Currently, I am working on developing heartbeat classification algorithms. I need some justification about feature engineering related to ECG/ heartbeat analysis. I am dealing with the features other than timing information (R-R intervals) such as entropy measures to quantify the irregularity of the signal and features measures the deviation from Gaussianity and linearity. My doubt is, how can we say these features are clinically relevant? what is the clinical meaning of these features that enables to improve the features? Are they aligned with the clinical evidence?. Though these features with the combination of a good learner are reasonably distinguishing the heartbeats, how can I interpret these features related to underlying physiology or clinical way? Thanks in advance to my dear researchers.
I am working with the EEG (electroencephalogram) signal, which is something similar to the ECG signal. For EEG signals, from my readings, researchers use wavelets decomposition to process signals and obtains some coefficients. From these coefficients, some summary statistics are derived, such as mean, standard deviation, kurtosis, skewness. I have not come across any justification from the clinical point of view though. Perhaps these summary statistics are derived to 'characterise' the behaviour of the biomedical signals when analysed at different resolution.
This is a very old question that underlies the huge literature on ECG hearbeat classification. As mentioned in one of the earlier answers, the key is to work with real clinicians, and avoid considering the problem as just another challenge for signal processing and machine learning experts. You have to engineer the features to fit the needs of cardiologists, not the other way around. Cardiologists are busy saving peoples' lives, therefore forget about expecting them to learn about, and trust, the latest findings of statistical signal processing.
A "normal" hearbeat (if such a thing exists) ECG recording is a sequence of positive and negative peaks (called "waves"), in known number, with known polarities, coming at known time intervals (with inter-patient and intra-patient variations, of course); most, if not all, "anomalous" heartbeats can be described by comparisons with the "normal" heartbeat in terms of sequence order, wave polarity, wave amplitude, presence or absence of this or that wave, etc. This is what MDs know, and this is the information that should be provided to them. In short, there is a wealth of prior knowledge that should be the basis of any attempt at designing a realistic automatic hearbeat classification tool for diagnosis purposes.
1) In a long-term collaboration with a leading international company in cardiac instrumentation, my group developed a methodology (called "bump modeling") based on functions (Gaussian mesa functions) that were specially designed to fit the shape of the ECG waves, thereby making an extensive use of prior knowledge:
- Building meaningful representations for nonlinear modeling of 1D- and 2D-signals: applications to biomedical signals
Neurocomputing, vol. 69, pp. 2180 - 2192 (2006)
- Automatic ECG wave extraction in long-term recordings using Gaussian mesa function models and nonlinear probability estimators
Computer Methods and Programs in Biomedicine, vol. 88, pp. 217-233 (2007)
- Automatic analysis of cardiac repolarization morphology using Gaussian mesa function modeling
Journal of Electrocardiology, vol. 41, pp. 588-594 (2008)
- Efficient modeling of ECG waves for morphology tracking,
Computers in Cardiology, Park City (2009).
2) For intracardiac recordings, we developed a new feature, based on prior knowledge, for arrythmia discrimination:
- Arrhythmia Discrimination in Implantable Cardioverter Defibrillators using Support Vector Machines applied to a New Representation of Electrograms
IEEE Transactions on Biomedical Engineering, vol. 56, pp. 1797 - 1803 (2011)
3) Bump modeling has been extended to the time-frequency analysis of EEG recordings, implemented in the Butif (Bump Time-frequency) Matlab toolbox:
- A machine learning approach to the analysis of time-frequency maps, and its application to neural dynamics
Neural Networks, vol. 20, pp. 194 - 209 (2007).
- Improving the specificity of EEG for Diagnosing Alzheimer's disease
International Journal of Alzheimer's Disease, doi: 0.4061/2011/259069 (2011).
- Blind Source Separation and Sparse Bump Modelling of Time-Frequency Representation of EEG Signals: New Tools for Early Detection of Alzheimer's Disease
Machine Learning for Signal Processing (MLSP 2005)
- Early Detection of Alzheimer's Disease by Blind Source Separation, Time Frequency Representation, and Bump Modeling of EEG Signals.
Lecture Notes in Computer Science, vol. 3696, pp. 683-692, Springer (2005) .
Thank you, Dr. Gérard Dreyfus, for your informative and kind response. I hope the articles you mentioned are really helpful for me to explore more on this area. Certainly, I will go through into this, really it's my pleasure to hear from you. Thank you.
- Badges
- Science topic