Phenotypic clustering of bipolar disorder supports stratification by lithium responsiveness over diagnostic subtypes

Introduction: The aim of this study was to determine whether the clinical profiles of bipolar disorder (BD) patients could be differentiated more clearly using the existing classification by diagnostic subtype or by lithium treatment responsiveness. Methods: We included adult patients with BD-I or II (N = 477 across four sites) who were treated with lithium as their principal mood stabilizer for at least 1 year. Treatment responsiveness was defined using the dichotomized Alda score. We performed hierarchical clustering on phenotypes defined by 40 features, covering demographics, clinical course, family history, suicide behaviour, and comorbid conditions. We then measured the amount of information that inferred clusters carried about (A) BD subtype and (B) lithium responsiveness using adjusted mutual information (AMI) scores. Detailed phenotypic profiles across clusters were then evaluated with univariate comparisons. Results: Two clusters were identified (n = 56 and n = 421), which captured significantly more information about lithium responsiveness (AMI range: 0.033 to 0.133) than BD subtype (AMI: 0.004 to 0.011). The smaller cluster had disproportionately more lithium responders (n = 47 [83.8%]) when compared to the larger cluster (103 [24.4%]; p = 0.006). Conclusions: Phenotypes derived from detailed clinical data may carry more information about lithium responsiveness than the current classification of diagnostic subtype. These findings support lithium responsiveness as a valid approach to stratification in clinical samples.