSCAN MRI data. WP4 also tested the ability of the novel analytic approaches developed in WP3 to predict clinical progression in the WP5 cohort and developed a prototype enrichment tool.
• WP5 recruited 328 FEP, 239 CHR, and 139 HC individuals, in whom clinical, neuroimaging cognitive, and genetic data were obtained. Follow-up of FEP/HC cohorts is now complete.
• WP6 established a new database containing contact information for more than 350 stakeholders among psychiatry professionals, professional associations, academic institutions and research bodies, patient organizations, and non-governmental organizations. It also created a PSYSCAN website (www.psyscan.eu) and a PSYSCAN twitter account for disseminating project findings.
Project Context and Objectives:
Context
Psychotic disorders are common, severely disabling, and associated with enormous clinical and socioeconomic impact. However, at present, we cannot predict their onset, nor their subsequent course. The onset of psychosis is preceded by attenuated psychotic symptoms and a decline in social and vocational functioning, with help-seeking individuals presenting with these features termed as being at clinical high-risk (CHR) for psychosis. Some CHR individuals will go on to develop a psychotic disorder; in others, symptoms persist without progression to frank illness, while in a third group symptoms will resolve. In those who do develop a psychosis, the subsequent course of illness is again unpredictable: some will make a good recovery, others will recover and then relapse in further episodes, and another subgroup will follow a chronic, unremitting course. The inability to predict what will happen to an individual at CHR, and to a patient who has just become psychotic, presents significant difficulties for tailoring psychiatric care to the needs of each patient.
Our inability to predict the onset and outcome of psychosis reflects a lack of objective and quantitative criteria to classify these patients. In fact, despite substantial advances in our understanding of the neurobiological basis of psychotic disorders, the assessment of patients with psychosis is still based on a clinical interview. Moreover, there is overwhelming evidence that psychotic disorders are pathophysiologically heterogeneous, which hampers the search for biomarkers to aid early diagnosis, stratification, and the measurement of disease progression. Indeed, this lack of biomarkers is perceived by industry as a key barrier to the development of novel treatments. Stratifying patients in an objective, quantitative way according to clinical and functional outcomes may generate patient clusters that are more clinically meaningful. Our industrial partners are particularly interested in the application of stratification to the recruitment of patients with established psychosis for clinical trials. It is possible that a new drug may work well in one subgroup of patients but not in another. In the absence of patient stratification, a true therapeutic effect may be undetectable when it is diluted across a heterogeneous sample.
Developing methods for stratifying individuals at CHR will is also of great importance. A key problem for both academic studies and pharmaceutical industry clinical trials is that, at present, these studies include all individuals who meet CHR criteria, even though only a small proportion (~16%) will later develop psychosis. As such, even studies with relatively large samples have limited ability test which prognostic factors or preventive treatments may reduce the risk of psychosis. Developing prediction tools will allow academics and industry to enrich CHR samples by selectively enrolling the subgroup at greatest risk of psychosis, and thus provide greater statistical power. This is attractive to academics and industry, as it reduces the size of the samples required to detect a significant prognostic or treatment effect. It would also reduce ethical concerns about giving preventive treatments to all CHR subjects when only a minority may actually need them.
Neuroimaging research has substantially advanced our understanding of the biological basis of psychotic disorders, with research in this field having the greatest potential for identifying quantitative objective measures in psychiatry. However, several factors have hampered the translation of findings from neuroimaging research to clinical, industry, and academic settings. Firstly, most neuroimaging research examines statistical differences between mean values from groups – findings which cannot be applied at the individual patient level. Secondly, the lack of standardisation of data acquisition or analysis approaches across different centres has made it difficult to combine samples to produce large scale multi-centre datasets with sufficient power to be predictive at the individual patient-level. Finally, previous research has typically examined single neuroimaging measures; however, the complex multifactorial aetiology of mental disorders (particularly psychotic disorders), indicates that that their onset and course will be best predicted by combining multiple neuroimaging and non-imaging.
To this end, the PSYSCAN project aimed to develop a neuroimaging-based tool to aid prediction and prognosis during the early course of illness and undertake activities to promote the translation of neuroimaging research findings into practice for the benefit of patients.
Primary objectives
1) To develop tools for the prediction of clinical outcomes in patients in the early phase of psychosis, with two main applications:
• PSYSCAN-Predict: Prediction of psychosis onset in CHR subjects.
• PSYSCAN-Stratify: Prediction of treatment response and illness course in FEP subjects.
2) To promote the translational application of neuroimaging in psychiatry and disseminate the results of the PSYSCAN project and the principles of using biomarkers to predict clinical outcomes in psychosis across European Member States.
To achieve these objectives, we proposed the following steps:
• Use legacy datasets from previous studies to identify neuroimaging measures associated with clinical outcomes in psychosis.
• Develop novel analytical methods that permit predictions of outcomes at the level of the individual patient, and thus have high translational potential.
• Assess whether prediction accuracy can be improved by basing this on multiple neuroimaging measures, plus clinical and cognitive measures, rather than a single neuroimaging measure alone.
• Employ outcomes that are clinically meaningful.
• Partner with imaging (IXICO) and cognitive testing (Cambridge Cognition) companies to standardise and harmonise neuroimaging data acquisition across European clinical research centres.
• Work with industrial partners (IXCO and Cambridge Cognition) to develop a novel electronic data capture system.
• Conduct multi-centre studies to acquire new datasets, using standardized protocols and a novel electronic data capturing system, and use these data to validate the PSYSCAN tools.
• Establish a network of representatives from key-stakeholder groups and networks within the field of psychosis and disseminate the tool to relevant stakeholders and end-users.
To achieve these goals, we established an international, multi-centre consortium, involving 22 academic and industry partners across Europe (London, Utrecht, Madrid, Cambridge, Maastricht, Edinburgh, Santander, Copenhagen, Galway, Naples, Amsterdam, Marburg, Tel Hashomer, Vienna Zürich, Heidelberg), Australia (Melbourne), Asia (Seoul, Hong Kong) and the Americas (Toronto, Sao Paulo).
Project Results:
Work Package 2: Legacy data
WP2 collated legacy datasets from previous studies of neuroimaging and non-neuroimaging measures in psychosis. In total, retrospective data were collected from over 3000 subjects, using 16 datasets from 11 of the consortium partners. These data comprised patients with established psychosis (n=722), patients with FEP (n=688), individuals at CHR (n=99), individuals at genetic risk for psychosis (22q11.2 deletion syndrome, n=30), unaffected siblings of patients with psychosis (n=295) and healthy controls (n=1642). The data included structural MRI, DTI, and fMRI data, as well as non-neuroimaging data (demographic and clinical data). Our analyses of these datasets generated six publications in peer-reviewed journals (see Dissemination section). The main results/foregrounds from these are summarised below:
Using neuroimaging data to enhance our understanding of the neurobiological basis of psychosis.
Structural MRI (sMRI) data were used to investigate differences in grey matter volume between healthy individuals and those with FEP (Viera et al., Psychological Medicine, 2021, 51(2): 340-350). Voxel-based morphometry was applied to sMRI data obtained from 572 FEP subjects and 502 age- and gender-comparable HC subjects. Analyses identified a widespread pattern of decreased grey matter in fronto-temporal, insular and occipital regions bilaterally among FEP individuals relative to controls; these decreases were not dependent on antipsychotic medication. The region with the most pronounced decrease - gyrus rectus - was negatively correlated with the severity of positive and negative symptoms. These findings provide evidence for reliable neuroanatomical alternations in FEP, expressed above and beyond site-related differences in antipsychotic medication, scanning parameters, and recruitment criteria.
Application of machine learning methods to neuroimaging data to develop predictive models for diagnostic and prognostic classification.
Three publications from WP2 involved the use of legacy data to evaluate machine learning (ML) methods to that could subsequently be incorporated into PSYSCAN tools.
The first of these (Vieira et al., Schizophrenia Bulletin, 2020 46(1): 17-26) investigated the extent to which the application of ML to neuroanatomical data allows detection of FEP after implementing methodological precautions to avoid overoptimistic results. This study tested both traditional ML and an emerging approach known as deep learning (DL) using 3 feature sets of interest: (1) surface-based regional volumes and cortical thickness, (2) voxel-based grey matter volume (GMV) and (3) voxel-based cortical thickness (VBCT). Analyses were repeated across 5 independent datasets (in total, 514 FEP and 444 within-site matched controls) to assess the reliability of findings with performance assessed using nested cross-validation (CV) and cross-site CV. Accuracies ranged from 50% to 70% for surfaced-based features; from 50% to 63% for GMV; and from 51% to 68% for VBCT. The best accuracies (70%) were achieved when DL was applied to surface-based features; however, these models generalized poorly to other sites. Findings from this study suggest that, when methodological precautions are adopted to avoid overoptimistic results, detection of individuals in the early stages of psychosis is more challenging than originally thought.
The second study investigated the extent to which combined, multimodal measures can improve the predictive accuracy of models (Lei et al., Human Brain Mapping, 2020, 41(5):1119-35). Using sMRI and resting-state functional MRI (rsfMRI) data acquired from 295 patients with schizophrenia and 452 HC, multiple features (GMV, WMV, amplitude of low-frequency fluctuation, regional homogeneity, structural covariance matrices, and functional connectivity matrices) were extracted. Results showed that higher accuracy of classification was achieved with functional data than structural data (mean 82.75% vs. 75.84%). Within each modality, the combination of images and matrices improves performance, resulting in mean accuracies of 81.63% for structural data and 87.59% for functional data, with the highest accuracy of classification (90.83%) observed when all combined structural and functional measures were used. Such findings indicate that combining multimodal measures within a single model is a promising direction for developing biologically informed diagnostic tools in schizophrenia.
