Bridging the gap between explainable AI and evidence-based medicine, a patient's and expert’s perspective in inherited cardiac arrhythmias

In the last few years transparent and explainable AI received a lot of attention. In earlier research we
developed a grey box approach, which is a combination of black and white box models. While the black box model allows to learn highly accurate models, the surrogate white box provides the transparency and explainability. This approach has also been proven to perform well self-labelling settings, which are often used in settings where only limited labelled data is available. Initial experiments have shown that we can enrich this approach by including ontologies or knowledge graphs. As such the machine learning model is not only providing a model to the user. It can also provide alternative models, which can also be considered an interesting hypothesis to explore further. The different hypothesis can then be explored further through active learning, i.e. strategically collecting extra data, and as such bridging the gap between explainable AI and evidence-based medicine.
The approach we want to explore is especially appealing in the context of rare diseases, where only limited labelled data is available. In this project we will focus on patients affected by a primary arrhythmia disorder, or persons at increased risk to develop this disorder because of a known genetic predisposition. We want to take the genetic analysis to a next level, to whole genome sequencing complemented with functional genomics on the one hand and/or analyzing the digenic to oligogenic contribution on the other hand to further increase the genetic diagnostic yield. Besides predicting the outcome and natural course of the disease we also want to investigate the efficacy and impact of a treatment on the individual patient. As of to date, deciding if and which treatment to apply for a particular patient, is a complex decision process. Guidelines, patients’ preferences and clinicians’ experience all play a role and both physical and mental treatment-related impairment can occur. Integration of mental health, socio-economic, clinical and genetic information before treatment, procedural information (ablation or defibrillator implantation), and mental and clinical outcome data, could allow us to predict response to a certain treatment in all its aspects. A tool to provide both patient and clinician support in which therapeutic strategy to opt for, might facilitate the decision-making process and improve outcome.