Valproic acid monitoring: Serum prediction using a machine learning framework from multicenter real-world data

Area of Science:

• Pharmacogenomics
• Computational Biology
• Clinical Chemistry

Background:

• Valproic acid (VPA) is a widely used antiepileptic and mood-stabilizing drug.
• Therapeutic drug monitoring of VPA is crucial but often requires invasive blood tests.
• Developing non-invasive methods for VPA monitoring is a significant clinical need.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting serum valproic acid (VPA) concentrations.
• To explore the potential of these models for non-invasive therapeutic drug monitoring.
• To identify key features influencing VPA levels and create simplified predictive models.

Main Methods:

• Utilized medical records from 2002-2019 from the Taiwan Chang Gung Research Database.
• Employed various machine learning algorithms to predict VPA concentrations (binary and continuous).
• Trained models on 5142 samples and validated on 644 independent samples, using accuracy with a 20 μg/ml tolerance as the primary metric.

Main Results:

• Models achieved high accuracy: 0.80-0.86 for binary and 0.72-0.88 for continuous VPA outcomes.
• Identified key predictors of higher VPA levels, including VPA dosage, specific diagnoses, and various blood parameters (albumin, calcium, creatinine, platelets, RDW-CV).
• Simplified models retained high accuracy, demonstrating efficiency.

Conclusions:

• Machine learning models show significant potential for predicting serum VPA concentrations using real-world data.
• These predictive models offer a promising avenue for reducing the frequency of invasive serum VPA monitoring in clinical practice.
• Further validation with external datasets is warranted to confirm generalizability.