Comparison of Window Shapes and Lengths in Short-Time Feature Extraction for Classification of Heart Sound Signals
Summary: arXiv:2604.13567v1 Announce Type: cross
Abstract
Heart sound signals, also known as phonocardiography (PCG) signals, provide a non-invasive means for the automatic diagnosis of potential cardiovascular pathologies. The classification task for these signals can effectively be approached using a bidirectional long short-term memory (biLSTM) network, which is trained on features extracted from labeled PCG signals. Given the non-stationary nature of these signals, it is recommended to extract features from multiple short-length segments using a sliding window of specific shapes and lengths.
However, some window types can produce unfavorable spectral side lobes that distort the extracted features. Therefore, it is essential to adapt the window shapes and lengths to improve classification performance.
Research Overview
This study proposes an experimental evaluation involving three distinct window shapes, each evaluated at three different lengths. Specifically, the biLSTM network is trained and tested on statistical features extracted from the PCG signals, and the performance metrics are reported concerning the variations in window shapes and lengths.
Window Shapes and Lengths
- Gaussian Window: Known for its favorable characteristics, the Gaussian window minimizes spectral leakage and is expected to enhance classification performance.
- Triangular Window: Competes closely with the Gaussian window, particularly when used at a length of 75 ms.
- Rectangular Window: While commonly used, the rectangular window has been shown to be the least effective option for this application.
Findings
Results from the study indicate that the best classification performance is achieved using the Gaussian window for segmenting the signals. The triangular window, when applied at a length of 75 ms, demonstrates competitive performance against the Gaussian option. In contrast, the rectangular window consistently underperformed, highlighting its inadequacy for effective feature extraction in heart sound signal classification.
Conclusion
In conclusion, adapting window shapes and lengths is crucial for optimizing the classification of heart sound signals. The findings suggest that the utilization of a 75 ms Gaussian window not only enhances performance but also surpasses the outcome of baseline methods. This research highlights the importance of feature extraction strategies in the realm of cardiovascular diagnosis, paving the way for future advancements in the automatic analysis of heart sounds.
Implications for Future Research
Future investigations could explore additional window shapes and lengths, as well as the integration of advanced machine learning techniques to further improve classification accuracy. The insights gained from this study lay the groundwork for ongoing developments in the automated assessment of cardiovascular health through heart sound analysis.