In the third study (Lei et al., Psychological Medicine, 2020, 50(11):1852-1861) we tested the relative diagnostic value of whole-brain images, connectome-wide functional connectivity and graph-based metrics when applied to a sample of 295 patients with schizophrenia and 452 HC. Results showed that connectome-wide functional connectivity allowed single-subject classification of patients and controls with higher accuracy (average: 81%) than both whole-brain images (average: 53%) and graph-based metrics (average: 69%). Classification based on connectome-wide functional connectivity was driven by a distributed bilateral network including the thalamus and temporal regions. Such findings indicate that connectome-wide functional connectivity permits differentiation of patients with schizophrenia from healthy controls at single-subject level with greater accuracy. This pattern of results is consistent with the 'dysconnectivity hypothesis' of schizophrenia, which states that the neural basis of the disorder is best understood in terms of system-level functional connectivity alterations.
Development and validation of new analytical tools for future studies.
The legacy datasets provided opportunities to develop new analytic tools which we have made freely available (https://github.com/garciadias/Neuroharmony). A key challenge to the development of reliable ML models, and their translational implementation in real-word clinical practice, is the integration of datasets collected using different scanners. Current approaches for harmonizing multi-scanner data, such as using ComBat software, require statistically representative samples which are not suitable for ML models aimed at clinical translation, where the focus is on the assessment of individual scans from previously unseen scanners. To overcome this challenge, we developed a tool (‘Neuroharmony’) that is capable of harmonizing single images from unseen/unknown scanners based on a set of image quality metrics (Garcia-Dias et al., NeuroImage, 2020, 15;220: 117127). This was developed using a mega-dataset of neuroanatomical data from 15,026 healthy subjects to train a machine learning model that captures the relationship between image quality metrics and the relative volume corrections for each region of the brain prescribed by the ComBat method. The tool proved effective in reducing systematic scanner-related bias from new individual images taken from unseen scanners without requiring any specifications about the image acquisition. This approach therefore represents a significant step forward in the quest to develop reliable imaging-based clinical tools.
Work Package 3: Development of new software and data analysis
WP3 used the PSYSCAN legacy data collated in WP2 to develop state-of-the-art pre-processing pipelines to build structural and functional brain networks from MRI data. This resulted in two publications, the results/foregrounds from which are summarised below. The pipelines were subsequently employed in processing the new imaging data acquired in WP5.
Design, development and integration of pipelines for functional connectivity and network analysis of resting state fMRI data
We established pre-processing pipelines to construct and analyse functional brain networks from resting state fMRI data. Pre-processing used a wavelet despiking algorithm (Patel et la, NeuroImage, 2015) to denoise data, alongside modules for slice order correction, image co-registration, and calculating networks from pre-processed timeseries. Full details of this pipeline were previously provided in D3.2 (submitted in December 2017).
These pipelines were reported in Morgan et al., Biol Psychiatry Cogn Neurosci Neuroimaging, 2020, 8; S2451-9022(20)30138-5. This paper also showed that a measure of fMRI connectivity could be used to accurately and replicably distinguish patients with schizophrenia from healthy controls.
Design, development, and integration of pipelines for anatomical connectivity and network analysis of structural MRI and diffusion tensor imaging data
We also established pre-processing pipelines for anatomical connectivity and network analysis of structural MRI and diffusion tensor imaging (DTI) data. Pre-processing of structural MPRAGE images was based on Freesurfer’s recon-all pipeline, whilst DTI image pre-processing followed a pipeline established by van den Heuvel. Full details of these pipelines were previously provided in D3.2 (submitted in December 2017).
These pipelines were reported in Morgan et al, PNAS, 2019, 116 (19) 9604-9609. In this paper, structural brain differences in psychosis were examined using morphometric similarity mapping (Seidlitz et al 2018), which quantifies the structural similarity between brain regions. Morphometric similarity was globally reduced in patients from three independent datasets (two PSYSCAN legacy datasets, from Maastricht and Dublin, plus the openly available Cobre dataset). This suggests that patients’ brain regions were more differentiated from each other and less interconnected than in controls. Similarity was especially decreased in frontal and temporal regions, see Figure 3. This anatomical pattern was correlated with expression of genes enriched for nervous system development and synaptic signalling and genes previously associated with schizophrenia and antipsychotic treatments, see Figure 4.
The paper has been particularly influential, with 67 citations to date and a number of other researchers now applying similar approaches in their own work (see ‘Impact and Dissemination’ section below for more details).
Development of software for analysis of brain network and connectivity metrics
In the paper by Morgan et al (2020), we described how a machine learning pipeline that uses Gaussian Processes was employed to distinguish patients with schizophrenia from healthy controls. These analyses used legacy data from two PSYSCAN sites- Maastricht and Dublin, plus the publicly available Cobre dataset as a replication dataset. We found that the neuroimaging measure that provided the classification highest accuracy (AUC=0.77 in the Maastricht dataset, AUC=0.85 for Dublin and AUC=0.75 for Cobre) was fMRI connectivity (see Figure 5). Importantly, these machine learning classifiers were informed by a replicable cortical pattern of functional connectivity differences (Figure 5C). This cortical map was correlated with replicable case-control differences in functional MRI degree centrality, and with a prior cortical map of adolescent development of functional connectivity, in-line with the concept of psychosis as a disorder of brain network development. Machine learning classifiers of schizophrenia also predicted intermediate probabilities of psychosis in patients’ siblings (Figure 6); in-line with the genetic component of schizophrenia.
Application of network analysis and machine learning tools to newly-acquired PSYSCAN data
The novel methods developed in WP3 (described above) have been applied to newly-acquired clinical datasets obtained in PSYSCAN WP5. Baseline T1w and rs-fMRI data collected from PSYSCAN FEP patients have been pre-processing using our prior pipelines to obtain structural and functional brain networks. In total, N=270 T1w scans and N=220 fMRI scans passed QC requirements and were used to estimate structural similarity and fMRI brain networks for those subjects.
The structural similarity and fMRI brain networks show high connectivity in regions that would normally be expected to have high connectivity and low connectivity in regions that would normally be expected to have low connectivity, see Figure 7.
Conclusions from WP3
The aim of WP3 was to design, code, assemble, document, and test new software for the application of machine learning to measures of connectivity and networks in neuroimaging data. This aim has been achieved: we developed state-of-the-art pre-processing pipelines to build structural and functional brain networks from MRI data, and applied these to PSYSCAN data to investigate how brain connectivity is altered in psychosis. A key finding was that resting state fMRI connectivity was the most useful of a number of neuroimaging measures for distinguishing patients from healthy volunteers, with a relatively high classification accuracy (approx. 80%). This suggests that it is a strong candidate measure for classifying patients according to clinical outcomes, and it was used for this purpose in the analysis of the new patient data acquired as part of WP5 (see below).
Work Package 4: Building the PSYSCAN tools
WP4 involved the handling of WP2 and WP5 data as well as incorporation of machine learning algorithms developed in WP3 to build the prototype PSYSCAN tools. Activities in WP4 have generated the following results/foregrounds:
Development of a new electronic data capture (EDC) system for psychosis.
A core aim of WP4 was to develop an iPad-based interface for collecting clinical, demographic, and cognitive data from participants, and then use this for data collection in WP5. Developed in partnership with Cambridge Cognition, this enables researchers to conduct assessments without the need for pencil and paper, with all data entered directly onto the researcher’s iPad. These data include psychopathology rated using established scales, such as the Positive and Negative Syndrome Scale (PANSS: Kay et al, 1987) and the Comprehensive Assessment of At-Risk Mental States (CAARMS: Yung et al., 2005), and cognitive function assessed using a bespoke computerized battery comprising the four tests from the CANTAB battery most relevant to psychosis (associative learning, working memory, sustained attention and emotion recognition). This approach allows cognitive performance to be assessed in shorter time frames than traditional cognitive batteries. In addition, because testing involves participants responding to stimuli by touching the iPad screen, there is no need for a highly trained researcher. We have called this digital cognitive mini battery PsyCog, and it has already been adopted for use in other research projects.
Development of a digital healthcare data collection platform configured for psychosis
A further objective of WP4 was to develop a platform for collating both retrospective (WP2) and prospective (WP5) data from PSYSCAN centres. To achieve this objective, we worked with IXICO to modify the TrialTrackerTM system so that it could integrate clinical, cognitive, and neuroimaging data. In addition, several plugins were modified to improve the functionality of the platform. These were largely in the form of fixes for issues discovered during everyday operation of the system during the course of the project. In order to facilitate the development and provision of software interfaces for novel image analysis methodology, a Python client for downloading the data from TrialTrackerTM directly was developed and made available to researchers at the PSYSCAN co-ordinating site (King's College London). This enabled analysis and manual QC of the data. An automated email plugin service was developed to report directly to sites the results of automatic metadata QC for each series uploaded to TrialTrackerTM.
