By Yina Moe-Lange
Deep Learning algorithms diagnose heart arrhythmias.
As the capacity to engineer more capable Artificial Intelligence (AI) systems increases rapidly, one of the first sectors to see a tangible AI impact is in the health care sector. The current applications of Machine Learning (ML) allow for the recognition of specific health problems at a high level of accuracy and at a much faster speed compared to humans.
A prime example of this came out earlier this summer. The Stanford Machine Learning Group developed a deep learning algorithm that can detect 14 types of arrhythmia from electrocardiogram (ECG) signals.
An in (doctor’s) office ECG may not always show the arrhythmia, so the idea is that patients with suspected arrhythmia receive a wearable ECG that can monitor the heart continuously for two weeks. The outcome of this is hundreds of hours of patient data, which the algorithm can examine every second of. Using iRhythm’s wearable ECG monitor, the Stanford group collected 30,000, 30-second clips that represented a variety of arrhythmias.
The abbreviations are the following: Ectopic Atrial Rhythm (EAR), Sinus Rhythm (SINUS), Second-degree atrioventricular block (AVB_TYPE2)
A group of three expert cardiologists were asked to reach a consensus on 300 undiagnosed clips. Separately, six individual cardiologists and the algorithm were asked to diagnose the same 300 clips. Comparing whether the individual cardiologists or the algorithm more closely matched the consensus opinion, it was found that “the algorithm is competitive with the cardiologists, and able to outperform cardiologists on most arrhythmias.”
As the Stanford group acknowledges, an advantage of using the algorithm is that it never gets fatigued and “can make arrhythmia detections instantaneously and continuously.” This type of work could extend access to diagnosis and treatment of arrhythmias and other heart conditions, specifically to those who do not have access to a high-level cardiologist in person.
These results are not the only place that algorithms have had success with heart health issues. At the University of Nottingham in the UK, a group used four ML algorithms to predict heart attacks. The four algorithms, random forests, logistic regression, gradient boosting and neural networks, all performed better than the American College of Cardiology/American Heart Association guidelines. They note that the best performing algorithm, neural networks, “correctly predicted at 7.6% more events than the ACC/AHA method, and it raised 1.6% fewer false alarms.”
It is important to note here, that the elimination of doctors is not the objective but rather incorporating the successes of the algorithms into diagnoses in order to augment and increase the effectiveness of doctors. Another benefit of using machines in conjunction with diagnosis is that they perform in a highly predictable fashion 24 hours a day. Their error rates are known and constant whereas different doctors may be more subjective and have greater variance in error rates.
These systems do not really have to be better than the best doctor – just more predictable in performance.