Unsupervised Heartbeat Detection and Classification in Phonocardiogram Signals Using a Dissimilarity Matrix Approach

Core Concepts

This research paper introduces a novel unsupervised method for detecting and classifying heart sounds (S1 and S2) in phonocardiogram (PCG) signals by leveraging the temporal repeatability of heartbeats and the dissimilarity between cardiac cycle phases.

Abstract

• Bibliographic Information: Torre-Cruz, J., Martinez-Muñoz, D., Ruiz-Reyes, N., Muñoz-Montoro, A.J., Puentes-Chiachio, M., & Canadas-Quesada, F.J. (2022). Unsupervised detection and classification of heartbeats using the dissimilarity matrix in PCG signals. Computer Methods and Programs in Biomedicine, 221, 106909. https://doi.org/10.1016/j.cmpb.2022.106909

• Research Objective: This study aims to develop an accurate and robust unsupervised method for automatically segmenting PCG signals, identifying the presence of heart sounds S1 and S2, which is crucial for diagnosing cardiovascular diseases.

• Methodology: The proposed method employs a two-stage approach. Stage I, "Rough heartbeat detection," utilizes a novel dissimilarity matrix derived from the spectrogram of the PCG signal. This matrix captures the dissimilarity between different time frames based on their spectral content. By analyzing the frame-level spectral divergence, the algorithm identifies potential locations of heart sounds S1 and S2. Stage II, "Fine heartbeat detection," refines the initial heartbeat locations and classifies them as S1 or S2. This stage leverages the relatively constant duration of the cardiac systole compared to the diastole. A sliding window algorithm verifies and corrects the heartbeat locations identified in Stage I, ensuring accurate segmentation and classification.

• Key Findings: The proposed method demonstrates superior performance compared to existing state-of-the-art methods in heartbeat detection and classification, achieving an average accuracy of 99.4% for heartbeat detection and 97.2% for heartbeat classification on PCG signals without cardiac abnormalities. Even in the presence of cardiac abnormalities and noise, the method maintains high accuracy, exceeding 92% in all tested scenarios.

• Main Conclusions: The study concludes that the proposed unsupervised method, based on the dissimilarity matrix and frame-level spectral divergence, provides a robust and accurate solution for heart sound segmentation in PCG signals. The method's ability to handle noise and signals with cardiac abnormalities highlights its potential for real-world applications.

• Significance: This research significantly contributes to the field of automated cardiac auscultation by introducing a novel and effective method for heart sound segmentation. The high accuracy and robustness of the proposed method, particularly in challenging scenarios, make it a promising tool for early cardiovascular disease detection, especially in low-resource settings.

• Limitations and Future Research: The study acknowledges the limitations of using a limited dataset and suggests further validation on larger and more diverse datasets. Future research directions include exploring the method's applicability in real-time scenarios and integrating it with other physiological signals for comprehensive cardiovascular health monitoring.

Stats

The proposed method achieves 99.4% average accuracy in heartbeat detection for PCG signals without cardiac abnormalities.
The method achieves 97.2% average accuracy in heartbeat classification for PCG signals without cardiac abnormalities.
For PCG signals with cardiac abnormalities, the worst average accuracy of the proposed method is above 92%.
The study used a band-pass filter with cut-off frequencies from 20 to 200 Hz, targeting the primary frequency range of heart sounds S1 and S2.
A tolerance margin (η) of 160 ms was used in the sliding window algorithm, based on the maximum duration of heart sounds reported in the literature.

Quotes

"The analysis of heart sounds by phonocardiography (PCG) has become a research topic of high interest in the biomedical signal processing community in recent years."
"In this paper, we aim to develop an unsupervised detection and classification of S1 and S2 heartbeat sounds combining the frame-level spectral divergence from the dissimilarity matrix applied to PCG recordings."
"The proposed method outperforms the detection and classification performance of other recent state-of-the-art methods."

Key Insights Distilled From

Unsupervised detection and classification of heartbeats using the dissimilarity matrix in PCG signals

by J. Torre-Cru... at arxiv.org 11-06-2024

https://arxiv.org/pdf/2411.03061.pdf

Deeper Inquiries

How might this dissimilarity matrix approach be adapted for use in analyzing other types of repetitive biological signals beyond PCG?

The dissimilarity matrix approach, grounded in the principle of identifying repeating patterns within a signal, holds significant potential for application beyond PCG analysis in various biological signals exhibiting repetitive or quasi-periodic characteristics. Here's how it can be adapted:

Electroencephalography (EEG): EEG signals, reflecting brain activity, exhibit characteristic wave patterns like alpha, beta, and theta waves, particularly prominent during different sleep stages. The dissimilarity matrix could be employed to segment EEG recordings, identifying these repeating patterns and potentially detecting anomalies like sleep spindles or epileptic spikes. Adaptations would involve optimizing the frequency bands and dissimilarity metric (e.g., considering phase synchrony measures) for EEG characteristics.

