AI-based ECG interpretation for Smartwatches
Cardiovascular disease is the leading cause of death worldwide, claiming 17.9 million lives per year. The estimated direct and indirect costs of cardiovascular disease in the United States exceed $245 billion annually. Cardiovascular disease alone accounts for 25% of all deaths in the US each year (more than 1.3 million deaths).
The Electrocardiogram (ECG) is one of the most used diagnostic tools in medicine globally. A key component of the complete clinical picture is an accurate ECG interpretation, which plays an essential role in patient management. ECG interpretation currently relies on visual assessment by skilled clinicians to recognise abnormalities. ECG interpretation is difficult for humans as it involves the ability to recognise subtle patterns and to have the visual acuity to recognise those patterns in real time. These are functions that humans traditionally struggle with but machines can excel at.
However, traditional ECG interpretation algorithms offer inferior interpretation performance when compared to experts in detecting and diagnosing abnormalities, in particular, arrhythmias such as Atrial Fibrillation (AF). In fact, incorrect algorithm interpretations by traditional methods can even negatively influence the experts’ interpretation. A study analysing ECGs from 1085 patients with a traditional rule-based computerised interpretation of AF, found that the interpretation of ECGs was incorrect for 92 patients and the physician had failed to correct it. These errors resulted in unnecessary anti-arrhythmic and anticoagulant treatment in 39 patients and unnecessary diagnostic testing for 90 of them. An incorrect final diagnosis of paroxysmal AF was also made for 43 of the patients (Bogun et al., 2004).
This is problematic due to the fact that there are millions of heart rhythm disturbances experienced each year, for which timely diagnosis and treatment is required. Most importantly, delays can have profound effects on the quality of life for those affected, with Atrial Fibrillation estimated to be responsible for a third of all ischaemic strokes.
AI-based ECG interpretation offers an improvement over traditional algorithms as it can learn from hundreds of thousands of examples that are usually directly annotated by expert cardiologists as part of standard care. This approach overcomes the need for traditional feature selection or feature engineering methods. The AI models can therefore learn complex hierarchical representations that are likely to be more closely reflective of the true biological relationships found in the real world. Allowing for more robust and repeatable automated ECG interpretation.
More recently, consumer devices such as smartwatches have incorporated ECG sensing capability, enabling continuous monitoring of heart activity and allowing users to track their health over time. Smartwatch ECG recordings also allow for more cost-effective remote patient monitoring, providing physicians with a better understanding of the patient's cardiac health remotely. This presents an opportunity for hundreds of millions of people to access affordable continuous monitoring technology without having to visit a hospital regularly and without requiring a trained health professional to supervise them.
However, the adoption of consumer based ECG devices will generate ECG data on a scale that requires rapid and accurate review, something that will not be feasible by trained professionals at scale. Therefore, AI models will be required to perform initial interpretation which is then used to triage cases that require professional review and intervention.
