enmark, and much of the western world, is facing an increasingly large problem with an ageing population - and thus, growing costs of taking care of the elderly population. In particular, healthcare costs and resources is an often discussed political issue as illustrated during Covid19. How will the state pay for, and support a healthcare system with increasing costs? \\
One such way is to optimise the healthcare system through the use of technology. I am talking about including technology to support decision making and diagnostics in healthcare scenarios - in particular by using machine learning. Machine learning has the potential to both decrease diagnostic times and increase diagnostic accuracy in medical imaging such as knee osteoarthritis diagnosis, skin cancer diagnosis and breast cancer detection. Deep learning for example, a recently incredibly popular area of machine learning, has been applied in healthcare studies and found highly accurate, but nevertheless, it is rarely seen that such technologies are used in our hospitals - so why is that? One clear issue is that of explainability, transparency and interpretability. When a machine learning model makes a prediction, what is the explanation for such a prediction? Is the explanation given by the model transparent and interpretable to the nurse or doctor looking at the explanation, without them being experts in machine learning, statistics and probability? As such, despite the high predictive accuracy of deep learning for example, the black-box nature of these models can make them appear unreliable and untrustworthy, which severely limits their potential and applicability in critical domains such as healthcare.\\
In this project, we will advance methods of causal modelling and causal machine learning, capable of creating and facilitating interpretability and explainability in the eyes of a nurse or a doctor. In particular, we want to develop hybrid solutions that are somewhere between classical causal modelling such as done with Structural Causal Models and descendants thereof, and deep learning as we know it today. By including casual relations in our machine learning models, we can create more transparency to an end-user. Under certain assumptions we might even be able to prompt a machine learning model for a counterfactual explanation (i.e. what would the prediction have looked like if the blood pressure had been X instead of Y for a given patient and an explanation thereof). Traditional causal models allow us to answer such questions - but they are limited in their assumptions around data collection and data richness and in their scaleability. These types of models are also very manual in their construction and require an incredibly large amount of domain knowledge. In contrast, at the other end of the spectrum we have deep learning - large and complex models which are black-box in their nature, but models which scale relatively well, and are capable of learning from observational data. We will be researching how to further bridge the gap between modern machine learning and causal models, i.e. causal machine learning methods.\\
We will be using skin cancer data from the SCIN CAG initative as a case study for testing and evaluating the methods. At the same time, we will be in continuous dialogue with healthcare professionals and researchers from SCIN CAG to ensure that the methods created, generates the trustability and explainability desired for a healthcare end user.