Development of a prototype tool
A prototype tool was developed, informed by the work performed on the legacy datasets. Briefly, the latter indicated that functional MRI (fMRI) data were more useful in terms of stratification than structural MRI (sMRI) or diffusion tensor imaging (DTI) data, and that within fMRI data, connectivity matrices derived from sparse inverse covariance matrix estimation were more useful than either thresholded connectivity matrices or graph- theory based summary measures. Therefore, the tool development work focused on using fMRI connectivity matrices. The latter were examined in fMRI data collected in the initial phase of WP5. In total, 253 baseline FEP scans available; however, after applying exclusions based on data quality and subject demographics, 159 scans from FEP subjects at the baseline timepoint were included in the analysis. In this sample, the majority (67.9%) were male, the mean age was 25.7 years, and the mean (± SD) total PANSS scores at baseline and 6-months were 54.9 ± 17.1 and 49.8 ± 15.5 respectively.
The tool was developed with the aim of predicting two outcomes in FEP subjects: response to treatment and remission. Both were based on the PANSS scores at a follow-up time-point. A subject was defined as in remission if a subset of 8 PANSS items (of 30 total) were all rated ≤ 3. A good response to treatment was defined as the total PANSS scores at follow-up being ≤ 80% of the total PANSS score at baseline. Using this criteria, 25.7% of FEP subjects demonstrated a response to treatment at the 6-month follow-up, although there was considerable variation across sites (range: 0 to 75%).
The comparability of fMRI data from multi-centre studies can be affected by subtle differences in scanner equipment or image acquisition across sites. To account for these potentially confounding effects, a number of site difference correction methods were evaluated. The most effective of these was found to be ComBat (combining batches), so this software was incorporated into the prototype tool.
Our original intention was to incorporate WP2 (legacy) data into referenced databases to use prospectively in WP5. WP2 data were made available for methods development but did not match the WP5 data in terms of either demographics and conditions among the subjects or in the types of imaging and non-imaging data available. Instead, the WP5 data were split (once sufficient quantities were available) into a set to be used for further methods development, and a testing set held back until the algorithms were finalised.
Integration of machine learning algorithms from WP3 with novel image analysis methodology.
The first type of machine learning algorithms developed accounted for site differences in the used imaging data and are described above. The second type of algorithm was used to develop predictive models based on the imaging data and for integration in the enrichment prototype. The algorithms that were developed and validated are described in the following section.
As noted above, our attempts to develop a tool to predict outcomes among individuals with FEP focused on using fMRI connectivity matrices. Both the outcomes we wished to predict are binary (remission/non remission and response/nonresponse) which suggested that trying to predict them by classification was the natural approach to take. However, an alternative was possible, as both remission and response status are based on follow-up PANSS scores. Thus, we could also use regression to predict the 8 PANSS items (for remission) or PANSS total score (for response) at follow-up. Then the predicted PANSS score(s) could be used to generate a predicted remission or response status by applying the criteria described previously to them (in conjunction with the ground truth baseline total PANSS for response).
Classification results: A 200-fold randomised cross validation was used, with 80% of the subjects used for training and 20% for testing in each fold. The strategy adopted was to test simple methods first. If these proved successful, we then tried to improve them by using more complex methods. Thus, the classification experiments began with an SVM with a linear kernel, which served as a basic benchmark for more advanced methods. However, none of the classification experiments produced accuracies above chance level for either outcome.
Regression results: The regression approach provided much better results than classification for predicting response to treatment, but it was still not possible to predict remission above chance level. Linear ridge regression, a simple regression method, was used as the benchmark, with the default parameters in the scikit-learn implementation (https://scikit-learn.org/stable/). Results for predicting response showed good performance even with no correction, with a balanced accuracy (BA) and Area Under the Curve (AUC) of 0.67 and 0.73 respectively. However, the small reduction in accuracy when applying ComBat (BA: 0.64; AUC: 0.72) suggested this was partly due to site effects; adding supervision to ComBat improved the results again (BA: 0.65; AUC: 0.73). We also attempted to apply the Manifold General Linear Model (MGLM) method to correct for site differences (an approach that also has the advantage of being able to regress out the effects of other nuisance variables, e.g. age and sex), but this performed very poorly (BA: 0.50; AUC: 0.38). The poor performance of the MGLM approach is likely because it completely removed linear site effects from the data, and, as site was correlated with the follow-up PANSS scores, this meant it removed all disease related effects too. We therefore propose that an equivalent of supervised ComBat for MGLM might be considered useful. The Yair method (another manifold-based correction approach) performed reasonably (BA: 0.60; AUC: 0.69) but not as well as ComBat.
Having tested these methods on the fMRI data, the same experiments were applied to anatomical covariance (sMRI) data, with the exception of the MGLM and Yair methods which do not function with non-SPD (symmetric positive definite) matrices. Despite the higher numbers of sMRI subjects (194, after applying the same exclusions as used with fMRI, except the ones regarding fMRI image quality) the accuracy was lower than for fMRI: the same accuracies were observed when no correction (BA: 0.64; AUC: 0.73) or supervised ComBat (BA: 0.64; AUC: 0.73) was applied, with ComBat performing slightly worse (BA: 0.61; AUC: 068). In order to provide a fairer comparison to fMRI, the sMRI experiments were repeated on a restricted sample (i.e. only images from individuals included in the fMRI analyses were examined in this experiment). Compared to using the full set of sMRI data, results were only slightly worse, with the AUC values for no site correction, ComBat and supervised Combat being 0.71 0.67 and 0.72 respectively. These results confirm the findings from WP3, namely, that the predictive ability of sMRI data is slightly lower than fMRI.
Using prediction of treatment response to enrich patient samples
We applied the technology described above to assess the ability to stratify or select subjects in a clinical trial based on a predicted clinical trajectory. This built on predictions of outcomes by allowing subjects with baseline data to be included or excluded from a trial by rejecting those subjects whose predicted outcomes fall above or below (depending on the trial scenario) a certain threshold. In this way it is possible to recruit for the arm of a clinical trial so that subjects will be healthier or more unwell than if they had been selected based on standard recruitment criteria only, resulting in greater statistical power and a smaller sample size. This is known as clinical trial enrichment.
We used the most promising methods identified (above) to predict the likelihood of whether an individual FEP patient would respond to future treatment. We then used this information to selectively enrich an FEP sample for individuals who are likely to respond to treatment, so that the sample size needed to detect a treatment effect was reduced. A “leave one out” cross-validation was used to maximise predictive accuracy and provide a single prediction of total PANSS at follow-up for each individual subject. This was done using stacking on fMRI data with supervised ComBat site correction. The results were sorted by the predicted total follow-up PANSS (decreasing). To calculate the resultant sample size, the top n per cent of subjects were taken by the enrichment prototype. The outcome metric was the ratio of the follow-up total PANSS score to that at baseline. We calculated the mean and SD of the PANSS ratios for these subjects (which we labelled as a group of non-responders). We assumed that the non-responders would have a mean PANSS ratio 75% that of the responders, and then calculated the sample size (per arm) needed to detect a difference between non-responders and responders with alpha=0.05 at a power of 80%. The sample size was calculated for a variety of n values, from 10 (top 10% of predicted follow-up PANSS used) to 100 (all subjects used, none rejected). The plot of sample size against n is shown below in Figure 8.
As expected, the more stringent the selection criteria (i.e. the larger the required symptom reduction for inclusion), the smaller the sample size needed to detect an effect. Although this results in the exclusion of patients who would otherwise be eligible for a study, this may be outweighed by the reduction in overall study costs, as fewer patients would have to be enrolled and followed up. An investigator or pharmaceutical company could choose inclusion criteria that provide a balance between these factors.
Work Package 5: Evaluation of PSYSCAN tools
WP5 was responsible for collecting new, homogenised data from CHR individuals and FEP patients in two naturalistic prospective studies. A healthy control (HC) group was also recruited for reference. Standardised and harmonised neuroimaging, psychopathology, demographic, and cognitive measures were collected at baseline and follow-up, with clinical and functional outcomes determined at 12-months (CHR, FEP, and HC) and 18- and 24-months (CHR). WP5 aimed to identify measures that can be used to stratify samples into subgroups with distinct clinical outcomes. A further objective was to assess whether this stratification is facilitated by using multiple measures across different modalities, as opposed to a single measure.
Recruitment and assessment of FEP participants was completed in March 2019, with a total of 328 patients. Of these, 228 were also assessed after 12-months of follow-up. A total of 239 CHR subjects were assessed at baseline. However, recruitment and follow up were delayed by the Covid pandemic, and the last 2-year follow-up assessments will be completed in March 2022. The drop-out rate in both cohorts is in line with our original predictions (~30%). Because it is necessary to follow up CHR subjects for 2 years in order to define their clinical outcomes, it has not yet been possible to examine predictors of outcomes in this cohort. To date, these analyses have focused on data from the FEP cohort, for whom the follow up data and outcomes are available. Preliminary analyses of these data (manuscript in preparation) have yielded the following results/foregrounds:
Remission and level of functioning trajectories
In 228 FEP patients, the number in symptomatic remission (defined as all PANSS items absent, minimal or mild) increased between baseline and 12-month follow-up from almost half to 67% (Figure 9). Over the same period, the proportion with a good level of functioning (defined as a SOFAS score > 60), increased from 10% to over 50%, with most of the improvement in the first two months. The extent to which baseline neuroimaging measures predict improvements in symptoms and level of functioning are currently being examine and will be reported in a forthcoming publication.
Potential Impact:
Scientific impact
One the key findings from WP2 is that among a range of neuroimaging metrics, a measure of functional connectivity (derived from fMRI data) provided the highest accuracy for distinguishing patients with schizophrenia from controls. This suggests that this metric may be particularly useful in studies designed to stratify patient groups according to clinical outcomes. A further important finding was that the combination of functional and structural measures provided a higher classification accuracy than using either modality alone. This suggests that patient stratification is likely to be more effective if multi-modal, rather than unimodal data are employed. Neuroharmony, the tool we developed for harmonizing datasets collected using different scanners and/or acquisition sequences, has now been made freely available to the scientific community (https://github.com/garciadias/Neuroharmony) and is already being used by several other research teams worldwide. This is expected to facilitate future multi-site studies.
