Machine learning models (MLP and XGB) can predict sub-therapeutic and supra-therapeutic sirolimus concentration risks in children with vascular anomalies, with the supra-therapeutic XGB model achieving AUROC of 0.825, enabling personalized exposure risk prediction to optimize dosing accuracy.
Key Findings
Results
The multilayer perceptron (MLP) model showed optimal performance for sub-therapeutic sirolimus concentration risk prediction.
MLP achieved AUROC = 0.646 and Brier score = 0.190 for sub-therapeutic risk prediction on the test set.
Temporal validation yielded AUROC = 0.678 and Brier score = 0.190 for the sub-therapeutic model.
The sub-therapeutic risk model included BMI, white blood cells (WBC), mean corpuscular hemoglobin (MCH), triglycerides (TG), and total bilirubin (TBIL) as features.
Six machine learning models were developed and compared, with MLP selected as optimal for sub-therapeutic prediction.
Results
The extreme gradient boosting (XGB) model showed optimal performance for supra-therapeutic sirolimus concentration risk prediction.
XGB achieved AUROC = 0.825 and Brier score = 0.143 for supra-therapeutic risk prediction on the test set.
Temporal validation yielded AUROC = 0.767 and Brier score = 0.190 for the supra-therapeutic model.
The supra-therapeutic risk model included height, platelet count (PLT), alanine aminotransferase (ALT), high-density lipoprotein cholesterol (HDL), and total cholesterol (TC) as features.
The XGB supra-therapeutic model outperformed the sub-therapeutic MLP model in terms of AUROC (0.825 vs 0.646).
Methods
The study dataset comprised 134 sirolimus therapeutic drug monitoring measurements from 49 pediatric patients with vascular anomalies.
Data were retrospectively collected from 49 patients with a total of 134 TDM measurements.
Data were randomly split into training (80%) and testing (20%) sets.
An additional temporal cohort was used for external validation.
SHAP (SHapley Additive exPlanations) analysis was used to interpret the optimal models.
Results
Different sets of clinical and laboratory variables were identified as risk factors for sub-therapeutic versus supra-therapeutic sirolimus concentrations.
Sub-therapeutic concentration risk factors included BMI, white blood cells (WBC), mean corpuscular hemoglobin (MCH), triglycerides (TG), and total bilirubin (TBIL).
Supra-therapeutic concentration risk factors included height, platelet count (PLT), alanine aminotransferase (ALT), high-density lipoprotein cholesterol (HDL), and total cholesterol (TC).
Feature importance was interpreted using SHAP analysis for each optimal model.
The two risk profiles involved distinct physiological domains, with sub-therapeutic risk more related to metabolic/hematologic indices and supra-therapeutic risk more related to hepatic and lipid parameters.
Background
This study is reported as the first to use machine learning models to predict the risk of sub- or supra-therapeutic sirolimus concentrations specifically in children with vascular anomalies.
Sirolimus (rapamycin) is an mTOR receptor inhibitor used against multiple types of vascular anomalies.
Both sub-therapeutic concentrations (below effective levels) and supra-therapeutic concentrations (leading to adverse reactions) may negatively impact patient treatment outcomes.
The study aimed to ensure sirolimus blood concentrations remain within the therapeutic range to enhance efficacy and safety.
Six machine learning models in total were developed and evaluated primarily by AUROC and Brier score.
Hu Y, Li W, Fan L, Zhou Z, Guo H, Chen F, et al.. (2026). Machine Learning Models Reveal New Risk Factors for Sub-/Supra-Therapeutic Concentrations of Sirolimus in Children with Vascular Anomalies.. Drug design, development and therapy. https://doi.org/10.2147/DDDT.S563637