Lithium is usually considered the most effective treatment for bipolar disorder. However, at least 30% of lithium-treated patients do not have a clinically significant response; and only 30% have a long-lasting full response. Since close laboratory monitoring is required, and adverse effects are not uncommon, identifying predictors of lithium response is highly desirable.
In this largest-ever study of lithium recipients with bipolar disorder, investigators used machine learning to examine medical records of 1266 patients with bipolar disorder treated with lithium for at least 1 year at 7 international specialty clinics.
In these patients, the authors report that 35% were full responders and 65% were nonresponders. This curious dichotomy (where were the partial responders that are so common in clinical practice) is not explained and is an important question about the study.
There was a considerable amount of variability in the amount of information available at each site, and in the treatment response at different sites, as a result only 138 clinical, demographic, and family history features were used for analysis. These features were based on eliminating all data for which more than 40% of data were missing.
The investigators could accurately categorize responders versus nonresponders with a specificity of 0.91 (those who met all the criteria were very likely to respond to lithium), a sensitivity of 0.53 (those who did not meet all the criteria often still responded to lithium).
The most powerful individual variables predicting positive lithium response were a completely episodic clinical course (resolution of all symptoms between episodes), absence of a chronic clinical course (a similar variable), and absence of rapid cycling (more than four episodes a year).
“Within the most robustly performing site, completely episodic clinical course was the most informative predictor of lithium response. This supports Grof’s highlighting of inter-episode remission quality as a central phenotypic element associated with lithium response.”
Despite reliance on retrospective, occasionally incomplete medical records data, these findings suggest that readily accessible clinical information might improve accurate case selection for lithium prescription and inform provider–patient discussions about medication choice. Combining clinical information with other biomarkers might ultimately increase the sensitivity and specificity of calculators predicting lithium response.
REFERENCES
Nunes A et al. Prediction of lithium response using clinical data. Acta Psychiatr Scand 2019 Oct 30; [e-pub]. (https://doi.org/10.1111/ACPS.13122)
Grof P. Sixty years of lithium responders Neuropsychobiology. 2010;62:8–16.
Related Content
