Adversarial training for diagnosis based on the electrocardiogram : Making deep learning analysis of the electrocardiogram more robust
Year: 2024
Type: dissertation
Abstract: Automated electrocardiogram (ECG) analysis using deep neural networks has seen promising results in recent years in various fields such as arrhythmia diagnosis and mortality prediction through age proxies. However, these deep learning models are susceptible to small indistinguishable perturbations that change the model prediction. These models also perform significantly worse when faced with domain shifts when applied to new data distributions not previously trained on. This thesis investigates how training on worst-case artificial samples through adversarial training can help generate models that can readily be adapted through fine-tuning to new datasets. Investigations are done by conducting robustness and transferability experiments using adversarial training on three different tasks in ECG analysis: atrial fibrillation diagnosis, sex classification, and age prediction. These experiments show that adversarial training creates models that generalize better and can be more efficiently adapted to new datasets with higher performance than non-adversarially trained models. This work thus paves the way for models that perform better on new datasets thus requiring less data for fine-tuning. This enables the use of high-performing models in settings where data is more scarce, thus enabling ECG analysis tools suitable in developing countries and smaller hospitals. The experiments also show that adversarial training improves the robustness of the trained models and that this robustness is retained to a smaller degree when transferred to new settings. This yields models that can withstand smaller directed perturbations without significantly changing their prediction. Thus this work also enables more consistent models, which in turn yields models that are more trustworthy.
Author: Eriksson, Arvid
Language: English
Topic: ECG Monitoring and Analysis
Subfield: Cardiology and Cardiovascular Medicine
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W6979764939