M110. PREDICTING RISK OF PSYCHOSIS IN A GENERAL POPULATION SAMPLE. (18th May 2020)

• Record Type: Journal Article
• Authors: Kounali, Daphne; Sullivan, Sarah; Heron, Jon; McLeod, Jon; Cannon, Mary; Cotter, David; Mongan, David; Lewis, Glyn; Jones, Peter; Zammit, Stanley
• Is Part Of: Schizophrenia bulletin. Volume 46(2020)Supplement 1
• Page Start: S176
• Page End: S177
• Publication Date: 2020-05-18
• Subjects: Schizophrenia -- Periodicals; Schizophrenia -- Research -- Periodicals; 616.898005
• DOI: 10.1093/schbul/sbaa030.422
• Languages: English
• ISSNs: 0586-7614

Abstract

Abstract: Background: At present clinical high-risk states for psychosis are determined by specialist mental health services using clinical tools such as the CAARMS, which largely rely on the detection of attenuated psychotic symptoms. However, the positive predictive value (PPV) of the CAARMS for transition to psychosis is only 25% in help seeking populations and as low as 5% in general population, non-help seeking samples. There is therefore a clear need to improve the prediction of psychotic disorder using other (non-symptom related) markers of risk. Our aim was to derive a risk prediction tool to determine risk of psychotic disorder using a large, population-based birth-cohort. Methods: We used data from the ALSPAC birth cohort, with data on a comprehensive range of predictors ascertained from early childhood through early adulthood, and on psychotic disorder up to age 24 (imputed up to N≈7000 with any data on psychotic experiences). We use a two-stage risk prediction model, where different sets of predictors are used for outcomes of increasing severity. In the first stage, we predicted a clinical need for care in those who had self-reported psychotic-like experiences prior to age 17 years. We assumed that most of this need for care subsample would be help-seeking and that they therefore provide a more accessible risk-enriched sample for the second stage of our prediction model, where more difficult to measure predictors are included for estimating the risk of new onset psychotic disorder. Here we report on the first stage of our prediction model where we predict a clinical need for care (defined as presence of frequent and distressing interviewer-rated psychotic experiences) at age 17–24 years in participants who self-reported any psychotic-like experiences prior to age 17. The prediction set consisted of sixty-one features across 4 domains: socio-demographic (12 features); cognitive (10 features); non-psychotic psychopathology (24 features); behavioural (10 features). We used machine-learning methods for predictor selection and model fitting, employing resampling to assess and validate model calibration and discrimination. Results: 13% of participants who self-reported psychotic experiences by age 17 were found to have a clinical need for care between ages 17–24, and 3.5% met criteria for newly ascertained psychotic disorder at 24 years. Use of two different machine learning methods for feature selection (random forests with a 10-fold cross-validation and elastic nets employing shrinkage) yielded similar results, although the elastic nets/ridge regression produced a more parsimonious model. The features selected included: adolescent self-harm, and childhood IQ, attention, processing speed and external locus of control. The AUC reduced very little compared to that of a model with 61 characteristics. This simpler model showed improved calibration and optimism-corrected predictive performance of AUC=0.73, sensitivity=0.75, specificity=0.60, and PPV=0.22. Discussion: Our risk calculator is comparable in performance to those produced in studies of prodromal psychosis in high-risk samples. This first-stage model achieved promising predictive performance. We are currently developing a prognostic score for psychotic disorder in those with a clinical need for care and augment the predictor set with genetic, lipidomic and proteomic markers and further cognitive tests. We will then assess the model's clinical utility and variation in predictive performance using linkage of ALSPAC data to clinical health care records with the aim to externally validate in other cohort studies.