LAUSR

Abstract Background Precise prognosis of clinical outcomes in individuals at clinical high-risk (CHR) of developing psychosis is imperative to guide treatment selection. While much effort has been put into the prediction of transition to psychosis in CHR individuals, prognostic models focusing on negative symptom progression in this population are widely missing. This is a major oversight bearing in mind that 82% of CHR individuals exhibit at least one negative symptom in the moderate to severe range at first clinical presentation, whereas 54% still meet this criteria after 12 months. Negative symptoms are strong predictors of poor functional outcome irrespective of other symptoms such as depression or anxiety. Prognostic tools are therefore urgently required to track negative symptom progression in CHR individuals in order to guide early personalized interventions. Here, we applied machine-learning to multi-site data from five European countries with the aim of predicting negative symptoms of at least moderate severity 9-month after study inclusion. Methods We analyzed data from the ‘Personalized Prognostic Tools for Early Psychosis Management’ (PRONIA; www.pronia.eu) study, which consisted of 94 individuals at clinical high-risk of developing psychosis (CHR). Predictive models either included baseline level of negative symptoms, measured with the Structured Interview for Prodromal Syndromes, whole-brain gyrification pattern, or both to forecast negative symptoms of moderate severity or above in CHR individuals. Using data from the clinical and gyrification model, further sequential testing simulations were conducted to stratify CHR individuals into different risk groups. Lastly, we assessed the models’ ability to predict functional outcomes in CHR individuals. Results Baseline negative symptom severity alone predicted moderate to severe negative symptoms with a balanced accuracy (BAC) of 68%, whereas predictive models trained on gyrification measures achieved a BAC of 64%. Stacking the two modalities allowed for an increased BAC of 72%. Additional sequential testing simulations suggested, that CHR patients could be stratified into a high risk group with 83% probability of experiencing at least moderate negative symptoms at follow-up and a medium/low risk group with a risk ranging from 25 to 38%, when using the two models sequentially. Furthermore, the models trained to predict negative symptom severity from baseline symptoms were less predictive of role (60% BAC) and social (62% BAC) functioning at follow-up. However, the model trained on gyrification data also predicted role (74% BAC) and social (73% BAC) functioning later on. The stacking model predicted role, and social functioning with 64% BAC and 66% BAC respectively. Discussion To the best of our knowledge this is the first study using state-of-the-art predictive modelling to prospectively identify CHR subjects with negative symptoms in the moderate to above moderate severity range who potentially require further therapeutic consideration. While the predictive performance will need to be validated in other samples and may be improved by expanding the models with additional predictors, we believe that this pragmatic approach will help to stratify individual risk profiles and optimize personal interventions in the future.