Preview

PDF, English Download (11MB) | Terms of use

Abstract

Diastolic heart failure is the most common cause of heart insufficiency worldwide. Diagnosis is typically established via hemodynamic measurements during invasive cardiac catheterization. However, this is associated with interventional risks for the patient. In this dissertation artificial intelligence (AI) models are proposed to predict left-ventricular filling pressures based on non-invasive cardiac magnetic resonance imaging (MRI). A total cohort of 66,936 patients receiving cardiac catheterization, including 11,699 cardiac MRI examinations, was investigated. The developed AI model could distinguish between elevated and normal filling pressures, providing valuable information on the heart’s diastolic function. The novel approach was found superior to established echocardiographic biomarkers and human experts. A secondary AI model was developed to automatically diagnose various types of cardiomyopathies from cardiac MRI. The detectable disease patterns were: hypertrophic, dilated and ischemic cardiomyopathy, cardiac amyloidosis and control. The AI only required a single MRI frame to reach a diagnosis, which could enable less time-intensive MRI protocols in the future. Both AI applications were introspected by attention mapping revealing the AI’s approach to solve those tasks, thus contributing to explainability. The AI model behind the filling pressure prediction was validated in three independent hospitals involving multiple MRI manufacturers, protocols and models. In essence, artificial neural networks can predict filling pressure and diagnosis from cardiac MRI, representing a highly scalable approach in the face of overburdened health systems, potentially impacting future diagnosis and treatment strategies.