Predicting the exposure of mycophenolic acid in children with autoimmune diseases using a limited sampling strategy: A retrospective study

Mycophenolic acid (MPA) is commonly used to treat autoimmune diseases in children, and therapeutic drug monitoring is recommended to ensure adequate drug exposure. However, multiple blood sampling is required to calculate the area under the plasma concentration-time curve (AUC), causing patient discomfort and waste of human and financial resources. This study aims to use machine learning and deep learning algorithms to develop a prediction model of MPA exposure for pediatric autoimmune diseases with optimizing sampling frequency. Pediatric autoimmune patients' data were collected at Nanfang Hospital between June 2018 and June 2023. Univariate analysis was applied for feature selection. Ten algorithms, including Random Forest, XGBoost, LightGBM, Gradient Boosting Decision Tree, CatBoost, Artificial Neural Network, Grandient Boosting Machine, Transformer, Wide&Deep, and TabNet, were employed for modeling based on two, three, or four concentrations of MPA. A total of 614 MPA AUC_0-12h samples from 209 patients were enrolled. Among the 10 models evaluated, the Wide&Deep model exhibited the best predictive performance. The predictive performance of the Wide&Deep model using four and three blood concentration points was similar (R ² ≈ 1 for four points; R ² = 0.95 for three points). No significant difference in accuracy within ±30% was observed between models utilizing three and four blood concentration points (p = 0.06). This study demonstrates that in the Wide&Deep model, MPA exposure can be accurately estimated with three sampling points in children with autoimmune diseases. This model could help reduce discomfort in pediatric patients without reducing the accuracy of MPA exposure estimates in clinical practice.