Predicting Serum Levels of Lithium-Treated Patients: A Supervised Machine Learning Approach

Area of Science:

• Pharmacogenomics and Computational Biology
• Clinical Pharmacology and Therapeutics

Background:

• Routine monitoring of serum lithium levels is standard practice.
• Existing lithium prediction models have limitations due to insufficient performance.
• Accurate prediction of lithium concentration is crucial for therapeutic drug monitoring.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting serum lithium concentration.
• To identify key clinical features influencing lithium levels.
• To assess the feasibility of reduced ML models for clinical application.

Main Methods:

• Utilized a large, real-world dataset from multicenter electronic medical records.
• Applied various ML algorithms for binary (0.6-1.2 vs. 0.0-0.6 mmol/L) and continuous lithium concentration prediction.
• Validated models using 5-fold cross-validation and independent testing, evaluating accuracy and feature importance.

Main Results:

• ML models achieved average accuracies of 0.70-0.73 for binary and 0.68-0.75 for continuous predictions.
• Identified seven important features: older age, lower systolic blood pressure, higher lithium doses, and specific concomitant medications (valproic acid, -pine drugs), and substance use disorders.
• Reduced models with fewer features maintained high average accuracies (0.67-0.74).

Conclusions:

• Machine learning effectively processes complex clinical data to predict lithium concentration.
• The developed ML models offer a potential tool to aid clinical decision-making.
• Predictive modeling may help optimize the frequency of serum lithium level monitoring.