Gender differences in bipolar disorder evaluated by inductive logic programming approach

Abstract

Objective: Bipolar Disorder (BD) literature is not conclusive about gender differences, with sparse data about the association between the variables studied. The statistical analysis has limitations that prevent more complex correlational evaluation. The Inductive Logic Programming (ILP) is machine learning and pattern recognition computational approach with the ability to solve problems describing relations (rules) extracted from the data set mined. This is a pilot study with both analyses searching for rules exploring multi-relational association in BD gender differences. Methods: The demographic and clinical (108 variables) data from Brazilian Bipolar Research Network (183 men and 417 women) were compared initially through biostatistics techniques and with assessments of ILP (Aleph method) using two approaches: considering all variables and just statistic significant variables. The ILP performance (all rules together) was measured evaluating the contingence table considering true and false positive and negative cases and searching complementary variables beyond already founded in statistical analysis. Results: The biostatistics analysis showed no differences in socio-demographic characteristics. BD women were more likely to have depression and mixed states, while men more frequently presented mania as first mood episode. Women also showed higher rates of lifetime manic episodes, comorbidities with anxiety disorders, migraine, and hypothyroidism. BD men were more likely to present alcohol abuse or dependence and substance abuse or dependence. The ILP approach with all variables generated 93 rules (41 for men and 52 for women) and 89 rules considering just significant statistic variables (33 for men and 56 for women). The observed performance range to identify women and men were: accuracy 89.7% to 98.8%, balanced accuracy 82.9% to 99.1%; sensibility 65.9% to 98.2% and F-measure 79.4% to 99.1%. No range was observed for specificity and precision, both 100%. For both genders, the rules unrelated to significant statistically observed variables were different periods of: age of disease onset, age at diagnosis and age at start of treatment. Discussion: Gender differences founded in biostatistics analysis were in agreement with the literature. The ILP results obtained were complementary to biostatistics generating rules with high precision to identify gender in BD and described several associations between them elucidating the dynamicity of BD phenomenology and phenotypes opening space for a more thorough analysis. The rules generated by ILP were still sensitive to show the effect of age reflecting the influence of woman’s reproductive cycle.