With regards to WP3, Morgan et al, PNAS, 2019, 116 (19) 9604-9609 found a novel pattern of cortical brain differences in schizophrenia which was replicable in 3 independent case-control datasets and associated with brain expression of schizophrenia related genes. This paper has been highly cited (67 citations to date, relative citation ratio 4.44 and field citation ratio 24.19) and several researchers are now applying the same approaches and code used here to study different disorder, demonstrating the clear impact of these results. The work has also been highlighted in public facing blogs, for example a blog post from the mental health charity MQ, and an article in American magazine WIRED describing how network neuroscience could help us better understand the brain and mental health conditions (https://www.wired.com/story/a-radical-new-model-of-the-brain-illuminates-its-wiring/). Finally, the paper has been recommended on Faculty Opinions by Dr Carrie Bearden, as offering “an innovative approach for investigation of the dysconnectivity hypothesis of psychosis” (https://facultyopinions.com/prime/735586068).
Further impacts from the study are expected when the final results from WP5 have been published. These will indicate the extent to which the analytical methods developed by PSYSCAN can be used to predict whether people at CHR will subsequently develop psychosis, and whether FEP patients will respond to clinical treatment.
Dissemination
WP6 has been responsible for the dissemination of outputs from the PSYSCAN project. This has involved engagement with representatives from key-stakeholder groups and networks in the field of psychosis. This successfully raised public awareness of the project. Information about PSYSCAN was also disseminated through a dedicated website (http://psyscan.eu/) a PSYSCAN twitter account (@PSYSCANproject), regular electronic newsletters, and presentations and symposia at international conferences (e.g. ECNP: European College of Neuropsychopharmacology; SIRS: Schizophrenia International Research Society).
A PSYSCAN Stakeholder Database was created comprising information and contact details of key stakeholders active in psychiatric care across Europe, North America and Israel. This includes mental health professionals, professional associations, academic institutions, research bodies (including The European Psychiatric Association; ECNP neuroimaging network; ROAMER), patient organizations (Fundación ASAM), non-governmental organizations involved in the provision of care to patients with psychosis, as well as pharma and medical technology providers. An interactive website (www.psyscan.eu) was created which receives more than 2,100 visits per year. It includes both an open-access area and pages that are available only to registered users (e.g. study coordinators, WP leaders and PSYSCAN researchers). The publicly-accessible part of the website provides an overview of the project, with a focus on its main aims, and how these may be of benefit to patients. The website also includes a social media component, allowing for sharing content and including details on our social media accounts. Through the investigator part of the website, PSYSCAN researchers can access protocols, consent forms, assessments, and worksheets needed to conduct the project, as well as consortium documents and project reports. PSYSCAN researchers can also access online training for project assessments and procedures (e.g. acquisition and management of hair samples) and videos from previous consortium meetings.
Three newsletters have been produced during the course of the project, with the final one to be circulated in October 2021. These provide general information on the aims and scope of the project, its current status, as well as contributions from principal investigators, WP leaders, and PSYSCAN investigators. The newsletters are disseminated among all those involved in the project and to family and patient advocacy groups, health and scientific institutions and associations, charities, and other stakeholders. The newsletters can also be accessed by the general public via the PSYSCAN website (http://psyscan.eu/category/newsletter).
A Twitter account for PSYSCAN was created in October 2016. The account has more than 240 followers and has published more than 400 tweets. Information regarding the objectives, development (e.g. number of patients included, beginning of recruitment in centres), and specific goals of the project has regularly been shared, linking contents to the webpage and using direct mentions to the Twitter accounts of relevant stakeholders to increase the reach of the project to new parties. The WP6 team also tweeted links to interesting scientific publications related to psychosis, CHR individuals, and schizophrenia. Special effort has been placed in disseminating publications developed by PSYSCAN partners related to the scope of the project and in linking contents to the project webpage, in an attempt to increase the number of visitors.
Several dissemination activities have been conducted in relevant meetings such as scientific symposia at the ECNP Congress and the SIRS Congress, where different consortium members have presented the scope and several outputs from PSYSCAN (see complete list of dissemination activities). A brainstorming session with the title “What’s the role of prognostic tools in management of clinical high risk for psychosis?” was held during the ECNP Virtual Congress 2020 on September 14, 2020. More than 360 participants attended the interactive session and a group of 30 participants participated actively in the discussion process. The brainstorming session fostered discussion on the potential applications of multimodal prediction tools for clinical high-risk for psychosis, the main unmet needs they should address, and the feasibility of their practical implementation in clinical settings.
The project is likely to produce >50 scientific papers that will be published in high impact factor journals. In the past few years, the Executive Board has already reviewed and approved more than 10 publication proposals. An article presenting the scope and overall aims of the project was published in March 2020 in Schizophrenia Bulletin, a reference journal in the field of schizophrenia research (Tognin et al., Schizophrenia Bulletin, 2020, 46 (2), 432-441). Additional publications based on legacy data have been published in relevant peer-reviewed journals as outlined above. Work from WP3 has also been presented at several international conferences and meetings. A full list of 16 talks and posters in total is attached. Notable presentations include talks at the NetSci conference (Indianapolis 2017 and Paris 2018), the international Complex Networks conference held in Cambridge in 2018, the Graphs for Neuroscience conference in Paris in 2019 and the European Conference on Schizophrenia Research in Berlin in 2021. A proposal for a symposium ‘PSYSCAN: Translating neuroimaging findings from research into clinical practice – describing the first data’ has recently been submitted to the 2022 Schizophrenia International Research Society (SIRS) conference. The proposed talks will describe the baseline characteristics of the PSYSCAN CHR sample, the remission/recovery trajectories of the PSYSCAN FEP cohort, analysis of FEP imaging data, and results of a cross-validation study predicting one-year outcome among PSYSCAN and EUFEST FEP participants.
Exploitation results
The PsyCog battery developed for the PSYSCAN study has already been adopted in other ongoing projects. We anticipate that this battery is likely to be widely used in early psychosis research as administration of cognitive tasks using this tool reduces the patient and researcher burden. It is possible that Cambridge Cognition may wish to commercialise this mini-battery, but discussions about this have yet to be completed.
The potential for exploiting the PSYSCAN tools will depend on the results of their application in the datasets in WP5, which have not yet been finalised. If these tools permit the prediction of clinical outcomes in psychosis with a high degree of accuracy, we anticipate that they will be made freely available to academics, and available to industry for a fee.
List of Websites:
Project website: http://psyscan.eu/
WP1 Leader
Prof. Philip McGuire
Affiliation: Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, UK.
Email: philip.mcguire@kcl.ac.uk
WP2 Leader
Prof. Andrea Mechelli
Affiliation: Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, UK.
Email: a.mechelli@kcl.ac.uk
WP3 Leaders
Prof. Ed Bullmore
Affiliation: Department of Psychiatry, University of Cambridge, Cambridge, UK.
Email: etb23@cam.ac.uk
Prof. Michael Brammer
Affiliation: Department of Neuroimaging, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, London, UK.
Email: michael.brammer@kcl.ac.uk
WP4 Leader
Dr Robin Wolz
Affiliation: IXICO LtD, London, UK.
Email: robin.wolz@ixico.com
WP5 Leader
Prof. René Kahn
Affiliation: Division of Neurosciences, Department of Psychiatry, University Medical Center, Utrecht, The Netherlands.
Email: rene.kahn@mssm.edu
WP6 Leader
Prof. Celso Arango
Affiliation: Servicio de Psiquiatría del Niño y del Adolescente, Hospital General Universitario Gregorio Marañon, Universidad Complutense Madrid, Spain.
Email: carango@hggm.es
CHR Project Manager
Dr Stefania Tognin
Affiliation: Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, UK.
Email: stefania.tognin@kcl.ac.uk
FEP Project Manager
Dr Erika van Hell
Affiliation: University Medical Center, Utrecht, The Netherlands.
Email: H.H.vanHell-2@umcutrecht.nl
CHR Network Coordinator
Prof. Paolo Fusar-Poli
Affiliation: Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, UK.
Email: paolo.fusar-poli@kcl.ac.uk
Scientific Coordinator
Dr Alexis Cullen
Affiliation: Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, UK.
Email: alexis.cullen@kcl.ac.uk
PSYSCAN is a multi-centre research programme, involving 22 academic and industry partners across Europe, Australia, Asia, and the Americas (see Figure 1). Its aim is to build tools that will help predict clinical outcomes in patients in the early stages of psychosis by integrating clinical, cognitive, and neuroimaging data. We anticipate that these tools will allow academics and industry to stratify patient samples according to clinical outcomes.
PSYSCAN comprises 6 highly integrated work packages (WPs, see Figure 2). WP1 relates to management and scientific coordination. The main aims of the remaining WPs were to:
• WP2: Collate legacy datasets comprising neuroimaging and non-neuroimaging measures in subjects at clinical high risk (CHR) for psychosis, patients with first-episode psychosis (FEP) and healthy controls (HC), then use these datasets to evaluate novel methods of analysis that may be employed in the new tools.
• WP3: Develop new software that applies machine learning to multi-modal data to stratify patients according to clinical outcomes.
• WP4: Develop patient stratification tools that employ the technology developed in WP3.
