The purpose of this study is to develop a new tool that helps doctors choose the right cyclosporine dose for patients undergoing bone marrow transplantation. The tool is designed to predict the best dose using sparse sampling, making it practical for everyday clinical care. It combines information about population pharmacokinetics of cyclosporine with advanced artificial intelligence techniques, including machine learning and deep learning. This tool aims to improve treatment, personalize dosing for each patient, and reduce the risk of graft-versus-host disease.

Cyclosporine (CsA) is a cornerstone immunosuppressive agent used for the prevention of graft-versus-host disease (GVHD) following allogeneic hematopoietic stem cell transplantation (HSCT). Despite its widespread use, cyclosporine has a narrow therapeutic index and exhibits substantial inter- and intra-individual pharmacokinetic variability. Subtherapeutic exposure increases the risk of GVHD and graft failure, whereas excessive exposure is associated with nephrotoxicity, neurotoxicity, hypertension, and other adverse events. Variability in cyclosporine pharmacokinetics is influenced by numerous patient-specific factors, including body weight, hematocrit, age, renal and hepatic function, concomitant medications (particularly azole antifungals), genetic factors, and post-transplant physiological changes.

Current therapeutic drug monitoring (TDM) practices are primarily reactive, with dose adjustments made only after measured drug concentrations fall outside the therapeutic range. Consequently, many patients fail to achieve target cyclosporine concentrations following the initial dose and require multiple dose modifications before therapeutic exposure is attained. Although Bayesian forecasting based on population pharmacokinetic (PopPK) models has improved dose individualization, existing models often assume linear covariate-parameter relationships, have limited external validation, and may not adequately capture the complex nonlinear interactions that influence cyclosporine pharmacokinetics in bone marrow transplant recipients.

This study aims to develop and externally validate individualized cyclosporine dosing models by integrating mechanistic population pharmacokinetic modeling with advanced machine learning and deep learning techniques. A retrospective cohort will be used for model development and internal validation, while a prospective observational cohort of transplant recipients receiving standard-of-care cyclosporine therapy will be used for external validation.

Demographic characteristics, transplantation-related variables, laboratory measurements, cyclosporine dosing history, therapeutic drug monitoring results, concomitant medications, and relevant clinical outcomes will be collected from routine clinical practice. A mechanistic PopPK model will first be developed to characterize cyclosporine pharmacokinetics. Machine learning algorithms, including XGBoost and LightGBM, together with deep learning models, will then be trained to improve dose prediction by modeling complex nonlinear relationships among patient-specific covariates and residual variability. Bayesian forecasting using the PopPK model will serve as the reference approach for comparison.

Model performance will be evaluated using predictive accuracy, bias, precision, root mean square error (RMSE), mean absolute error (MAE), mean prediction error (MPE), coefficient of determination (R²), and the proportion of predicted concentrations or doses within predefined acceptable error limits. External validation will assess model generalizability in an independent prospective cohort. Model interpretability will be evaluated using SHAP (Shapley Additive Explanations) to identify the most influential variables contributing to individualized dose predictions.

The final validated hybrid model will be implemented as an R Shiny web-based clinical decision-support application capable of providing individualized initial cyclosporine dose recommendations, prediction intervals, and model explanation before the first therapeutic drug monitoring measurement. The study is expected to demonstrate whether hybrid PopPK-machine learning and deep learning approaches provide superior predictive performance compared with conventional Bayesian forecasting, thereby supporting precision dosing of cyclosporine, improving early therapeutic target attainment, reducing dose adjustments and drug-related toxicity, and establishing the foundation for future interventional clinical trials.

Inclusion Criteria:

• • CsA therapy indicated alone or in combination for GVHD prophylaxis.

  • Aged 2-65 years.
  • Clinically stable after first HSCT.

Exclusion Criteria:

• • Inaccurate sampling or dose administration times.

  • Patients with missing key covariates.
  • Patients lacking sufficient pharmacokinetic or TDM data

The study population consists of pediatric and adult patients aged 2-65 years who underwent first allogeneic hematopoietic stem cell transplantation (HSCT) and received cyclosporine for graft-versus-host disease (GVHD) prophylaxis. Participants will be identified retrospectively from electronic medical records and therapeutic drug monitoring (TDM) databases. Eligible patients must have complete demographic, clinical, laboratory, dosing, and cyclosporine TDM data. Patients with inaccurate dose administration or blood sampling times, missing essential covariates, or insufficient pharmacokinetic or TDM data will be excluded.