Reinforcement Learning for treatment and prediction in AML patient progression

Real data in general, and clinical time course data in particular, are inherently noisy, sparse, and irregularly sampled, making them especially difficult to handle. Within this project, we aim to provide accurate predictions and optimal interventions in individual patient disease progression based on such sparse dynamic time course data enriched with multivariate clinical information. To achieve this, we complement Machine Learning (ML) methods with Mechanistic Modelling (MM) of the disease to train algorithms capable of handling the nature of clinical data. Specifically, we use Reinforcement Learning, Transformers, and LSTMs to learn, predict, and efficiently sample from time courses. We apply this to longitudinal data of patients with Acute Myeloid Leukemia and generalize the approach by transferring this knowledge to models of Chronic Myelogenous Leukemia. We aim to demonstrate how temporal data can be utilized efficiently to optimize information gain and inform clinical treatment decisions. Furthermore, we use this framework to explore the reciprocal relationship between ML and MM to complement and strengthen each other.

Involved scientists

• Dr Friedemann Uschner
• Prof Dr Ingmar Glauche

Funding

IMB budget