• WP5: Conduct a large multi-centre naturalistic study, collecting longitudinal neuroimaging and non-neuroimaging data from CHR, FEP and HC groups, to provide data for developing and testing the PSYSCAN tool.
• WP6: Interact with representatives from key stakeholder groups and networks within the field of psychosis and disseminate the results of the PSYSCAN project.
The PSYSCAN project ran from February 2014 to July 2021. Despite substantial disruptions due to the Covid pandemic, it has been highly successful:
• WP2 collated 16 legacy datasets, comprising over 3000 subjects. Analyses of these have facilitated the development of machine learning algorithms used in WP3 and generated six publications.
• WP3 created state-of-the-art pre-processing pipelines to build structural and functional brain networks which have been applied to data acquired in WP5 to investigate how brain connectivity is altered in psychotic disorders.
• WP4 developed a novel iPad-based electronic data capture system that was used to collect clinical, demographic, and cognitive data in WP5 participants, and an online database (implementing automated quality control checks) for integrating legacy and newly-acquired PSYSCAN MRI data. WP4 also tested the ability of the novel analytic approaches developed in WP3 to predict clinical progression in the WP5 cohort and developed a prototype enrichment tool.
• WP5 recruited 328 FEP, 239 CHR, and 139 HC individuals, in whom clinical, neuroimaging cognitive, and genetic data were obtained. Follow-up of FEP/HC cohorts is now complete.
• WP6 established a new database containing contact information for more than 350 stakeholders among psychiatry professionals, professional associations, academic institutions and research bodies, patient organizations, and non-governmental organizations. It also created a PSYSCAN website (www.psyscan.eu) and a PSYSCAN twitter account for disseminating project findings.
Project Context and Objectives:
Context
Psychotic disorders are common, severely disabling, and associated with enormous clinical and socioeconomic impact. However, at present, we cannot predict their onset, nor their subsequent course. The onset of psychosis is preceded by attenuated psychotic symptoms and a decline in social and vocational functioning, with help-seeking individuals presenting with these features termed as being at clinical high-risk (CHR) for psychosis. Some CHR individuals will go on to develop a psychotic disorder; in others, symptoms persist without progression to frank illness, while in a third group symptoms will resolve. In those who do develop a psychosis, the subsequent course of illness is again unpredictable: some will make a good recovery, others will recover and then relapse in further episodes, and another subgroup will follow a chronic, unremitting course. The inability to predict what will happen to an individual at CHR, and to a patient who has just become psychotic, presents significant difficulties for tailoring psychiatric care to the needs of each patient.
Our inability to predict the onset and outcome of psychosis reflects a lack of objective and quantitative criteria to classify these patients. In fact, despite substantial advances in our understanding of the neurobiological basis of psychotic disorders, the assessment of patients with psychosis is still based on a clinical interview. Moreover, there is overwhelming evidence that psychotic disorders are pathophysiologically heterogeneous, which hampers the search for biomarkers to aid early diagnosis, stratification, and the measurement of disease progression. Indeed, this lack of biomarkers is perceived by industry as a key barrier to the development of novel treatments. Stratifying patients in an objective, quantitative way according to clinical and functional outcomes may generate patient clusters that are more clinically meaningful. Our industrial partners are particularly interested in the application of stratification to the recruitment of patients with established psychosis for clinical trials. It is possible that a new drug may work well in one subgroup of patients but not in another. In the absence of patient stratification, a true therapeutic effect may be undetectable when it is diluted across a heterogeneous sample.
Developing methods for stratifying individuals at CHR will is also of great importance. A key problem for both academic studies and pharmaceutical industry clinical trials is that, at present, these studies include all individuals who meet CHR criteria, even though only a small proportion (~16%) will later develop psychosis. As such, even studies with relatively large samples have limited ability test which prognostic factors or preventive treatments may reduce the risk of psychosis. Developing prediction tools will allow academics and industry to enrich CHR samples by selectively enrolling the subgroup at greatest risk of psychosis, and thus provide greater statistical power. This is attractive to academics and industry, as it reduces the size of the samples required to detect a significant prognostic or treatment effect. It would also reduce ethical concerns about giving preventive treatments to all CHR subjects when only a minority may actually need them.
Neuroimaging research has substantially advanced our understanding of the biological basis of psychotic disorders, with research in this field having the greatest potential for identifying quantitative objective measures in psychiatry. However, several factors have hampered the translation of findings from neuroimaging research to clinical, industry, and academic settings. Firstly, most neuroimaging research examines statistical differences between mean values from groups – findings which cannot be applied at the individual patient level. Secondly, the lack of standardisation of data acquisition or analysis approaches across different centres has made it difficult to combine samples to produce large scale multi-centre datasets with sufficient power to be predictive at the individual patient-level. Finally, previous research has typically examined single neuroimaging measures; however, the complex multifactorial aetiology of mental disorders (particularly psychotic disorders), indicates that that their onset and course will be best predicted by combining multiple neuroimaging and non-imaging.
To this end, the PSYSCAN project aimed to develop a neuroimaging-based tool to aid prediction and prognosis during the early course of illness and undertake activities to promote the translation of neuroimaging research findings into practice for the benefit of patients.
Primary objectives
1) To develop tools for the prediction of clinical outcomes in patients in the early phase of psychosis, with two main applications:
• PSYSCAN-Predict: Prediction of psychosis onset in CHR subjects.
• PSYSCAN-Stratify: Prediction of treatment response and illness course in FEP subjects.
2) To promote the translational application of neuroimaging in psychiatry and disseminate the results of the PSYSCAN project and the principles of using biomarkers to predict clinical outcomes in psychosis across European Member States.
To achieve these objectives, we proposed the following steps:
• Use legacy datasets from previous studies to identify neuroimaging measures associated with clinical outcomes in psychosis.
• Develop novel analytical methods that permit predictions of outcomes at the level of the individual patient, and thus have high translational potential.
• Assess whether prediction accuracy can be improved by basing this on multiple neuroimaging measures, plus clinical and cognitive measures, rather than a single neuroimaging measure alone.
• Employ outcomes that are clinically meaningful.
• Partner with imaging (IXICO) and cognitive testing (Cambridge Cognition) companies to standardise and harmonise neuroimaging data acquisition across European clinical research centres.
• Work with industrial partners (IXCO and Cambridge Cognition) to develop a novel electronic data capture system.
• Conduct multi-centre studies to acquire new datasets, using standardized protocols and a novel electronic data capturing system, and use these data to validate the PSYSCAN tools.
• Establish a network of representatives from key-stakeholder groups and networks within the field of psychosis and disseminate the tool to relevant stakeholders and end-users.
To achieve these goals, we established an international, multi-centre consortium, involving 22 academic and industry partners across Europe (London, Utrecht, Madrid, Cambridge, Maastricht, Edinburgh, Santander, Copenhagen, Galway, Naples, Amsterdam, Marburg, Tel Hashomer, Vienna Zürich, Heidelberg), Australia (Melbourne), Asia (Seoul, Hong Kong) and the Americas (Toronto, Sao Paulo).
Project Results:
Work Package 2: Legacy data
WP2 collated legacy datasets from previous studies of neuroimaging and non-neuroimaging measures in psychosis. In total, retrospective data were collected from over 3000 subjects, using 16 datasets from 11 of the consortium partners. These data comprised patients with established psychosis (n=722), patients with FEP (n=688), individuals at CHR (n=99), individuals at genetic risk for psychosis (22q11.2 deletion syndrome, n=30), unaffected siblings of patients with psychosis (n=295) and healthy controls (n=1642). The data included structural MRI, DTI, and fMRI data, as well as non-neuroimaging data (demographic and clinical data). Our analyses of these datasets generated six publications in peer-reviewed journals (see Dissemination section). The main results/foregrounds from these are summarised below:
Using neuroimaging data to enhance our understanding of the neurobiological basis of psychosis.
Structural MRI (sMRI) data were used to investigate differences in grey matter volume between healthy individuals and those with FEP (Viera et al., Psychological Medicine, 2021, 51(2): 340-350). Voxel-based morphometry was applied to sMRI data obtained from 572 FEP subjects and 502 age- and gender-comparable HC subjects. Analyses identified a widespread pattern of decreased grey matter in fronto-temporal, insular and occipital regions bilaterally among FEP individuals relative to controls; these decreases were not dependent on antipsychotic medication. The region with the most pronounced decrease - gyrus rectus - was negatively correlated with the severity of positive and negative symptoms. These findings provide evidence for reliable neuroanatomical alternations in FEP, expressed above and beyond site-related differences in antipsychotic medication, scanning parameters, and recruitment criteria.
Application of machine learning methods to neuroimaging data to develop predictive models for diagnostic and prognostic classification.
Three publications from WP2 involved the use of legacy data to evaluate machine learning (ML) methods to that could subsequently be incorporated into PSYSCAN tools.
The first of these (Vieira et al., Schizophrenia Bulletin, 2020 46(1): 17-26) investigated the extent to which the application of ML to neuroanatomical data allows detection of FEP after implementing methodological precautions to avoid overoptimistic results. This study tested both traditional ML and an emerging approach known as deep learning (DL) using 3 feature sets of interest: (1) surface-based regional volumes and cortical thickness, (2) voxel-based grey matter volume (GMV) and (3) voxel-based cortical thickness (VBCT). Analyses were repeated across 5 independent datasets (in total, 514 FEP and 444 within-site matched controls) to assess the reliability of findings with performance assessed using nested cross-validation (CV) and cross-site CV. Accuracies ranged from 50% to 70% for surfaced-based features; from 50% to 63% for GMV; and from 51% to 68% for VBCT. The best accuracies (70%) were achieved when DL was applied to surface-based features; however, these models generalized poorly to other sites. Findings from this study suggest that, when methodological precautions are adopted to avoid overoptimistic results, detection of individuals in the early stages of psychosis is more challenging than originally thought.
