Prediction of Lithium Therapeutic Target Attainment in Adolescents Using an XGBoost Model: A Single-Center Retrospective Study.

Objective: Lithium carbonate is a key psychiatric treatment limited by a narrow therapeutic index and significant pharmacokinetic variability. We aimed to develop a machine learning (ML) model to predict therapeutic target attainment in adolescents treated with lithium. Methods: First, the dataset was split into training, internal validation, and held-out testing sets. Second, we used the Boruta algorithm to select key clinical features from the original variables. Third, ten different machine learning algorithms and ensemble strategies were comprehensively screened and optimized using cross-validation. Finally, the best-performing model was selected based on AUC and accuracy, the optimal decision threshold was determined, and the final model was interpreted using SHAP analysis. Additionally, model calibration and net benefit were evaluated using calibration metrics and decision curve analysis (DCA), respectively. The final model was deployed as a web-based tool. Results: A total of 406 serum concentration measurements from 279 adolescents were included. Boruta selected 5 key features from 49 original variables: weight, height, daily dose, albumin (ALB), and alkaline phosphatase (ALP). The top three base models were Logistic Regression, AdaBoost, and XGBoost. The XGBoost model was ultimately selected for its superior overall performance and interpretability. It achieved an AUC of 0.846 (95% CI: 0.742-0.928), an accuracy of 0.770 (95% CI: 0.671-0.861), and a recall of 0.754 (95% CI: 0.630-0.865) on the testing set. Furthermore, the model demonstrated acceptable calibration and showed a positive net benefit on decision curve analysis. Discussion: The developed XGBoost model may serve as a reference for estimating lithium therapeutic target attainment and identifying adolescents at risk of non-attainment.