Abstract
Conference Title: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018) Conference Start Date: 2018, April 4 Conference End Date: 2018, April 7 Conference Location: Washington, DC, USA A data-driven model able to automatically localize the origin of abnormal ventricular activation from electrocardiograms will have the potential to improve clinical interventions for a variety of cardiac arrhythmias, such as premature ventricular contraction (PVC). Unfortunately, due to the large anatomical and physiological variations across individuals, a patient-specific model is necessary for an accurate localization. Such models are difficult to implement in clinical practice because the training data on each patient have to be obtained through invasive and expensive procedures. This paper proposes to overcome this shortage of clinical data by transferring the knowledge extracted from a large set of patient-specific simulation data to a small set of clinical data on the patient. A domain adaption approach is used to learn the "domain shift" between simulation and clinical data, while a novel error modeling method is developed to take into account potential simulation errors during domain adaption. The presented method is evaluated on three PVC patients. In comparison to training a patient-specific model purely from clinical data, or a combination of simulation and clinical data without considering domain shift, the presented method is able to improve the localization accuracy, especially when there is only a small number of clinical data available for training.