Electromyography (EMG): EMG records electrical activity associated with muscle contractions, often displaying repetitive patterns during activities like walking or hand gestures. The dissimilarity matrix could be used to segment EMG signals, identifying individual muscle activation patterns and potentially aiding in the diagnosis of neuromuscular disorders. Adaptations might involve tailoring the window size to match the duration of typical muscle activation bursts.

Gait Analysis: Human gait exhibits a cyclical pattern. By recording signals from sensors placed on the body (accelerometers, gyroscopes), the dissimilarity matrix could be applied to segment gait cycles, identify subtle variations in stride length or timing, and potentially aid in assessing gait abnormalities.

Respiratory Signals: Respiratory rate and depth follow a repetitive pattern. The dissimilarity matrix could be applied to respiratory signals obtained via methods like respiratory inductance plethysmography, enabling the identification of individual breaths, detection of irregularities like apnea, and potentially aiding in the diagnosis of respiratory disorders.
Key Considerations for Adaptation:

Signal Preprocessing: Appropriate filtering and noise reduction techniques are crucial to isolate the repetitive patterns of interest within the specific biological signal.
Choice of Dissimilarity Metric: The choice of dissimilarity metric (e.g., Kullback-Leibler divergence, Euclidean distance, Dynamic Time Warping) should be tailored to the specific characteristics of the signal and the patterns being analyzed.
Window Size Selection: The window size used to calculate the dissimilarity matrix should be optimized based on the expected duration of the repeating patterns within the signal.

Could the reliance on the constant systole duration criterion be a limitation in cases of certain cardiac conditions where this assumption doesn't hold true?

You are absolutely correct. The reliance on the constant systole duration criterion, while generally valid for a healthy heart, can pose a limitation in cases of certain cardiac conditions where this assumption might not hold true.
Here are specific scenarios where this approach might falter:

Tachycardia and Bradycardia: In conditions of significantly elevated heart rate (tachycardia) or very low heart rate (bradycardia), the relative durations of systole and diastole can change. The algorithm might misinterpret shortened or prolonged systole durations as anomalies.

Arrhythmias: Heart rhythm irregularities like atrial fibrillation or ventricular tachycardia can lead to highly variable beat-to-beat intervals and inconsistent systole durations, making it challenging for the algorithm to accurately segment heartbeats and potentially leading to misclassifications.

Valvular Heart Diseases: Conditions like aortic stenosis (narrowing of the aortic valve) or mitral regurgitation (leakage of the mitral valve) can alter the pressure dynamics within the heart chambers, potentially affecting the duration of systole and leading to segmentation errors.
Possible Mitigations:

Adaptive Systole Duration: Instead of relying on a fixed systole duration, the algorithm could incorporate an adaptive mechanism that adjusts the expected duration based on the overall heart rate trends observed in the signal.
Incorporating Additional Features: The algorithm could benefit from incorporating additional features beyond timing information, such as spectral characteristics of the heart sounds, to improve robustness in cases of variable systole durations.
Machine Learning for Pattern Recognition: Training a machine learning model on a diverse dataset of PCG recordings, including those with various cardiac conditions, could enable the algorithm to learn and recognize more complex patterns and variations in systole duration.

If our understanding of the human heart's rhythmic patterns is key to this technology, what other "rhythms" in the universe might we decode using similar approaches?

The success of this dissimilarity matrix approach in decoding the heart's rhythmic patterns sparks intriguing possibilities for applying similar techniques to decipher other "rhythms" present in the universe. Here are a few captivating examples:

Pulsars: These rapidly rotating neutron stars emit beams of electromagnetic radiation that appear as periodic pulses when observed from Earth. Analyzing the subtle variations in these pulse periods using dissimilarity matrices could provide insights into the pulsar's spin-down rate, magnetic field structure, and potential interactions with companion objects.

Variable Stars: Certain stars exhibit periodic or semi-periodic variations in their brightness due to internal processes or eclipses by companion objects. Applying dissimilarity analysis to their light curves could help astronomers classify different types of variable stars, determine their pulsation modes, and uncover details about their evolution.

Seismic Waves: Earthquakes generate seismic waves that propagate through the Earth, exhibiting characteristic patterns depending on the earthquake's location, depth, and magnitude. Dissimilarity matrices could be employed to analyze seismic data, potentially improving earthquake detection, location accuracy, and our understanding of Earth's internal structure.

Climate Patterns: Earth's climate system exhibits various natural oscillations, such as El Niño-Southern Oscillation (ENSO) and the Pacific Decadal Oscillation (PDO), which influence global weather patterns. Applying dissimilarity analysis to climate data could help identify these oscillations, understand their underlying mechanisms, and improve climate prediction models.

Cosmic Microwave Background Radiation: The Cosmic Microwave Background (CMB) radiation, a remnant of the Big Bang, exhibits subtle temperature fluctuations across the sky. Analyzing these fluctuations using dissimilarity matrices could potentially reveal information about the early universe's structure, expansion rate, and the nature of dark matter and dark energy.
The key lies in identifying phenomena that exhibit repetitive or quasi-periodic behavior and then leveraging the power of dissimilarity analysis to extract meaningful information from the subtle variations within those patterns.