Principal Investigator: Professor Roland Eils
Charite – Universitatsmedizin Berlin, Berlin, GermanyTags: 51157, cardiovascular disease risk, coronary disease, Machine Learning, modelling of competing cardiovascular risks, neural networks, survival analysis
Cardiovascular diseases are the leading cause of death globally. In cardiovascular disease, an early detection results in less complications and better prognosis. Currently, disease detection is either based on the investigation of symptoms or by chance in routine assessments. This strategy may be improved, however, if the risk for disease development, depending on the current patient state, was better understood. Modelling the cardiovascular disease risk from patient’s data is the major aim of this research project. The current patient state can be described in various levels of detail, where the utilization of static descriptors such as age, sex and BMI, is the most course grained, but also the most common. Today, however, there are already far more complex and richer sources of information available that might be exploited for risk prediction. Electrocardiograms, heart-images, blood sample analytics and the genome contain detailed information on the patient that may not be captured by high-level descriptors like age and sex. An exploitation of this condensed information may help to improve the definition of the patient’s status and thereby lead to the identification of patient groups in need of earlier treatment.
Additionally, we aim to redefine risk itself. Cardiovascular disease is a very broad disease category with diverse levels of severity and incidence. Distinguishing the risk for heart-attack from the risk of arrhythmia is therefore desirable. To do so, we pursue a novel approach in the field of cardiovascular risk modelling and set out to learn the risk for defined disease events (e.g. heart-attack specifically) from our augmented description of the patient status (e.g. the images, electrocardiograms, genomics and metabolomics). For this purpose, we apply neural networks incorporating different levels of patient data for risk-distribution learning. Thereby, a sub-goal is the quantification of the effects, that the incorporation of certain types of data, for instance electrocardiograms, have on the performance of the model.
Once trained, our network will contain implicit knowledge on both, the patterns in the input contributing to a certain event, as well as inputs leading to similar events. This knowledge, hidden within the model might lead to insights in risk group identification and is by itself one of the results of this project.
While the clinical application of our model is out of scope of this work, we hope to spark interest in further projects and contribute to cardiovascular risk stratification.