Team: Martijn Schut, Mark van de Wiel, Matthan Caan

In current medical decision-making, clinical risk and prediction models play an increasingly important role in determining, accelerating and improving diagnosis, prognosis and/or intervention. The development of diagnostic or predictive models based on clinical, laboratory, genetic and/or imaging variables is now at the core of many clinical research questions in both oncology and neurology. Although medical data is widely available today, these data still pose challenges for clinical models. They are often heterogeneous, multimodal (text, images, tables) and high-dimensional (lots of measurement data) with relatively small numbers of patients. Fortunately, there is ample expert knowledge and external data available. In the first sub-project on integrative learning, we will develop techniques to systematically integrate this knowledge into clinical models to improve their performance. To achieve maximum effect, these techniques are directly applied to two ‘use cases’ within Adore: predicting the progression of adrenoleukodystrophy and the disease activity of multiple sclerosis.The success and use of predictive models is determined, in addition to accuracy, by interpretability. Put bluntly: a clinician’s confidence in a model is greater if it is explainable. For complex machine learning models this is much more difficult than for traditional prediction models. In the second sub-project on model explainability, we develop visual and statistical tools to assess the relevance of predictors per patient and to summarise this across a group of patients. This gives us a better idea of what is relevant to whom. In this project we also start with an Adore use case: predicting long-term treatment effects of lymphoma cancer.

The two sub-projects in this proposal and the Adore Cell2Sample project come together in a third subproject in which we hope to improve predictive models based on cell type-specific genetic properties. If these challenges were addressed in isolation, the clinical scope of the techniques would still be limited because they are initially developed for a single specific disease, intervention, or clinical outcome. However, the setting of Adore, where two major medical specialties are brought together on an unprecedented scale, offers unique opportunities to generically address these challenges for the development and implementation of techniques. We develop such techniques, implement them, learn what works or doesn’t work in practice by applying them to both onco- and neurological data, and we develop computational techniques/tools that we implement in software packages for general use within onco- and neurological fields, in particular also for the Adore community.