Machine Learning Models Reveal New Risk Factors for Sub-/Supra-Therapeutic Concentrations of Sirolimus in Children with Vascular Anomalies

Ya-Hui Hu,1,* Wan-Xia Li,2,3,* Lin Fan,1,* Zhou Zhou,3,* Hong-Li Guo,1 Feng Chen,1 Jian-Jun Zou,3 Yi Ji,4 Jin Xu,1 Wei-Min Shen4 1Pharmaceutical Sciences Research Center, Department of Pharmacy, Children’s Hospital of Nanjing Medical University, Nanjing, 210008, China; 2School of Basic Medicine and Clinical Pharmacy, China Pharmaceutical University, Nanjing, 210009, China; 3Department of Pharmacy, Nanjing First Hospital, Nanjing Medical University, Nanjing, 210006, China; 4Department of Burns and Plastic Surgery, Children’s Hospital of Nanjing Medical University, Nanjing, 210008, China*These authors contributed equally to this workCorrespondence: Yi Ji, Children’s Hospital of Nanjing Medical University, 72 Guangzhou Road, Nanjing, 210008, People’s Republic of China, Email [email protected] Wei-Min Shen, Children’s Hospital of Nanjing Medical University, 72 Guangzhou Road, Nanjing, 210008, People’s Republic of China, Email [email protected]: Sirolimus, also known as rapamycin, is an mTOR receptor inhibitor that suppresses cell proliferation and angiogenesis, demonstrating efficacy against multiple types of vascular anomalies. However, sub-therapeutic concentrations (below effective levels) and supra-therapeutic concentrations (leading to adverse reactions) of sirolimus may both negatively impact patient treatment outcomes. This study aimed to establish optimal models to predict the risk of sirolimus exposure using machine learning, ensure that sirolimus blood concentrations remain within the therapeutic range, and thus enhance the efficacy and safety of sirolimus therapy for children with vascular anomalies.Methods: We retrospectively analyzed 134 sirolimus therapeutic drug monitoring (TDM) measurements from 49 patients. Data were randomly split into training (80%) and testing (20%) sets, with an additional temporal cohort for external validation. Six machine learning models were developed to predict sub-therapeutic and supra-therapeutic risks, and evaluated primarily by the area under the receiver operating characteristic curve (AUROC) and Brier score. The optimal model was interpreted using SHapley Additive exPlanations (SHAP) analysis.Results: The sub-therapeutic risk model included body mass index (BMI), white blood cells (WBC), mean corpuscular hemoglobin (MCH), triglycerides (TG), and total bilirubin (TBIL); while the supra-therapeutic model comprised height, platelet count (PLT), alanine aminotransferase (ALT), high-density lipoprotein cholesterol (HDL), and total cholesterol (TC). The multilayer perceptron (MLP) and extreme gradient boosting (XGB) models showed optimal performance for sub-therapeutic (AUROC = 0.646, Brier = 0.190) and supra-therapeutic (AUROC = 0.825, Brier = 0.143) risk prediction, respectively, with consistent results in temporal validation (AUROC: 0.678, Brier = 0.190 and AUROC: 0.767, Brier = 0.190).Conclusion: This study is the first to use machine learning models to predict the risk of sub- or supra-therapeutic sirolimus concentrations in vascular anomalies children. By enabling personalized exposure risk prediction, the dosing accuracy of sirolimus for the treatment of children with vascular anomalies can be optimized, thereby enhancing effectiveness and safety.Keywords: sirolimus, vascular anomalies, machine learning, concentration risk prediction, children