Onset probability prediction of schizophrenia based on a multimodal approach

Abstract

Psychosis is a severe mental condition characterized by a complex set of disturbances of thinking, perception, affect and social behaviour. It is usually preceded by a prodromal phase lasting months to years and in which patients are clinically identified has being ‘At Risk Mental State’ (ARMS). Retrospective studies have showed that ARMS individuals have a 30% risk of transition to psychosis within the first 2 years after presentation to clinical services. Moreover, several neuroimaging, genetic and environmental biomarkers have been independently associated with the onset of psychosis in the ARMS. However, at present there is yet no established method for predicting which individuals will develop the illness and which will not – which would allow cost-efficient targeting of early intervention therapies. Furthermore, a few studies have demonstrated the feasibility to predict psychosis transition from an ARMS using structural magnetic resonance imaging (sMRI) data and machine learning (ML) methods. However, the reliability of these findings is unclear due to possible sampling bias. Moreover, the value of genetic and environmental data in predicting transition to psychosis from an ARMS is yet to be explored. In this study I aimed at predicting transition to psychosis from an ARMS using ML and quantitative data – neuroimaging, genetics, and environment – as predictors. I used several samples (one for each modality – neuroimaging, genetics or environment) drawn from a pool of 246 subjects identified as being at an ARMS when they first sought clinical help (i.e. at baseline). Subjects were clinically identified as transitioned to psychosis (ARMS-T, 60 subjects) if they later presented a first episode of psychosis (FEP) or as nottransitioned to psychosis (ARMS-NT, 186 subjects) if they did not present a FEP within at least a period of 2 years. Structural magnetic resonance imaging, genome-wide genotypes and environmental risk assessment data was collected from the ARMS subjects at baseline. Then, the modality-specific value in predicting transition to psychosis was evaluated using a) several feature types [regional and voxel-based grey matter and white matter volumes, and regional cortical thickness, and brain gyrification, sulci depth and complexity indexes (neuroimaging); a polygenic risk score (PRS) for schizophrenia, a list of psychosisassociated single nucleotide polymorphisms (SNP), and a list of psychosis-associated genes for which several brain tissue-specific expression quantitative trait loci (eQTL) scores were extracted (genetics); and an environmental risk score (ERS) for schizophrenia, and a list of environmental risks factors (environment)], b) several feature manipulation strategies [feature dimensionality reduction through principal component analysis, no feature selection, and forward feature selection (neuroimaging), and embedded feature selection (genetics and environment)], c) several ML algorithms [linear support vector machines (neuroimaging), elastic-net and simple logistic regression (genetics and environment)], d) several cross-validation (CV) strategies [5-fold CV and leave-one scanning acquisition protocol-out (neuroimaging) and leave-one per group, i.e. 1 ARMS-T and 1 ARMS-NT,-out (neuroimaging, genetics and environment)], e) sample balancing, i.e. same number of ARMS-T and ARMS-NT subjects, and f) bootstrapping, i.e. 5 (neuroimaging) or 100 (genetics and environment) semi-random subsamples drawn from the original pool. Then, only the modalities whose classification models showed a balanced accuracy across bootstrapped samples statistically better than chance level were included in a multimodal classification model. Overall, this study’s results showed that only genetics, and when using a set of psychosisassociated SNPs, could predict the transition to psychosis from an ARMS marginally better than chance, albeit with no clinical significance, (balanced accuracy = 53%, diagnostic odds ratio = 3.3 – averaged across bootstrapped samples). Furthermore, the environmental and neuroimaging alone could not predict psychosis from an ARMS, statistically better than chance. Therefore, no multimodal classification model was trained/tested. Moreover, and unexpectedly, I could not replicate previous findings showing the usefulness of structural MRI in predicting transition to psychosis from an ARMS using ML. Therefore, my results suggest that: a) genetic data may be promising for predicting transition to psychosis from an ARMS; and b) the value of structural MRI data in predicting psychosis from an ARMS, as suggested by previous evidence, should be reconsidered. Finally, this study serves as a proofof-concept on how multimodal quantitative data can be used to predict psychosis development already from a prodromal stage and should be replicated in larger ARMs samples.