Abstract:
In this research, we attempt to create a state-of-the-art model for the classification of the PTBXL dataset and improve upon already established metrics. Firstly, we use the previous best model, ST-CNN-GAP-5 and introduce three distinct loss functions—Weighted Loss, Focal Loss, and Asymmetric Loss—and evaluate their impact on classification metrics. Additionally, the research investigates the integration of Support Vector Machines (SVMs) and neural networks as separate classifiers for each disease, demonstrating comparable outcomes to the original model. Subsequently, the study introduces a transformative shift, incorporating transformer models for feature extraction and using the original model for classification. Two strategies are explored: utilizing multiple transformers or a single transformer before and after the temporal layers of ST-CNN-GAP-5. The results revealed variations, with no improvement on previous metrics. Furthermore, the study delves into the analysis of single and combined occurrences of diseases, exploring a single-label approach and a combined healthy and disease samples approach. The findings highlight the challenges and opportunities in handling multi-label disease samples. Lastly, a hierarchical model is proposed to address classification challenges arising from mixed-label samples.
