Graph Neural Networks-Based Digital Twin Modeling of WBC and ANC Dynamics for Personalized Time-Series Prediction

This study introduces a digital twin model approach with Graph Neural Networks (GNNs) to forecast white blood cell (WBC) and absolute neutrophil count (ANC) levels during 6-mercaptopurine (6-MP) chemotherapy for treating childhood acute lymphoblastic leukemia (ALL). For incorporating clinical and lifestyle factors like daily medication dosage, anthropometric measurements (weight, height, and body surface area), diet, exercise, and vitamin D intake, the model is constructed using a synthetic patient dataset based on personalized computational Pharmacokinetics (PK) / Pharmacodynamics (PD) mathematical models. To capture temporal relationships, each data point is connected to earlier observations and is defined as a node within a time-dependent graph structure. The model, which was created with PyTorch Geometric, was evaluated by using MAE, MSE, and RMSE metrics after being trained with inputs, which are normalized by z-scores. The study results point that the developed GNN-based method works well as a digital twin tool for customized chemotherapy simulations and can predict changes in hematological parameters over time.