The second study investigated the extent to which combined, multimodal measures can improve the predictive accuracy of models (Lei et al., Human Brain Mapping, 2020, 41(5):1119-35). Using sMRI and resting-state functional MRI (rsfMRI) data acquired from 295 patients with schizophrenia and 452 HC, multiple features (GMV, WMV, amplitude of low-frequency fluctuation, regional homogeneity, structural covariance matrices, and functional connectivity matrices) were extracted. Results showed that higher accuracy of classification was achieved with functional data than structural data (mean 82.75% vs. 75.84%). Within each modality, the combination of images and matrices improves performance, resulting in mean accuracies of 81.63% for structural data and 87.59% for functional data, with the highest accuracy of classification (90.83%) observed when all combined structural and functional measures were used. Such findings indicate that combining multimodal measures within a single model is a promising direction for developing biologically informed diagnostic tools in schizophrenia.
In the third study (Lei et al., Psychological Medicine, 2020, 50(11):1852-1861) we tested the relative diagnostic value of whole-brain images, connectome-wide functional connectivity and graph-based metrics when applied to a sample of 295 patients with schizophrenia and 452 HC. Results showed that connectome-wide functional connectivity allowed single-subject classification of patients and controls with higher accuracy (average: 81%) than both whole-brain images (average: 53%) and graph-based metrics (average: 69%). Classification based on connectome-wide functional connectivity was driven by a distributed bilateral network including the thalamus and temporal regions. Such findings indicate that connectome-wide functional connectivity permits differentiation of patients with schizophrenia from healthy controls at single-subject level with greater accuracy. This pattern of results is consistent with the 'dysconnectivity hypothesis' of schizophrenia, which states that the neural basis of the disorder is best understood in terms of system-level functional connectivity alterations.
Development and validation of new analytical tools for future studies.
The legacy datasets provided opportunities to develop new analytic tools which we have made freely available (https://github.com/garciadias/Neuroharmony). A key challenge to the development of reliable ML models, and their translational implementation in real-word clinical practice, is the integration of datasets collected using different scanners. Current approaches for harmonizing multi-scanner data, such as using ComBat software, require statistically representative samples which are not suitable for ML models aimed at clinical translation, where the focus is on the assessment of individual scans from previously unseen scanners. To overcome this challenge, we developed a tool (‘Neuroharmony’) that is capable of harmonizing single images from unseen/unknown scanners based on a set of image quality metrics (Garcia-Dias et al., NeuroImage, 2020, 15;220: 117127). This was developed using a mega-dataset of neuroanatomical data from 15,026 healthy subjects to train a machine learning model that captures the relationship between image quality metrics and the relative volume corrections for each region of the brain prescribed by the ComBat method. The tool proved effective in reducing systematic scanner-related bias from new individual images taken from unseen scanners without requiring any specifications about the image acquisition. This approach therefore represents a significant step forward in the quest to develop reliable imaging-based clinical tools.
Work Package 3: Development of new software and data analysis
WP3 used the PSYSCAN legacy data collated in WP2 to develop state-of-the-art pre-processing pipelines to build structural and functional brain networks from MRI data. This resulted in two publications, the results/foregrounds from which are summarised below. The pipelines were subsequently employed in processing the new imaging data acquired in WP5.
Design, development and integration of pipelines for functional connectivity and network analysis of resting state fMRI data
We established pre-processing pipelines to construct and analyse functional brain networks from resting state fMRI data. Pre-processing used a wavelet despiking algorithm (Patel et la, NeuroImage, 2015) to denoise data, alongside modules for slice order correction, image co-registration, and calculating networks from pre-processed timeseries. Full details of this pipeline were previously provided in D3.2 (submitted in December 2017).
These pipelines were reported in Morgan et al., Biol Psychiatry Cogn Neurosci Neuroimaging, 2020, 8; S2451-9022(20)30138-5. This paper also showed that a measure of fMRI connectivity could be used to accurately and replicably distinguish patients with schizophrenia from healthy controls.
Design, development, and integration of pipelines for anatomical connectivity and network analysis of structural MRI and diffusion tensor imaging data
We also established pre-processing pipelines for anatomical connectivity and network analysis of structural MRI and diffusion tensor imaging (DTI) data. Pre-processing of structural MPRAGE images was based on Freesurfer’s recon-all pipeline, whilst DTI image pre-processing followed a pipeline established by van den Heuvel. Full details of these pipelines were previously provided in D3.2 (submitted in December 2017).
These pipelines were reported in Morgan et al, PNAS, 2019, 116 (19) 9604-9609. In this paper, structural brain differences in psychosis were examined using morphometric similarity mapping (Seidlitz et al 2018), which quantifies the structural similarity between brain regions. Morphometric similarity was globally reduced in patients from three independent datasets (two PSYSCAN legacy datasets, from Maastricht and Dublin, plus the openly available Cobre dataset). This suggests that patients’ brain regions were more differentiated from each other and less interconnected than in controls. Similarity was especially decreased in frontal and temporal regions, see Figure 3. This anatomical pattern was correlated with expression of genes enriched for nervous system development and synaptic signalling and genes previously associated with schizophrenia and antipsychotic treatments, see Figure 4.
The paper has been particularly influential, with 67 citations to date and a number of other researchers now applying similar approaches in their own work (see ‘Impact and Dissemination’ section below for more details).
Development of software for analysis of brain network and connectivity metrics
In the paper by Morgan et al (2020), we described how a machine learning pipeline that uses Gaussian Processes was employed to distinguish patients with schizophrenia from healthy controls. These analyses used legacy data from two PSYSCAN sites- Maastricht and Dublin, plus the publicly available Cobre dataset as a replication dataset. We found that the neuroimaging measure that provided the classification highest accuracy (AUC=0.77 in the Maastricht dataset, AUC=0.85 for Dublin and AUC=0.75 for Cobre) was fMRI connectivity (see Figure 5). Importantly, these machine learning classifiers were informed by a replicable cortical pattern of functional connectivity differences (Figure 5C). This cortical map was correlated with replicable case-control differences in functional MRI degree centrality, and with a prior cortical map of adolescent development of functional connectivity, in-line with the concept of psychosis as a disorder of brain network development. Machine learning classifiers of schizophrenia also predicted intermediate probabilities of psychosis in patients’ siblings (Figure 6); in-line with the genetic component of schizophrenia.
Application of network analysis and machine learning tools to newly-acquired PSYSCAN data
The novel methods developed in WP3 (described above) have been applied to newly-acquired clinical datasets obtained in PSYSCAN WP5. Baseline T1w and rs-fMRI data collected from PSYSCAN FEP patients have been pre-processing using our prior pipelines to obtain structural and functional brain networks. In total, N=270 T1w scans and N=220 fMRI scans passed QC requirements and were used to estimate structural similarity and fMRI brain networks for those subjects.
The structural similarity and fMRI brain networks show high connectivity in regions that would normally be expected to have high connectivity and low connectivity in regions that would normally be expected to have low connectivity, see Figure 7.
Conclusions from WP3
The aim of WP3 was to design, code, assemble, document, and test new software for the application of machine learning to measures of connectivity and networks in neuroimaging data. This aim has been achieved: we developed state-of-the-art pre-processing pipelines to build structural and functional brain networks from MRI data, and applied these to PSYSCAN data to investigate how brain connectivity is altered in psychosis. A key finding was that resting state fMRI connectivity was the most useful of a number of neuroimaging measures for distinguishing patients from healthy volunteers, with a relatively high classification accuracy (approx. 80%). This suggests that it is a strong candidate measure for classifying patients according to clinical outcomes, and it was used for this purpose in the analysis of the new patient data acquired as part of WP5 (see below).
Work Package 4: Building the PSYSCAN tools
WP4 involved the handling of WP2 and WP5 data as well as incorporation of machine learning algorithms developed in WP3 to build the prototype PSYSCAN tools. Activities in WP4 have generated the following results/foregrounds:
Development of a new electronic data capture (EDC) system for psychosis.
A core aim of WP4 was to develop an iPad-based interface for collecting clinical, demographic, and cognitive data from participants, and then use this for data collection in WP5. Developed in partnership with Cambridge Cognition, this enables researchers to conduct assessments without the need for pencil and paper, with all data entered directly onto the researcher’s iPad. These data include psychopathology rated using established scales, such as the Positive and Negative Syndrome Scale (PANSS: Kay et al, 1987) and the Comprehensive Assessment of At-Risk Mental States (CAARMS: Yung et al., 2005), and cognitive function assessed using a bespoke computerized battery comprising the four tests from the CANTAB battery most relevant to psychosis (associative learning, working memory, sustained attention and emotion recognition). This approach allows cognitive performance to be assessed in shorter time frames than traditional cognitive batteries. In addition, because testing involves participants responding to stimuli by touching the iPad screen, there is no need for a highly trained researcher. We have called this digital cognitive mini battery PsyCog, and it has already been adopted for use in other research projects.
