Automated ECG interpretation

Screenshot of a software for digital ECG processing

Automated ECG interpretation is the use of artificial intelligence and pattern recognition software and knowledge bases to carry out automatically the interpretation, test reporting, and computer-aided diagnosis of electrocardiogram tracings obtained usually from a patient.

History

The first automated ECG programs were developed in the 1970s, when digital ECG machines became possible by third-generation digital signal processing boards. Commercial models, such as those developed by Marquette Electronics—General Electric (GE), Hewlett-Packard-Philips and Mortara incorporated these programs into clinically used devices.

Further, automated interpretation of ECGs was driven by advancements in microprocessor technology. In 1979, the introduction of the Motorola 68000 32-bit microprocessor enabled a leap forward in ECG device capabilities.

During the 1980s and 1990s, extensive research was carried out by companies and by university labs in order to improve the accuracy rate, which was not very high in the first models. For this purpose, several signal databases with normal and abnormal ECGs were built by institutions such as MIT and used to test the algorithms and their accuracy.

Since the late 2010s, deep neural networks trained directly on large collections of digital ECGs have been applied to automated interpretations.

Phases

Basic signal features of time and amplitude which are measured and form the basis for automated ECG analysis
  1. A digital representation of each recorded ECG channel is obtained, by means of an analog-to-digital converter and a special data acquisition software or a digital signal processing (DSP) chip.
  2. The resulting digital signal is processed by a series of specialized algorithms, which start by conditioning it, e.g., removal of noise, baselevel variation, etc.
  3. Feature extraction: mathematical analysis is now performed on the clean signal of all channels, to identify and measure a number of features which are important for interpretation and diagnosis, this will constitute the input to AI-based programs, such as the peak amplitude, area under the curve, displacement in relation to baseline, etc., of the P, Q, R, S and T waves, the time delay between these peaks and valleys, heart rate frequency (instantaneous and average), and many others. Some sort of secondary processing such as Fourier analysis and wavelet analysis may also be performed in order to provide input to pattern recognition-based programs.
  4. Logical processing and pattern recognition, using rule-based expert systems, probabilistic Bayesian analysis or fuzzy logics algorithms, cluster analysis, artificial neural networks, genetic algorithms and others techniques are used to derive conclusions, interpretation and diagnosis.
  5. A reporting program is activated and produces a proper display of original and calculated data, as well as the results of automated interpretation.
  6. In some applications, such as automatic defibrillators, an action of some sort may be triggered by results of the analysis, such as the occurrence of an atrial fibrillation or a cardiac arrest, the sounding of alarms in a medical monitor in intensive-care unit applications, and so on.

Applications

The manufacturing industries of ECG machines is now entirely digital, and many models incorporate embedded software for analysis and interpretation of ECG recordings with 3 or more leads. Consumer products, such as home ECG recorders for simple, 1-channel heart arrhythmia detection, also use basic ECG analysis, essentially to detect abnormalities. Some application areas are:

Implications and limitations

The automated ECG interpretation is a useful tool when access to a specialist is not possible. Although considerable effort has been made to improve automated ECG algorithms, the sensitivity of the automated ECG interpretation is of limited value in the case of STEMI equivalent as for example with "hyperacute T waves", de Winter ST-T complex, Wellens phenomenon, Left ventricular hypertrophy, left bundle branch block or in presence of a pacemaker. Automated monitoring of ST-segment during patient transport is increasingly used and improves STEMI detection sensitivity, as ST elevation is a dynamical phenomenon.

See also

References

  1. ↑ Macfarlane, Peter W.; Kennedy, Julie (23 September 2021). "Automated ECG Interpretation—A Brief History from High Expectations to Deepest Networks". Hearts. 2 (4): 433–448. doi:10.3390/hearts2040034. ISSN 2673-3846.
  2. ↑ "HeartLine - News from the Division of Cardiology at the University of California, San Francisco - Center for Biosignal Research Decoding Music of the Heart" (PDF). The 1979 release of the Motorola 68000 32-bit microprocessor chip was a gamechanger. "This chip allowed us to design a circuit board with the horsepower to do everything," said Dr. Mortara, who invented the hardware and software for the new interpretive ECG device. "It no longer had to be connected to an outside computer, and the interpretation was rendered immediately [by the machine], right at the bedside.
  3. ↑ Siontis, Konstantinos C.; Noseworthy, Peter A.; Attia, Zachi I.; Friedman, Paul A. (July 2021). "Artificial intelligence-enhanced electrocardiography in cardiovascular disease management". Nature Reviews Cardiology. 18 (7): 465–478. doi:10.1038/s41569-020-00503-2. ISSN 1759-5002. PMC 7848866. PMID 33526938.
  4. ↑ BioPac Systems. Application Note: Automated ECG Analysis
  5. ↑ Al-Fahoum, AS; Howitt, I. Combined wavelet transformation and radial basis neural networks for classifying life threatening cardiac arrhythmias, Med. Biol. Eng. Comput. 37 (1999), pp. 566–573.
  6. ↑ Mautgreve, W., et al. HES EKG expert-an expert system for comprehensive ECG analysis and teaching. Proc. Computers in Cardiology: Jerusalem, Israel 19–22 September 1989. (USA: IEEE Comput. Soc. Press, 1990. p. 77–80).
  7. ↑ Bortolan, G., et al. ECG classification with neural networks and cluster analysis. Proc. Computers in Cardiology. Venice, Italy, 23–26 September 1991. (USA: IEEE Comput. Soc. Press, 1991. p. 177-80).
  8. ↑ Sabbatini, R.M.E. Applications of artificial neural networks in biological signal processing. MD Computing, 3(2), 165-172 March 1996.
  9. ↑ Difficult ECGs in STEMI: lessons learned from serial sampling of pre- and in-hospital ECGs, Ayer et al., JECG, 2014
  10. ↑ ECG Interpretation - STEMI and equivalent, ebook
  11. ↑ The Prominent T wave: Electrocardiographic differential diagnosis[link removed], Sommers et al., American Journal of Emergency Medicine
  12. ↑ A New ECG Sign of Proximal LAD Occlusion, de Winter, NEJM, 2008

Sources


Translated and reproduced by permission of the author.