Abstract by Tianwu Yang

The model-informed precision dosing (MIPD) could provide individual dosing recommendations to optimise the drug's clinical response and minimise toxic effects. Compared with the traditional dose individualisation method, MIPD adjusts the dose using quantitative evidence from mathematical models, typically population pharmacokinetic (PopPK) models. These models not only describe the typical PK profile, but also describe different sources of variability, e.g., inter-individual variability (IIV). They are potentially useful for personalising the dose recommendations for patients with haematological disease. Imatinib is a tyrosine kinase inhibitor widely used to treat patients with chronic myeloid leukaemia (CML) and certain subtypes of acute lymphoblastic leukaemia (ALL). Busulfan is a widely used myeloablative agent in conditioning regimens for hematopoietic stem cell transplantation (HSCT). Based on previous studies, both drugs could benefit from individual dosing. Thus, the overall aim of this PhD project is to improve personalised dosing of these two drugs through pharmacometrics models.

In Study 1, a systematic review of published PopPK models for imatinib was conducted, and 15 models were identified. A real-world observational dataset (N = 39) comprising both children and adult patients receiving imatinib was used to externally evaluate the models' performance. Some published models could achieve satisfactory performance on bias, with a median prediction error within ± 5%, but none met the predefined precision requirement. Published models performed less satisfactorily on children than on adult observations. Standard allometric scaling could improve the adults' model's predictive performance in children, but further PopPK studies in paediatric populations are still required to get better predictions.

In Study 2, a retrospective TDM dataset from the local Danish centre was collected, including 67 paediatric patients receiving busulfan. A PopPK analysis and two machine learning (ML) algorithms (i.e., Random Forest and XGBoost) were conducted on this dataset, and their predictive performance was evaluated through an external dataset from another centre (N = 171). The final PopPK model was described through a 1-compartment model with a time-varying clearance pattern. Several sample schedules with fewer sampling points could achieve similar predictive performance using the developed PopPK model. Both the PopPK model and the ML models showed satisfactory predictive performance on external data, with around 80% of the data predicted with ≤ ±30 % of the prediction error. Moreover, the predictive performance of the PopPK model could be further improved by Bayesian forecasting. The ML models required full-time span data to capture the entire pharmacokinetic profile.

In Study 3, a PopPK model of imatinib was developed from a dataset of 19 adult patients, and neural ordinary differential equation (NODE) models were employed to investigate the model's structure. A one-compartment model with sequential zero- and first-order absorption structure best described the imatinib data, and the NODE models independently indicated and confirmed this absorption structure. The typical values of clearance and volume of distribution, with 95% confidence intervals, were estimated at 12.3 (10.3 – 14.3) L/h and 276 (232 – 363) L, respectively. Based on the simulation study of the developed PopPK model, more than half of the patients receiving a standard dose of imatinib (400 mg) had a trough concentration below the target exposure (1,000 ng/mL). For these patients, the results highlighted the benefits of personalised dosing of imatinib, probably MIPD.

In summary, this PhD thesis explored pharmacometrics-based personalised medicine for haematological patients taking imatinib and busulfan. The published imatinib PopPK models were reviewed and evaluated, and two PopPK analyses were conducted for both imatinib and busulfan. Novel ML methods (including NODE) were explored, demonstrating their potential to improve traditional pharmacometrics modelling. Overall, the pharmacometrics approach can support personalised medicine in haematology by characterising pharmacokinetic variability, making individual dose recommendations, extrapolating from adults to paediatric populations, reducing required sample sizes, and simulating scenarios for alternative dosing regimens. Integrating these pharmacometrics models into clinical practice could be useful for haematological patients to achieve satisfactory therapeutic outcomes.