Denis Cener

Semester Work
Supervisors: David Br├╝ggemann and Prof. Ender Konukoglu

Combining Deep Learning and Bayesian Networks in a Generative Model for TAVI Outcome Prediction

Severe aortic stenosis is a highly prevalent cardiovascular disease with possibly serious implications for the affected person. Transcatheter aortic valve implantation (TAVI) poses an alternative treatment option compared to classical surgery. Despite advantages of TAVI, post-surgical complications persist. Therefore, risk scores and outcome predictions based on patient factors are of essential interest. In this work, we expand upon a previously developed discriminative model for TAVI outcome prediction which combined patient images with baseline factors. The original model was based on classical Deep Learning and convolutional neural networks for image feature extraction. We have developed a new, generative model which expands upon the old one and allows us to consider patients in the database, even if their images are missing. This is made possible by a probabilistic point of view which allows marginalization of missing data. Marginalization is a principled way of dealing with missing data which is a common occurrence in the medical context. Auxiliary features which one can extract from an image serve as help when the image itself is missing or not usable. First, we used the model to generate synthetic datasets which we used for general model validation. We explored image- and categorical prior marginalization for this case. Afterwards, we focussed on the real data where we implemented image marginalization. This allowed increase of considered patients from 758 to 1081. The new model achieved an AUC of 0.73 while the previous one reached an AUC of 0.72. Overall, we could demonstrate that a generative model combining classical Deep Learning techniques and Bayesian Networks is capable of automatic TAVI outcome prediction. Future parameter tuning, extension of the model and a more extensive validation is still desired.