S64. COGNITIVE IMPAIRMENTS AND PREDICTION OF FUNCTIONAL OUTCOME IN INDIVIDUALS AT CLINICAL HIGH-RISK FOR PSYCHOSIS

Background Research in individuals at clinical-high risk for psychosis (CHR-P) has focused on developing algorithms to predict transition to psychosis. However, it is becoming increasingly important to address other outcomes, such as the level of functioning of CHR-P participants. To address this important question, this study investigated the relationship between baseline cognitive performance and functional outcome between 6–12 months in a sample of CHR-P individuals using a machine-learning approach to identify features that are predictive of long-term functional impairments. Methods Data was available for 111 CHR-P individuals at 6–12 months follow-up. In addition, 47 CHR-negative (CHR-N) participants who did not meet CHR criteria and 55 healthy controls (HCs) were recruited. CHR-P status was assessed using the Comprehensive Assessment of At-Risk Mental States (CAARMS) and the Schizophrenia Proneness Instrument, Adult version (SPI-A). Cognitive assessments included the Brief Assessment of Cognition in Schizophrenia (BACS) and the Penn Computerized Neurocognitive Battery (CNB). Global, social and role functioning scales were used to measure functional status. CHR-P individuals were divided into good functional outcome (GFO, GAF ≥ 65) and poor functional outcome groups (PFO, GAF < 65). Feature selection was performed using LASSO regression with the LARS algorithm and 10-fold cross validation with GAF scores at baseline as the outcome variable. The following features were identified as predictors of GAF scores at baseline: verbal memory, verbal fluency, attention, emotion recognition, social and role functioning and SPI-A distress. This model explained 47% of the variance in baseline GAF scores. In the next step, Support Vector Machines (SVM), Linear Discriminant Analysis (LDA), Logistic Regression (LR), Gaussian Naïve Bayes (GNB), and Random Forest (RF) classifiers with 10-fold cross validation were then trained on those features with GAF category at follow-up used as the binary label column. Models were compared using a calculated score incorporating area under the curve (AUC), accuracy, and AUC consistency across runs, whereby AUC was given a higher weighting than accuracy due to class imbalance. Results CHR-P individuals had slower motor speed, reduced attention and processing speed and increased emotion recognition reaction times (RTs) compared to HCs and reduced attention and processing speed compared to CHR-Ns. At follow-up, 66% of CHR-P individuals had PFO. LDA emerged as the strongest classifier, showing a mean AUC of 0.75 (SD = 0.15), indicating acceptable classification performance for GAF category at follow-up. PFO was detected with a sensitivity of 75% and specificity of 58%, with a total mean weighted accuracy of 68%. Discussion The CHR-P state was associated with significant impairments in cognition, highlighting the importance of interventions such as cognitive remediation in this population. Our data suggest that the development of features using machine learning approaches is effective in predicting functional outcomes in CHR-P individuals. Greater levels of accuracy, sensitivity and specificity might be achieved by increasing training sets and validating the classifier with external data sets. Indeed, machine learning methods have potential given that trained classifiers can easily be shared online, thus enabling clinical professionals to make individualised predictions.