Early Detection of Bipolar Disorder Four Years Before Onset in a 22-Year Population-Based Birth Cohort Using Advanced Machine Learning Techniques

Abstract

Background: Bipolar disorder (BD) is the sixth leading cause of disability worldwide. Early detection of BD may help clinicians to intervene early and to prevent illness progression. We aimed to build a BD prediction model using machine learning techniques in a large population-based birth cohort at the individual level. Methods: A total of 3,748 subjects were studied from birth up to the age of 22 years in a prospective population-based birth cohort. We used the Elastic Net algorithm with 10-fold cross-validation to predict who would develop BD by the age 22-years assessment using clinical and demographic variables at each follow-up visit before diagnosis (perinatal, 11 years, 15 years, and 18 years). After that, we used the best predictive model to calculate the subgroups of subjects at higher and lower risk of developing BD and analyzed the clinical differences among them. Findings: A total of 107 (2∙8%) individuals within the cohort presented with BD type I, 26 (0∙6%) presented with BD type II, and 87 (2∙3%) presented with BD not otherwise specified. The model with variables assessed at the 18-years follow-up visit achieved the best performance: an AUC of 0∙82 (CI 0∙75–0∙88), balanced accuracy of 0∙75, sensitivity of 0∙72, and specificity of 0∙77. The most important variables to predict BD at the 18-years follow-up visit were suicide risk, a diagnosis of generalized anxiety disorder, parental physical abuse, and financial problems. Additionally, the high-risk subgroup of BD showed a high frequency of drug use and depressive symptoms. Interpretation: We developed a machine learning-based risk calculator for BD at the age of 22 years, which may be used as a tool to clinical decision making, incorporating both demographic and clinical variables. Future studies integrating data from different biological levels, such as genetics and metabolomics, could potentially help to build more accurate models. Furthermore, other researches should determine whether the features selected by the model as the most important have a causal relation with BD or whether they are prodromal symptoms. Funding Statement: This article is based on data from the study Birth Cohort − conducted by Graduate Program in Epidemiology at Universidade Federal de Pelotas with the collaboration of the Brazilian Public Health Association (ABRASCO). Furthermore, the 1993 birth cohort study was supported by the Wellcome Trust and Brazilian National Research Council (CNPq) with Brazilian Ministry of Health (DECIT – CNPq). Declaration of Interests: The authors declare no competing interests. Ethics Approval Statement: The 1993 Pelotas Birth Cohort study was approved by the local ethics committee. All participants provided written informed consent.