Adaptive Sojourn Time HSMM for Heart Sound Segmentation.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence in Medicine

Background:

• Heart sounds present interpretation challenges due to rapid temporal onsets and frequencies outside the human audible range.
• Computer-assisted decision systems require robust signal processing for accurate heart sound analysis.
• Existing heart sound segmentation algorithms struggle with precise event delineation and parameter optimization.

Purpose of the Study:

• To introduce a novel algorithm for heart sound segmentation utilizing a hidden semi-Markov model (HSMM).
• To enhance sojourn time parameter inference through a maximum likelihood approach for improved segmentation accuracy.
• To validate the algorithm's performance across diverse datasets, including a newly released pediatric heart sound dataset with detailed annotations.

Main Methods:

• Development of a new heart sound segmentation algorithm based on a hidden semi-Markov model (HSMM).
• Implementation of a maximum likelihood approach to infer optimal sojourn time parameters.
• Testing and validation using PhysioNet, Pascal, and a novel, meticulously annotated pediatric heart sound dataset.

Main Results:

• The proposed HSMM algorithm demonstrated significant improvements in heart sound segmentation performance across all tested datasets.
• Achieved a higher F-score (e.g., [Formula: see text] on PhysioNet) compared to the current state-of-the-art ([Formula: see text]).
• The algorithm proved effective even with training data exhibiting variability not perfectly matching testing data.

Conclusions:

• The novel approach to adapting sojourn time parameters in HSMMs offers an effective solution for heart sound segmentation.
• This method significantly enhances the robustness and accuracy of computer-assisted cardiac diagnostic systems.
• The released pediatric dataset with precise event annotations contributes valuable resources for future research in heart sound analysis.