| dc.description.abstract | In this day and age, the fascination surrounding deep learning and AI is at its absolute peak. Both in terms of hype and controversy the current interest level is unprecedented, with exciting developments happening at a lightning pace. Yet, as is often the case when capitalist motives are the driving force behind progress, use cases that could potentially save lives are left behind. Specifically, deep learning has particularly exciting potential in the field of ECG analysis. In our research, we investigated the most prominent model type for this purpose, namely the Convolutional Neural Network (CNN). To that end, a deep learning pipeline was developed based on the renowned PTB-XL dataset. The CNN was tasked with classifying ECG signals according to the 5 diagnostic classes; Normal, Myocardial Infarction (MI), ST/T Change, Conduction Disturbance (CD) and Hypertrophy. Several experiments testing factors such as Pooling and Batch Normalization were conducted. Simultaneously, different levels of parameters such as dropout and hidden dimensions were also examined. Our findings indicated that Average Pooling was the most influential factor and that its combination with Batch Normalization produced the most effective results. The thesis also discusses ethical considerations regarding the use of such models in clinical practice, and approaches aimed at alleviating privacy concerns, such as synthetic datasets. Lastly, we emphasize the importance of developing explainable methods to better facilitate the use of deep learning models in the medical domain. In this context, the inclusion of doctors and radiologists can be considered of utmost importance. | en_US |
