Predominant polarity classification and associated clinical variables in bipolar disorder: A machine learning approach

• Published in: Journal of affective disorders Vol. 245; pp. 279 - 282
• Main Authors: Belizario, Gabriel Okawa, Junior, Renato Gomes Borges, Salvini, Rogerio, Lafer, Beny, Dias, Rodrigo da Silva
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 15.02.2019
• Subjects: Adult; Age Factors; Alcoholism - psychology; Algorithms; Area Under Curve; Bipolar disorder; Bipolar Disorder - classification; Bipolar Disorder - psychology; Classifier; Cohort Studies; Cross-Sectional Studies; Delusions - psychology; Demography; Feeding and Eating Disorders - psychology; Female; Hallucinations - psychology; Hospitalization - statistics & numerical data; Humans; Machine Learning; Male; Middle Aged; Predominant polarity; ROC Curve; Substance-Related Disorders - psychology; Tobacco Use Disorder - psychology; Bipolar disorder; Predominant polarity; Classifier; Machine Learning
• ISSN: 0165-0327; 1573-2517
• DOI: 10.1016/j.jad.2018.11.051

•Predominant Polarity (PP) may be an important specifier of Bipolar Disorder (BD).•The present study attempted to employ a machine learning (ML) approach to analyze PP data in a BD sample.•The results suggest that the ML approach could be effective in determining a patient´s PP, even without including number and polarity of past episodes.•Although not previously reported, some variables, such as tobacco use and comorbid eating disorders, appear to be closely associated with PP. Bipolar disorder (BD) is a severe psychiatric disorder characterized by periodic episodes of manic and depressive symptomatology. Predominant polarity (PP) appears to be an important specifier of BD. The present study employed machine learning (ML) algorithms to accurately determine a patient´s PP without the inclusion of number and polarity of past episodes, while exploring associations between PP and demographic and clinical variables. From a cohort of 148 BD patients, demographic and clinical variables were collected using a customized questionnaire and the SCID-CV. The algorithm employed was the Random-Forest method. The algorithm was programed to classify patients into either depressive or manic predominant polarities and to reveal which variables were associated to the specifier. The algorithm attained an AUC ROC of 74.72% (95% CI = 72.29–77.15%) in classifying patients into either manic or depressive PP. The variables selected by the algorithm were: (1) age at first depressive episode; (2) number of hospitalizations; (3) BD Type II; (4) manic onset; (5) delusions; (6) psychotic features at onset; (7) tobacco addiction; (8) family history of BD; (9) hallucinations; and (10) comorbid anxiety disorders, (11) alcohol dependence, (12) eating disorders and (13) substance dependence. The study is limited due to the small sample size, the inclusion of only self-reported and clinician-observed clinical variables and its cross-sectional design. The results suggest that the ML approach could be effective in determining a patient´s PP. Furthermore, although not previously reported, some variables, such as tobacco use and comorbid eating disorders, appear to be closely associated with PP.