A new algorithm developed by researchers at US Stanford University is capable of detecting 14 types of cardiac arrhythmia. This the algorithm does as accurately as cardiologists and in some cases even more accurately. The researchers expect the algorithm to improve the diagnosis and treatments of people with cardiac arrhythmias.
People with suspected arrhythmias are usually given an electrocardiogram (ECG) in hospital. In some cases, however, the problem does not show up on this ECG, after which patients are given a portable device to take home that permanently monitors their heart for a fortnight. This provides hundreds of hours of data, which will have to be meticulously analysed by the cardiologist to detect indications of cardiac arrhythmias. Not only is this process very time-consuming, serious problems are not always easy to distinguish from harmless irregularities in patients' heart rhythm.
Detecting subtle differences
"The differences in heart rhythm can be very subtle but have a huge impact on treatment modalities," says Pranav Rajpurkar, one of the researchers involved in the project. "For example, two forms of cardiac arrhythmias known as 'second-degree atrioventricular blocks' look almost the same. However, one requires no treatment, while the other requires immediate action."
To solve this problem, researchers from the Stanford Machine Learning Group led by professor of computer science Andrew Ng have developed a deep learning algorithm. Such an algorithm is based on the human brain and is capable of learning new skills on its own. The more examples the algorithm processes, the more accurate it can do its job.
Training algorithm
The researchers trained the algorithm in question to recognise 14 different types of cardiac arrhythmia. To do this, they used a large dataset that came from iRhythm, a US company that provides wearable technology for detecting cardiac arrhythmias. This dataset consisted of 30,000 30-second recordings of different patients with various cardiac arrhythmias. The recordings were made with the Zio Patch, a wearable device that monitors patients' hearts over a two-week period.
The deep learning algorithm trained with this dataset for a total of over seven months. After this period, the algorithm was tested. This involved six cardiologists independently reviewing a set of 336 recordings of heart rhythms from 328 unique patients. Then, the deep learning algorithm analysed the same set of recordings. This trial showed that the algorithm was able to detect all 14 heart rhythm disorders as accurately as the cardiologists. Some disorders are even detected more accurately by the algorithm.
Unburdening doctors
Incidentally, the algorithm is not intended to replace doctors, but rather to support them. By deploying the algorithm, the researchers hope to relieve doctors, leaving them more time available for other care tasks. The time savings this could potentially realise worldwide are enormous; Stanford University reports that over 300 million ECGs are estimated to be performed per year. The researchers also note that a deep learning algorithm offers some advantages over human cardiologists. For instance, a deep learning algorithm never gets tired, can detect cardiac arrhythmias instantly and can be put to work 24/7.
They also hope to improve the quality of care provided to patients with cardiac arrhythmias. For instance, the researchers hope to combine their algorithm with low-cost ECG devices so that ECG can be used more widely as a diagnostic tool. According to the researchers, this is of interest, among others, to patients living in remote areas, as access to cardiology care is often limited here. For instance, patients often live far from the hospital, which makes travelling to a doctor time-consuming.
Author: Wouter Hoeffnagel
Source: Stanford University
Source: arXiv