Development of a digital healthcare data collection platform configured for psychosis
A further objective of WP4 was to develop a platform for collating both retrospective (WP2) and prospective (WP5) data from PSYSCAN centres. To achieve this objective, we worked with IXICO to modify the TrialTrackerTM system so that it could integrate clinical, cognitive, and neuroimaging data. In addition, several plugins were modified to improve the functionality of the platform. These were largely in the form of fixes for issues discovered during everyday operation of the system during the course of the project. In order to facilitate the development and provision of software interfaces for novel image analysis methodology, a Python client for downloading the data from TrialTrackerTM directly was developed and made available to researchers at the PSYSCAN co-ordinating site (King's College London). This enabled analysis and manual QC of the data. An automated email plugin service was developed to report directly to sites the results of automatic metadata QC for each series uploaded to TrialTrackerTM.
Development of a prototype tool
A prototype tool was developed, informed by the work performed on the legacy datasets. Briefly, the latter indicated that functional MRI (fMRI) data were more useful in terms of stratification than structural MRI (sMRI) or diffusion tensor imaging (DTI) data, and that within fMRI data, connectivity matrices derived from sparse inverse covariance matrix estimation were more useful than either thresholded connectivity matrices or graph- theory based summary measures. Therefore, the tool development work focused on using fMRI connectivity matrices. The latter were examined in fMRI data collected in the initial phase of WP5. In total, 253 baseline FEP scans available; however, after applying exclusions based on data quality and subject demographics, 159 scans from FEP subjects at the baseline timepoint were included in the analysis. In this sample, the majority (67.9%) were male, the mean age was 25.7 years, and the mean (± SD) total PANSS scores at baseline and 6-months were 54.9 ± 17.1 and 49.8 ± 15.5 respectively.
The tool was developed with the aim of predicting two outcomes in FEP subjects: response to treatment and remission. Both were based on the PANSS scores at a follow-up time-point. A subject was defined as in remission if a subset of 8 PANSS items (of 30 total) were all rated ≤ 3. A good response to treatment was defined as the total PANSS scores at follow-up being ≤ 80% of the total PANSS score at baseline. Using this criteria, 25.7% of FEP subjects demonstrated a response to treatment at the 6-month follow-up, although there was considerable variation across sites (range: 0 to 75%).
The comparability of fMRI data from multi-centre studies can be affected by subtle differences in scanner equipment or image acquisition across sites. To account for these potentially confounding effects, a number of site difference correction methods were evaluated. The most effective of these was found to be ComBat (combining batches), so this software was incorporated into the prototype tool.
Our original intention was to incorporate WP2 (legacy) data into referenced databases to use prospectively in WP5. WP2 data were made available for methods development but did not match the WP5 data in terms of either demographics and conditions among the subjects or in the types of imaging and non-imaging data available. Instead, the WP5 data were split (once sufficient quantities were available) into a set to be used for further methods development, and a testing set held back until the algorithms were finalised.
Integration of machine learning algorithms from WP3 with novel image analysis methodology.
The first type of machine learning algorithms developed accounted for site differences in the used imaging data and are described above. The second type of algorithm was used to develop predictive models based on the imaging data and for integration in the enrichment prototype. The algorithms that were developed and validated are described in the following section.
As noted above, our attempts to develop a tool to predict outcomes among individuals with FEP focused on using fMRI connectivity matrices. Both the outcomes we wished to predict are binary (remission/non remission and response/nonresponse) which suggested that trying to predict them by classification was the natural approach to take. However, an alternative was possible, as both remission and response status are based on follow-up PANSS scores. Thus, we could also use regression to predict the 8 PANSS items (for remission) or PANSS total score (for response) at follow-up. Then the predicted PANSS score(s) could be used to generate a predicted remission or response status by applying the criteria described previously to them (in conjunction with the ground truth baseline total PANSS for response).
Classification results: A 200-fold randomised cross validation was used, with 80% of the subjects used for training and 20% for testing in each fold. The strategy adopted was to test simple methods first. If these proved successful, we then tried to improve them by using more complex methods. Thus, the classification experiments began with an SVM with a linear kernel, which served as a basic benchmark for more advanced methods. However, none of the classification experiments produced accuracies above chance level for either outcome.
Regression results: The regression approach provided much better results than classification for predicting response to treatment, but it was still not possible to predict remission above chance level. Linear ridge regression, a simple regression method, was used as the benchmark, with the default parameters in the scikit-learn implementation (https://scikit-learn.org/stable/). Results for predicting response showed good performance even with no correction, with a balanced accuracy (BA) and Area Under the Curve (AUC) of 0.67 and 0.73 respectively. However, the small reduction in accuracy when applying ComBat (BA: 0.64; AUC: 0.72) suggested this was partly due to site effects; adding supervision to ComBat improved the results again (BA: 0.65; AUC: 0.73). We also attempted to apply the Manifold General Linear Model (MGLM) method to correct for site differences (an approach that also has the advantage of being able to regress out the effects of other nuisance variables, e.g. age and sex), but this performed very poorly (BA: 0.50; AUC: 0.38). The poor performance of the MGLM approach is likely because it completely removed linear site effects from the data, and, as site was correlated with the follow-up PANSS scores, this meant it removed all disease related effects too. We therefore propose that an equivalent of supervised ComBat for MGLM might be considered useful. The Yair method (another manifold-based correction approach) performed reasonably (BA: 0.60; AUC: 0.69) but not as well as ComBat.
Having tested these methods on the fMRI data, the same experiments were applied to anatomical covariance (sMRI) data, with the exception of the MGLM and Yair methods which do not function with non-SPD (symmetric positive definite) matrices. Despite the higher numbers of sMRI subjects (194, after applying the same exclusions as used with fMRI, except the ones regarding fMRI image quality) the accuracy was lower than for fMRI: the same accuracies were observed when no correction (BA: 0.64; AUC: 0.73) or supervised ComBat (BA: 0.64; AUC: 0.73) was applied, with ComBat performing slightly worse (BA: 0.61; AUC: 068). In order to provide a fairer comparison to fMRI, the sMRI experiments were repeated on a restricted sample (i.e. only images from individuals included in the fMRI analyses were examined in this experiment). Compared to using the full set of sMRI data, results were only slightly worse, with the AUC values for no site correction, ComBat and supervised Combat being 0.71 0.67 and 0.72 respectively. These results confirm the findings from WP3, namely, that the predictive ability of sMRI data is slightly lower than fMRI.
Using prediction of treatment response to enrich patient samples
We applied the technology described above to assess the ability to stratify or select subjects in a clinical trial based on a predicted clinical trajectory. This built on predictions of outcomes by allowing subjects with baseline data to be included or excluded from a trial by rejecting those subjects whose predicted outcomes fall above or below (depending on the trial scenario) a certain threshold. In this way it is possible to recruit for the arm of a clinical trial so that subjects will be healthier or more unwell than if they had been selected based on standard recruitment criteria only, resulting in greater statistical power and a smaller sample size. This is known as clinical trial enrichment.
We used the most promising methods identified (above) to predict the likelihood of whether an individual FEP patient would respond to future treatment. We then used this information to selectively enrich an FEP sample for individuals who are likely to respond to treatment, so that the sample size needed to detect a treatment effect was reduced. A “leave one out” cross-validation was used to maximise predictive accuracy and provide a single prediction of total PANSS at follow-up for each individual subject. This was done using stacking on fMRI data with supervised ComBat site correction. The results were sorted by the predicted total follow-up PANSS (decreasing). To calculate the resultant sample size, the top n per cent of subjects were taken by the enrichment prototype. The outcome metric was the ratio of the follow-up total PANSS score to that at baseline. We calculated the mean and SD of the PANSS ratios for these subjects (which we labelled as a group of non-responders). We assumed that the non-responders would have a mean PANSS ratio 75% that of the responders, and then calculated the sample size (per arm) needed to detect a difference between non-responders and responders with alpha=0.05 at a power of 80%. The sample size was calculated for a variety of n values, from 10 (top 10% of predicted follow-up PANSS used) to 100 (all subjects used, none rejected). The plot of sample size against n is shown below in Figure 8.
As expected, the more stringent the selection criteria (i.e. the larger the required symptom reduction for inclusion), the smaller the sample size needed to detect an effect. Although this results in the exclusion of patients who would otherwise be eligible for a study, this may be outweighed by the reduction in overall study costs, as fewer patients would have to be enrolled and followed up. An investigator or pharmaceutical company could choose inclusion criteria that provide a balance between these factors.
Work Package 5: Evaluation of PSYSCAN tools
WP5 was responsible for collecting new, homogenised data from CHR individuals and FEP patients in two naturalistic prospective studies. A healthy control (HC) group was also recruited for reference. Standardised and harmonised neuroimaging, psychopathology, demographic, and cognitive measures were collected at baseline and follow-up, with clinical and functional outcomes determined at 12-months (CHR, FEP, and HC) and 18- and 24-months (CHR). WP5 aimed to identify measures that can be used to stratify samples into subgroups with distinct clinical outcomes. A further objective was to assess whether this stratification is facilitated by using multiple measures across different modalities, as opposed to a single measure.
Recruitment and assessment of FEP participants was completed in March 2019, with a total of 328 patients. Of these, 228 were also assessed after 12-months of follow-up. A total of 239 CHR subjects were assessed at baseline. However, recruitment and follow up were delayed by the Covid pandemic, and the last 2-year follow-up assessments will be completed in March 2022. The drop-out rate in both cohorts is in line with our original predictions (~30%). Because it is necessary to follow up CHR subjects for 2 years in order to define their clinical outcomes, it has not yet been possible to examine predictors of outcomes in this cohort. To date, these analyses have focused on data from the FEP cohort, for whom the follow up data and outcomes are available. Preliminary analyses of these data (manuscript in preparation) have yielded the following results/foregrounds:
Remission and level of functioning trajectories
In 228 FEP patients, the number in symptomatic remission (defined as all PANSS items absent, minimal or mild) increased between baseline and 12-month follow-up from almost half to 67% (Figure 9). Over the same period, the proportion with a good level of functioning (defined as a SOFAS score > 60), increased from 10% to over 50%, with most of the improvement in the first two months. The extent to which baseline neuroimaging measures predict improvements in symptoms and level of functioning are currently being examine and will be reported in a forthcoming publication.
