Related Experiment Video
Updated: Jun 22, 2025
04:24
A Novel Digital Platform for a Monitored Home-based Cardiac Rehabilitation Program
Published on: April 19, 2019
11.5K
Machine Learning Algorithms for Processing and Classifying Unsegmented Phonocardiographic Signals: An Efficient Edge
Roberto De Fazio1, Lorenzo Spongano1,2, Massimo De Vittorio1,2
1Department of Innovation Engineering, University of Salento, 73100 Lecce, Italy.
Sensors (Basel, Switzerland)
|June 27, 2024
Summary
This study developed machine learning classifiers for phonocardiogram (PCG) signals, achieving high accuracy in detecting heart conditions like coronary artery disease and mitral valve prolapse without segmentation. Neural networks offer a good balance of performance and memory efficiency for these affordable heart monitoring tools.
Area of Science:
- Biomedical Engineering
- Cardiology
- Machine Learning
Background:
- Phonocardiogram (PCG) signals offer an affordable method for monitoring heart conditions.
- Accurate classification of PCG signals is crucial for diagnosing various cardiac pathologies.
- Existing methods may require complex heart sound segmentation, limiting their applicability.
Purpose of the Study:
- To train and test machine learning classifiers (SVM, k-NN, NN) for binary and multiclass classification of PCG signals.
- To evaluate classifier performance without relying on heart sound segmentation algorithms.
- To assess the trade-off between classifier performance and memory occupation for potential implementation in resource-limited systems.
Main Methods:
- Utilized two Physionet/CinC 2016 datasets comprising 482 (binary) and 826 (multiclass) PCG signals.
- Pre-processed PCG signals including spike removal, denoising, filtering, and normalization.
- Extracted features from 5-second frames with a 1-second shift for training and testing SVM, k-NN, and NN classifiers.
Main Results:
- Binary classification achieved accuracies from 92.4% to 98.7% with memory usage from 92.7 kB to 11.1 MB.
- Multiclass classification (Normal, CAD, MVP, Benign) yielded accuracies from 95.3% to 98.6% with memory usage from 233 kB to 14.1 MB.
- Neural Networks (NNs) provided the best performance-memory trade-off; k-NN offered top performance at higher memory cost.
- Denoising improved signal-to-noise ratio (SNR) by 15-30 dB, minimally impacting classifier performance.
Conclusions:
- Machine learning classifiers, particularly NNs, can effectively classify PCG signals for heart condition monitoring without segmentation.
- The developed models demonstrate high accuracy and relatively low memory occupation, suitable for resource-constrained devices.
- This approach offers a promising, affordable tool for widespread cardiac condition screening and diagnosis.
