A Markov-Switching Model Approach to Heart Sound Segmentation and Classification

Area of Science:

• Biomedical Signal Processing
• Machine Learning for Healthcare
• Cardiology

Background:

• Accurate segmentation of heart sound signals is crucial for classifying pathological events.
• Existing methods like Hidden Markov Models (HMMs) struggle with noisy clinical data and rely on pre-extracted features.

Purpose of the Study:

• To develop a robust method for heart sound segmentation directly from raw signals, even in noisy conditions.
• To improve the classification of pathological heart sound events.

Main Methods:

• Proposed a Markov-switching autoregressive (MSAR) process to model raw heart sound signals.
• Extended MSAR to a switching linear dynamic system (SLDS) to jointly model signal dynamics and noise.
• Developed a novel algorithm fusing switching Kalman filter and duration-dependent Viterbi algorithm for improved state decoding.

Main Results:

• The MSAR-SLDS approach significantly outperformed Hidden Semi-Markov Models (HSMM) in segmenting raw heart sound signals.
• Achieved high average precision (86.1%) in identifying abnormal heartbeats on a noisy dataset.
• Demonstrated comparable performance to feature-based HSMM for segmentation.

Conclusions:

• The proposed MSAR-SLDS approach offers noticeable performance improvements in heart sound segmentation and classification on large, noisy datasets.
• This method shows potential for developing automated heart monitoring systems for pre-screening cardiac pathologies.