Potential Impact:
Scientific impact
One the key findings from WP2 is that among a range of neuroimaging metrics, a measure of functional connectivity (derived from fMRI data) provided the highest accuracy for distinguishing patients with schizophrenia from controls. This suggests that this metric may be particularly useful in studies designed to stratify patient groups according to clinical outcomes. A further important finding was that the combination of functional and structural measures provided a higher classification accuracy than using either modality alone. This suggests that patient stratification is likely to be more effective if multi-modal, rather than unimodal data are employed. Neuroharmony, the tool we developed for harmonizing datasets collected using different scanners and/or acquisition sequences, has now been made freely available to the scientific community (https://github.com/garciadias/Neuroharmony) and is already being used by several other research teams worldwide. This is expected to facilitate future multi-site studies.
With regards to WP3, Morgan et al, PNAS, 2019, 116 (19) 9604-9609 found a novel pattern of cortical brain differences in schizophrenia which was replicable in 3 independent case-control datasets and associated with brain expression of schizophrenia related genes. This paper has been highly cited (67 citations to date, relative citation ratio 4.44 and field citation ratio 24.19) and several researchers are now applying the same approaches and code used here to study different disorder, demonstrating the clear impact of these results. The work has also been highlighted in public facing blogs, for example a blog post from the mental health charity MQ, and an article in American magazine WIRED describing how network neuroscience could help us better understand the brain and mental health conditions (https://www.wired.com/story/a-radical-new-model-of-the-brain-illuminates-its-wiring/). Finally, the paper has been recommended on Faculty Opinions by Dr Carrie Bearden, as offering “an innovative approach for investigation of the dysconnectivity hypothesis of psychosis” (https://facultyopinions.com/prime/735586068).
Further impacts from the study are expected when the final results from WP5 have been published. These will indicate the extent to which the analytical methods developed by PSYSCAN can be used to predict whether people at CHR will subsequently develop psychosis, and whether FEP patients will respond to clinical treatment.
Dissemination
WP6 has been responsible for the dissemination of outputs from the PSYSCAN project. This has involved engagement with representatives from key-stakeholder groups and networks in the field of psychosis. This successfully raised public awareness of the project. Information about PSYSCAN was also disseminated through a dedicated website (http://psyscan.eu/) a PSYSCAN twitter account (@PSYSCANproject), regular electronic newsletters, and presentations and symposia at international conferences (e.g. ECNP: European College of Neuropsychopharmacology; SIRS: Schizophrenia International Research Society).
A PSYSCAN Stakeholder Database was created comprising information and contact details of key stakeholders active in psychiatric care across Europe, North America and Israel. This includes mental health professionals, professional associations, academic institutions, research bodies (including The European Psychiatric Association; ECNP neuroimaging network; ROAMER), patient organizations (Fundación ASAM), non-governmental organizations involved in the provision of care to patients with psychosis, as well as pharma and medical technology providers. An interactive website (www.psyscan.eu) was created which receives more than 2,100 visits per year. It includes both an open-access area and pages that are available only to registered users (e.g. study coordinators, WP leaders and PSYSCAN researchers). The publicly-accessible part of the website provides an overview of the project, with a focus on its main aims, and how these may be of benefit to patients. The website also includes a social media component, allowing for sharing content and including details on our social media accounts. Through the investigator part of the website, PSYSCAN researchers can access protocols, consent forms, assessments, and worksheets needed to conduct the project, as well as consortium documents and project reports. PSYSCAN researchers can also access online training for project assessments and procedures (e.g. acquisition and management of hair samples) and videos from previous consortium meetings.
Three newsletters have been produced during the course of the project, with the final one to be circulated in October 2021. These provide general information on the aims and scope of the project, its current status, as well as contributions from principal investigators, WP leaders, and PSYSCAN investigators. The newsletters are disseminated among all those involved in the project and to family and patient advocacy groups, health and scientific institutions and associations, charities, and other stakeholders. The newsletters can also be accessed by the general public via the PSYSCAN website (http://psyscan.eu/category/newsletter).
A Twitter account for PSYSCAN was created in October 2016. The account has more than 240 followers and has published more than 400 tweets. Information regarding the objectives, development (e.g. number of patients included, beginning of recruitment in centres), and specific goals of the project has regularly been shared, linking contents to the webpage and using direct mentions to the Twitter accounts of relevant stakeholders to increase the reach of the project to new parties. The WP6 team also tweeted links to interesting scientific publications related to psychosis, CHR individuals, and schizophrenia. Special effort has been placed in disseminating publications developed by PSYSCAN partners related to the scope of the project and in linking contents to the project webpage, in an attempt to increase the number of visitors.
Several dissemination activities have been conducted in relevant meetings such as scientific symposia at the ECNP Congress and the SIRS Congress, where different consortium members have presented the scope and several outputs from PSYSCAN (see complete list of dissemination activities). A brainstorming session with the title “What’s the role of prognostic tools in management of clinical high risk for psychosis?” was held during the ECNP Virtual Congress 2020 on September 14, 2020. More than 360 participants attended the interactive session and a group of 30 participants participated actively in the discussion process. The brainstorming session fostered discussion on the potential applications of multimodal prediction tools for clinical high-risk for psychosis, the main unmet needs they should address, and the feasibility of their practical implementation in clinical settings.
The project is likely to produce >50 scientific papers that will be published in high impact factor journals. In the past few years, the Executive Board has already reviewed and approved more than 10 publication proposals. An article presenting the scope and overall aims of the project was published in March 2020 in Schizophrenia Bulletin, a reference journal in the field of schizophrenia research (Tognin et al., Schizophrenia Bulletin, 2020, 46 (2), 432-441). Additional publications based on legacy data have been published in relevant peer-reviewed journals as outlined above. Work from WP3 has also been presented at several international conferences and meetings. A full list of 16 talks and posters in total is attached. Notable presentations include talks at the NetSci conference (Indianapolis 2017 and Paris 2018), the international Complex Networks conference held in Cambridge in 2018, the Graphs for Neuroscience conference in Paris in 2019 and the European Conference on Schizophrenia Research in Berlin in 2021. A proposal for a symposium ‘PSYSCAN: Translating neuroimaging findings from research into clinical practice – describing the first data’ has recently been submitted to the 2022 Schizophrenia International Research Society (SIRS) conference. The proposed talks will describe the baseline characteristics of the PSYSCAN CHR sample, the remission/recovery trajectories of the PSYSCAN FEP cohort, analysis of FEP imaging data, and results of a cross-validation study predicting one-year outcome among PSYSCAN and EUFEST FEP participants.
Exploitation results
The PsyCog battery developed for the PSYSCAN study has already been adopted in other ongoing projects. We anticipate that this battery is likely to be widely used in early psychosis research as administration of cognitive tasks using this tool reduces the patient and researcher burden. It is possible that Cambridge Cognition may wish to commercialise this mini-battery, but discussions about this have yet to be completed.
The potential for exploiting the PSYSCAN tools will depend on the results of their application in the datasets in WP5, which have not yet been finalised. If these tools permit the prediction of clinical outcomes in psychosis with a high degree of accuracy, we anticipate that they will be made freely available to academics, and available to industry for a fee.
List of Websites:
Project website: http://psyscan.eu/
WP1 Leader
Prof. Philip McGuire
Affiliation: Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, UK.
Email: philip.mcguire@kcl.ac.uk
WP2 Leader
Prof. Andrea Mechelli
Affiliation: Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, UK.
Email: a.mechelli@kcl.ac.uk
WP3 Leaders
Prof. Ed Bullmore
Affiliation: Department of Psychiatry, University of Cambridge, Cambridge, UK.
Email: etb23@cam.ac.uk
Prof. Michael Brammer
Affiliation: Department of Neuroimaging, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, London, UK.
Email: michael.brammer@kcl.ac.uk
WP4 Leader
Dr Robin Wolz
Affiliation: IXICO LtD, London, UK.
Email: robin.wolz@ixico.com
WP5 Leader
Prof. René Kahn
Affiliation: Division of Neurosciences, Department of Psychiatry, University Medical Center, Utrecht, The Netherlands.
Email: rene.kahn@mssm.edu
WP6 Leader
Prof. Celso Arango
Affiliation: Servicio de Psiquiatría del Niño y del Adolescente, Hospital General Universitario Gregorio Marañon, Universidad Complutense Madrid, Spain.
Email: carango@hggm.es
CHR Project Manager
Dr Stefania Tognin
Affiliation: Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, UK.
Email: stefania.tognin@kcl.ac.uk
FEP Project Manager
Dr Erika van Hell
Affiliation: University Medical Center, Utrecht, The Netherlands.
Email: H.H.vanHell-2@umcutrecht.nl
CHR Network Coordinator
Prof. Paolo Fusar-Poli
Affiliation: Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, UK.
Email: paolo.fusar-poli@kcl.ac.uk
Scientific Coordinator
Dr Alexis Cullen
Affiliation: Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, UK.
Email: alexis.cullen@kcl.ac.uk
