New Model Shows High Accuracy in Predicting Response to Lithium

Share this content:
LITHIA showed near-perfect accuracy in classifying patients into groups that did or did not respond to lithium therapy.
LITHIA showed near-perfect accuracy in classifying patients into groups that did or did not respond to lithium therapy.

In a proof-of-concept study published in Bipolar Disorders, researchers at the University of Cincinnati, Ohio, found that a novel machine learning system had high accuracy in predicting lithium response in patients with bipolar disorder.1

For the millions of individuals who have bipolar disorder – including up to 6% of the US population — there is an elevated risk of mortality and morbidity during manic episodes, underscoring the need for prompt and effective treatment.2,3 Although lithium has long been first-line therapy for bipolar disorder, a significant number of patients do not show a satisfactory response to the drug.

However, because there is no reliable method – such as the use of objective biological markers – for predicting treatment response to mood stabilizers, these patients are needlessly exposed to the risk of associated adverse effects during the trial-and-error process of identifying the optimal therapy.4 The sensitivities and specificities of current prediction methods for patients with mood disorders range from 70% to 90%, which means that 3 in 10 patients with a mood disorder are incorrectly diagnosed as healthy, and 1 in 10 healthy patients is incorrectly diagnosed as having a mood disorder.5

The current investigators used a cascading genetic fuzzy tree (GFT) design called the LITHium Intelligent Agent (LITHIA) to develop a new linguistic machine learning system. They then tested the ability of the system to predict the response to lithium using retrospective data from 20 patients who had undergone both functional magnetic resonance imaging (fMRI) and proton magnetic resonance spectrometry 1H-MRS) at a baseline session, followed by an 8-week course of lithium treatment for their first episode of mania. “Prediction was based on a set of objective pre-treatment neuroimaging parameters, including relative fMRI activation and MRS metabolite levels in various brain regions in a data-driven design,” they wrote.

According to the results, 75% of patients exhibited a response to lithium. Responders and non-responders did not differ significantly in gender, age, mood state, or presence of psychotic symptoms (P >.05). While both groups had statistically similar baseline Young Mania Rating Scale (YMRS) scores, responders had significantly lower scores at week 8 compared with non-responders [t(18)=6.58, P <.001].

LITHIA showed near-perfect accuracy in classifying patients into groups that did or did not respond to lithium therapy. In addition, the system predicted symptom reductions at 8 weeks with an accuracy of at least 88% in training and 80% in validation. The findings further demonstrate that LITHIA had a higher predictive capacity compared with 8 common classification methods.

Newer machine learning approaches are “capable of using high-dimensional data… to identify combinations of inputs that optimally discriminate between groups,” the current investigators stated. “As single neuroimaging parameters have largely failed the biomarker test in psychiatry, perhaps combining different measures in similar machine learning paradigms will yield better response prediction and diagnostic accuracy,” they concluded.

References

  1. Fleck DE, Ernest N, Adler CM, et al. Prediction of lithium response in first- episode mania using the LITHium Intelligent Agent (LITHIA): Pilot data and proof- of- concept. Bipolar Disord. 2017;19(4):259-272. doi:10.1111/bdi.12507
  2. Judd LL, Akiskal HS. The prevalence and disability of bipolar spectrum disorders in the US population: re-analysis of the ECA database taking into account subthreshold cases. J Affect Disord. 2003;73(1-2):123-131. doi:10.1016/S0165-0327(02)00332-4
You must be a registered member of Psychiatry Advisor to post a comment.

Sign Up for Free e-newsletters