Exploring temperamental and clinical predictors of lithium treatment outcomes in bipolar disorder using diverse machine learning approaches

Clinical Snapshot

Key Findings

• P Value: Not specified for individual predictors

• Effect Size: Random forest AUC = 1.00; Elastic net and XGBoost F1 = 0.86

• Primary Outcome: Lithium treatment response prediction using clinical and temperamental predictors

• Nnt Or Sensitivity: Sensitivity and specificity not reported; only 12.3% of participants were lithium responders

• Confidence Interval: Not reported

Clinical Application

Temperament assessment tools are readily available but require specialized training; clinical variables easily obtainable Lithium is PBS-subsidized in Australia; findings may inform TGA-approved prescribing but require validation in Australian populations before clinical implementation Adults with bipolar disorder considering or currently receiving lithium treatment

Abstract

BACKGROUND: Lithium is a core treatment for bipolar disorder (BD), yet clinical response varies across patients. Testing accessible predictors of lithium response is critical for personalization strategies in real-world settings. METHODS: In this cross-sectional study, 179 individuals with BD were assessed using a broad range of standardized clinical instruments and the Brief TEMPS-M to evaluate affective temperaments. Lithium response was retrospectively classified via the Alda Scale. Multivariate linear regressions and three supervised machine learning classifiers (random forest, elastic net, and XGBoost) were applied to identify predictors of treatment response. Class imbalance was addressed by down-sampling during training. RESULTS: Responders (12.3 % of participants) were more likely to be employed, diagnosed with BD-II, had fewer lifetime depressive and hypomanic episodes, fewer suicide attempts, and significantly lower anxious temperament scores. In multivariable models, lifetime psychotic and mixed features, as well as higher counts of hypomanic and manic episodes, were independently associated with poorer lithium response. Machine learning models confirmed the importance of these variables, possibly with remarkably high predictive validity (random forest AUC = 1.00; elastic net and XGBoost F1 = 0.86). Hyperthymic temperament was positively associated with lithium response, whereas cyclothymic and depressive traits were associated with poorer lithium response. CONCLUSIONS: The present exploratory study should be intended as a primer for larger samples and more balanced replication studies employing reliable machine learning techniques, allowing for firm conclusions about suggestive clinical predictors of lithium response in BD, and their complex interaction.

References

1. 1.Miola, A., Fornaro, M., Di Stefano, V., Pullano, I., Monaco, F., & Steardo, L. (2026). Exploring temperamental and clinical predictors of lithium treatment outcomes in bipolar disorder using diverse machine learning approaches. Journal of Affective Disorders, 121152. https://doi.org/10.1016/j.jad.2026.121152