ECG segmentation algorithm based on bidirectional hidden semi-Markov model.

Area of Science:

• Biomedical Signal Processing
• Machine Learning in Healthcare
• Cardiovascular Diagnostics

Background:

• Accurate electrocardiogram (ECG) wave segmentation is vital for diagnosing cardiovascular diseases (CVDs).
• Traditional methods may face challenges in segmenting complex ECG waveforms, especially in dynamic monitoring scenarios.
• The need for robust and automated ECG segmentation techniques is increasing with the prevalence of CVDs.

Purpose of the Study:

• To propose a novel bidirectional hidden semi-Markov model (BI-HSMM) for precise ECG wave segmentation.
• To evaluate the performance of the BI-HSMM method on standard ECG databases and real-world wearable dynamic ECG (DCG) signals.
• To demonstrate the utility of the BI-HSMM in aiding the detection and monitoring of cardiovascular diseases.

Main Methods:

• Developed a BI-HSMM incorporating probability distributions of ECG waveform durations.
• Extracted four feature vectors as observation sequences for the hidden Markov model (HMM).
• Employed logistic regression (LR) for parameter training and an improved Viterbi algorithm for segmentation, utilizing forward prediction and backward backtracking.

Main Results:

• Achieved high accuracy (97.98%) and F1 scores (P wave: 98.37%, QRS wave: 97.60%, T wave: 97.79%) on the QT database.
• Demonstrated superior performance on wearable dynamic ECG (DCG) signals with 99.71% detection accuracy and >99% F1 scores for all waveforms.
• Validated the BI-HSMM's significant ability to segment both resting and DCG signals effectively.

Conclusions:

• The proposed BI-HSMM offers a significant advancement in ECG wave segmentation accuracy.
• This method is highly effective for both standard resting ECG and real-time wearable DCG signal analysis.
• The BI-HSMM shows considerable promise for improving the early detection and continuous monitoring of cardiovascular diseases.