An electrocardiogram (ECG) is a tool used to measure the heartbeat using an electrical current. By following the current from 10 electrodes placed around the body, a traditional 12-lead ECG can give a physician an idea about your heart health. However, current analysis methods may not get the most bang for your buck.
The development of machine learning, a form of simple artificial intelligence, has taken many fields by storm. These programs can analyze hard to read or incredibly dense data and identify unique characteristics within those data. For ECGs, the data typically requires expert clinicians using special programs. With machine learning, you could do the same thing, and more, with just the program and an ECG. To this end, a team from the University of Pittsburgh created a machine learning program to analyze ECG results and do it in a way that outperforms trained physicians and produces more expansive results.
Their idea was to not only look at the temporal aspect of the ECG readout (what most programs use) but the spatial as well. Current analytical methods are only reliable in diagnosing ST-elevation myocardial infarction, aka a heart attack. Patients with another condition would continue to lengthy tests and exams before proper diagnosis. This new method can cut out the guesswork, and get a diagnosis with the ECG alone.
Machine learning programs require large datasets, so the team took ECG data from the EMPIRE study between 2013 and 2015. While most other programs used “model” data (aka perfect data) for training and validation, this study focused on using real-world ECG results. The team used three algorithms (logistic regression, gradient boosting machine, and artificial neural network) to train their program. Validation tests showed the resulting model could reliably classify ECG results with more sensitivity than expert clinicians and modern software.
Other groups have developed machine learning programs to analyze ECGs before, but the team is quick to criticize their use of non-real-world data. Their program was developed using real-world ECG readings and had four benefits of note: increased classification strength, integrable into ECG machines, lower expertise requirements, and increased interpretative strength for hard to analyze ECG results.
The study concludes, “In conclusion, using features extracted from only the prehospital 12-lead ECG, we arrived at a generalizable ML model that outperforms both commercial interpretation software and experienced clinician interpretation.”
Sources: Nature Communications, Osmosis
