Description

With the introduction of electronic health records (EHRs), health care is finally catching up with the rest of society where digitization of core processes has become the norm. The EHR has increased the availability of observational health-care data, that are highly heterogeneous in nature and demand complex methods for their analysis and statistical machine learning. How to deal with such health-care data and what can be achieved by their analyses is seen by many as a big challenge. At the same time, artificial intelligence (AI) research has made considerable progress, in particular in tackling real-world problems. At the moment, a large number of AI researchers are focusing their research on low-hanging fruit, such as applying deep-learning methods to diagnostic imaging problems. However, medicine and health care is not just a matter of interpreting a digital image: it involves highly complex decision making to ensure that the trajectory a patient with a disease needs to take for the diagnosis, treatment, recovery, and finally outcome is optimal in some sense. As a consequence, researchers have to draw methods from the entire field of AI, not just deep learning. In addition, health-care data is usually problematic because of failure of systematic coding, use of free text to describe essential aspects of the disease follow-up, missing data, and lots of coding mistakes. However, health care and medicine are built upon a rich body of knowledge, concerning the pathophysiology of diseases, molecular, genetic, cytological, and histological characterization of stages of a disease, described by temporal and spatial disease patterns. For example to help patients with a chronic diseases managing their disorder and to prevent exacerbations, one needs knowledge about common causes of an exacerbation, typical symptoms and signs, and effective treatment to prevent or suppress the worsening of these signs and symptoms. Much of this clinical knowledge is evidence based, based on research but unable to guarantee optimal outcome. Nevertheless, clinical decisions on disease management are based on the best available evidence and it makes sense to incorporate such knowledge when building AI solutions. Clinicians and health-care researchers have recently spotted the potential of AI for clinical decision making, clearly inspired by success stories from the popular press and novel health-care projects by Big Tech. This has created a new enthusiasm for medical AI in the health-care community. The workshop builds on the rationale that learning from scratch is not possible at the current state of the art, while model-based and knowledge-based methods have been shown to support effectively analysis of data to address complex decision making problems in both static and dynamic settings. Validity and usability, as well as ethical and legal implications, of decision making based on models are also important issues in health care, in the sense that models completely learnt from data are ill-justified, cannot be explained, and therefore hard to accept by the health-care community. The workshop offers a venue for researchers and practitioners to show how model-based artificial intelligence, theory, models and algorithms can provide help physicians and clinicians to make actionable and effective decisions.