Exploring the relationship between cognition and symptomatology among patients with schizophrenia and individuals at risk for psychosis: a machine learning study

Introduction: Besides positive and negative symptoms, many patients with schizophrenia (SCZ) show neuro- and socio-cognitive deficits to different extents. However, the way in which cognitive deficits might contribute to the development of more severe symptoms in SCZ or conditions that might predate schizophrenia has been poorly investigated only via univariate methods, which flatten deliver only group- level findings. Employing Machine Learning (ML), this study aims to identify classification signatures of symptom levels based on neuro- and socio-cognitive deficits in SCZ and to investigate their potential prognostic relevance in At Risk Mental State (ARMS) and First-Episode Psychosis (FEP) individuals. Method: We recruited 91 SCZ (discovery sample, DS), 19 ARMS and 17 FEP (validation sample, VS). We divided SCZ in two groups (higher/lower) based on the severity of Positive, Negative, General and Total scales of the Positive and Negative Syndrome Scale (PANSS). For each scale, we trained 3 ML algorithms in the DS aimed at discriminating SCZ with higher/lower symptoms based on (i) 45 neurocognitive variables, (ii) 34 socio-cognitive variables, and (iii) the combination of (i) and (ii) decisions. Algorithms reaching more than 60% Balanced Accuracy (BAC) and permuted significance (all p<0.05) were applied to the VS to predict symptoms 12 months after baseline. Results: Three models were significant: 1) Neuro-cognitive variables classified SCZ with higher/lower PANSS total symptoms (BAC=61.7%). Decisions were most influenced by semantic fluency. 2) Socio- cognitive variables discriminated SCZ with higher/lower PANSS positive (BAC=62.2%) and 3) with higher/lower PANSS negative symptoms (BAC=63.7%). Social inference abilities in sarcastic contexts were the most relevant variable. The socio-cognitive algorithm successfully predicted follow-up negative symptoms in ARMS (BAC=69.4%) and FEP (BAC=76.4%). The other algorithms did not validate in any of the VS populations. Conclusion: We identified specific socio-cognitive alterations that specifically impact levels of negative symptoms in schizophrenia via ML; the generalizability of this ML algorithm to VS populations offers preliminary evidence of its prognostic relevance, suggesting that these specific socio-cognitive variables may be relevant for the symptom course also during conditions that may predate schizophrenia.