ECG segmentation and fiducial point extraction using multi hidden Markov model.

Area of Science:

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

Background:

• Accurate fiducial point (FP) extraction is crucial for electrocardiogram (ECG) signal analysis.
• Traditional methods like Classic Hidden Markov Models (HMMs) have limitations in precision.
• Existing benchmarks include partially collapsed Gibbs sampler (PCGS) and wavelet transform methods.

Purpose of the Study:

• To introduce a novel Multi Hidden Markov Model (MultiHMM) for enhanced ECG fiducial point extraction.
• To evaluate the performance of the MultiHMM against established methods using public and swine ECG databases.
• To demonstrate the superiority of MultiHMM in terms of accuracy and error variability.

Main Methods:

• Proposed a Multi Hidden Markov Model (MultiHMM) approach where each ECG beat segment is modeled by a separate ergodic continuous density HMM.
• Trained individual HMMs with varying state numbers separately.
• Estimated FP by comparing log-likelihoods of consecutive HMMs and identifying the path with maximum likelihood.

Main Results:

• The MultiHMM method achieved lower Root Mean Square Error (RMSE) values on both Swine (13ms) and QT (10ms) databases compared to Classic HMM, PCGS, and Wavelet methods.
• MultiHMM demonstrated significantly smaller error variability than other approaches.
• The proposed method outperformed all benchmark methods in ECG fiducial point extraction accuracy.

Conclusions:

• The MultiHMM approach represents a significant advancement in ECG fiducial point extraction.
• The method's high accuracy and reliability make it suitable for practical clinical applications.
• MultiHMM offers a robust alternative to existing techniques for analyzing ECG signals.