Artificial intelligence is becoming increasingly popular in medicine. A new computer program is now able to predict the rapid relaxation of the left ventricle - a measure of diastolic function - with a high degree of accuracy using a simple ECG. Could echocardiography become superfluous in the future?
Around 26 million people worldwide suffer from heart failure. A general distinction is made between left ventricular heart failure with reduced (HFrEF) and preserved (HFpEF) ejection fraction. In the latter case, the diastolic function is impaired, i.e. the filling of the heart chamber is slowed down. In many cases, this form of cardiac insufficiency develops completely unnoticed. It is therefore important to detect it early on in order to avoid consequential cardiac damage. To do this, cardiologists usually use echocardiography and measure the rate of relaxation of the left ventricle in the diastole (also called e'). Normally this is over 8 cm/s.
A value that is too low is considered an important early warning sign of damage to the heart muscle. A research group led by Nobuyuki Kagiyama at the University Hospital in West Virginia (USA) has been looking for a method of diagnosing diastolic dysfunction using ECG. They used machine learning technology, which trains an algorithm to recognize certain patterns of diastolic dysfunction in the ECG.
First of all, the researchers needed a set of ECG and echocardiography data that they could feed into the system. They received this data from 651 US test subjects who were scheduled to receive echocardiography to examine heart function. They also received a 12-lead ECG as part of the study. The exclusion criteria for participation in the study were signs of acute ischemia or cardiac arrhythmia. The researchers used traditional and processed ECG data and clinical parameters - in this case, age, dyslipidemia, and hypertension - as input for the algorithm. The trained algorithm was then tested on a small group of 163 subjects recruited from the same pool as the training cohort ("internal test").
The main part of the study took place in Canada, on a new cohort of 388 participants that differed clinically and demographically significantly from the US population ("external test"). The Canadian cohort was about 10 years older, had more male subjects, and fewer subjects with a known previous cardiovascular disease. The algorithm, therefore, worked in "new, unknown territory".
In the "internal" test on the US subjects, the algorithm already scored well: the difference between predicted e' and e' measured in echocardiography was only 0.23 cm/s. In the "external" test on the Canadian subjects, the difference was slightly larger at 1.38 cm/s, but it was still within acceptable limits.
To further validate the predicted rate of relaxation of the left ventricle (e'), the scientists used the statistical model of the ROC curve (ROC = Receiver Operating Characteristics). Analysis of the ROC curves showed that the algorithm was able to detect reduced e' (defined as septal e' below 7 cm/s and/or lateral e' below 10 cm/s) in the external cohort with 77% probability (sensitivity: 78%, specificity 77%). The algorithm also correctly diagnosed left ventricular diastolic dysfunction with a 46% probability according to the 2016 guidelines of the US American Society of Echocardiography (sensitivity: 93%, specificity: 51%).
In further analysis, the researchers found that traditional ECG data were the most important for prediction. The most relevant parameter here was the amplitude of the S-wave in the derivative V3. The clinical data, on the other hand, had only a subordinate value.
The study shows that a computer algorithm with simple ECG data can detect diastolic dysfunction of the left ventricle quite reliably. Nevertheless, the question arises whether this technique will ever find its way into clinical practice. This is because echocardiography is still considered the gold standard in the evaluation of heart function. It is inexpensive and quickly done.
ECG-based screening would therefore only make sense where either no echocardiography is available or where ultrasound cannot yet detect any disturbance of heart function. This would be the case, for example, with incipient diastolic dysfunction, which is difficult to detect in echocardiography. Here, an ECG algorithm could provide early indications and give physicians and patients advance warning. However, further studies with the algorithm are still necessary. In particular, the program needs further training to improve the hit rate. In the future, an "intelligent" ECG could then be used as a supplement in the diagnosis of heart failure.
Source:
Kagiyama N et al Machine Learning Assessment of Left Ventricular Diastolic Function Based on Electrocardiographic Features. JACC. Vol. 76, No. 8, 2020. Originally published 25 Aug 2020. DOI: https://doi.org/10.1016/j.jacc.2020.06.061