O5. Classification of Schizophrenia Using Machine Learning With Multimodal Markers

Background: Schizophrenia risk is associated with genetic variation and with early life environmental factors. Moreover, cognitive abnormalities are key features of this disorder. Our aim was to use multimodal machine learning to assess the performance of an ensemble of genetic, environmental and cognitive variables on schizophrenia classification. Methods: 337 healthy controls (HC) and 103 schizophrenia patients (SCZ) underwent a full neuropsychological evaluation, a broad environmental assessment and a genetic risk estimation with polygenic risk scores computation from the Psychiatric Genomics Consortium (PGC2) study. Scores from these three data modalities entered a Support Vector Machine algorithm aimed at classifying HC vs. SCZ. Specifically, we applied decision-based data fusion strategies to integrate the individual predictions of each modality in a nested cross-validation framework. Results: Results revealed that cognitive features classified SCZ with the highest Balanced Accuracy (BAC, 88.7%) and that the most selected cognitive indices were intelligence quotient and attention. Environmental features classified SCZ with a 65.1% BAC, and the most predictive indices were parental socio-economic status and presence of developmental anomalies. Genetic features discriminated HC from SCZ only at 55.5% BAC. Late fusion combining decision scores from individual modalities classified SCZ with a 77.5% BAC. Discussion: Our results suggest that an ensemble of cognitive, environmental and genetic features can classify SCZ with high accuracy and offer insights on cognitive and environmental factors that can be targeted in early identification programs. However, the near chance-level classification ability of the genetic modality alone calls for the implementation of more complex models of interaction between multiple risk factors