12.4 CAN MACHINE LEARNING HELP IDENTIFY CHILDHOOD RISK INDICATORS OF FUTURE DEVELOPMENT OF BIPOLAR DISORDER? RESULTS FROM A 10-YEAR LONGITUDINAL FOLLOW-UP STUDY

Abstract

The main aim of the present study was to assess whether machine learning can empirically identify early predictors of future bipolar disorder outcomes based on presenting childhood characteristics and identifying the important factors for that prediction. We used the extra trees algorithm with repeated sampling cross-validation to ascertain generalizability of the model (sensitivity and specificity) to unseen data and to ascertain a distribution of model performance. We also identified important features in the classification. The final sample consisted of 492 children, 52% male, ranging in age—at their first evaluation—from 6 to 19 years (μ = 11). The sensitivity of our model—the rate at which it correctly predicted that a subject would develop bipolar disorder—was 70%. Its specificity was found to be 68%. We also determined that childhood features were particularly good at differentiating between patients who developed bipolar disorder and patients who did not, including the School Adjustment Inventory for Children and Adolescents (SAICA) school behavior problems, CBCL rule-breaking t score, CBCL attention problems t score, socioeconomic status (SES), and CBCL externalizing t score. We applied machine learning prediction models to longitudinal data and successfully predicted the development of bipolar disorder in children observed at baseline and followed up to 10 years to their late adolescence and young adulthood. The finding that childhood school and social behavior problems were predictive of the future development of bipolar disorder are consistent with the literature.