Artificial intelligence for intelligent care How machine learning algorithms can enhance the personalised treatment of patients with haemophilia A

This thesis explores the integration of machine learning with pharmacometrics to enhance the treatment of haemophilia A, a rare X-linked bleeding disorder characterized by an increased risk of spontaneous bleeding. The introduction provides an overview of the disorder, the history of prophylactic treatment, and the role of pharmacokinetics (PK) in personalized care, concluding with a discussion of future therapies such as non-factor replacement. The potential of machine learning is introduced, emphasizing algorithms like random forests, neural networks, and Gaussian Processes, alongside challenges specific to pharmacometrics such as data sparsity and interpretability. Recent ML applications in pharmacometrics are reviewed, discussing usage as part of data preparation, hypothesis generation, and predictive modelling. The thesis then presents deep compartment models (DCMs), a hybrid framework combining neural networks with differential equations, which simplifies PK modelling, handles sparse data effectively, and outperforms traditional methods in speed and accuracy. Variational inference (VI) is proposed as an alternative to conventional mixed-effects estimation, yielding stable and precise results. Applications of machine learning algorithms for improving haemophilia A treatment are detailed, including models for predicting factor VIII (FVIII) pharmacokinetics in prophylactic and perioperative settings and for optimizing dosing based on bleeding risk using repeated time-to-event (RTTE) models. The OPTI-CLOT web-portal is introduced as a platform for personalized dosing recommendations. The thesis concludes by advocating for hybrid machine learning approaches to address challenges in personalized treatment, offering insights for broader adoption in rare disease management.