October 18, 2022 | In the most extensive atrial fibrillation (AF) experimental study to date, a team of French and Canadian investigators found that smartwatches often misdiagnose AF patients with at least one preexisting heart condition.
Atrial fibrillation is characterized by the irregular and often rapid heart rhythms of the two upper heart chambers. Due to its sporadic nature, the condition is traditionally very challenging to diagnose and, if left untreated, can manifest into blood clots and deadly strokes. With the emergence of wearable and implantable health monitoring devices, scientists hoped these tools would increase early detection, leading to better prognoses.
While previous studies suggest that smartwatches can diagnose AF with a reasonable degree of certainty (Frontiers in Cardiovascular Medicine, DOI: 10.3389/fcvm.2022.836375 and EP Europace, DOI: 10.1093/europace/euab192), their success mainly was demonstrated in a limited and otherwise healthy patient population.
Published in the Canadian Journal of Cardiology (DOI: 10.1016/j.cjca.2022.08.222), the French and Canadian investigators monitored 734 hospitalized patients to determine if an Apple Watch Series 5 could reliably detect AF in individuals with additional cardiac rhythm anomalies.
“With the growing use of smartwatches in medicine, it is important to know which medical conditions and ECG abnormalities could impact and alter the detection of AF by the smartwatch in order to optimize the care of our patients,” said Marc Strik, MD, Ph.D., lead investigator at the French Bordeaux University Hospital LIRYC Institute, in a press release. “Smartwatch detection of AF has great potential, but it is more challenging in patients with pre-existing cardiac disease.”
In the study, each patient received a 12-lead electrocardiogram followed by a 30-second single-lead Apple Watch ECG recording. The automated Apple Watch application classified patients' readings under three labels: “atrial fibrillation,” “no signs of atrial fibrillation,” or “diagnosis unclear.”
Two electrophysiologists—physicians specializing in abnormal cardiac electrical activity—conducted a blind interpretation of the readings. One physician was given the Apple Watch recordings and asked to classify each reading under three categories: “AF,” “absence of AF,” or “diagnosis unclear.” The second physician received 100 randomly selected readings from the same batch to determine the degree to which the physicians' interpretations aligned with one another.
Of the 734 participants, the Apple Watch failed to produce a diagnosis for 20% of patients, primarily due to its inability to monitor exceptionally slow or fast heart rates. Furthermore, of those patients in sinus rhythm, 81% were correctly diagnosed, 1% were misdiagnosed with AF, and 17% received no diagnosis.
The smartwatch accurately identified AF in approximately 69% of patients, 9% were misdiagnosed as being in sinus rhythm, and 22% received no diagnosis. These findings are in stark contrast to the whopping 97% of patients in AF who were accurately diagnosed by the electrophysiologists using the same readings.
The researchers also found that patients in sinus rhythm who had premature atrial or ventricular contractions (PVCs), sinus node dysfunction, or atrioventricular block were much more likely to receive a false positive for AF. Additionally, patients in AF were three times more likely to receive a false positive for sinus rhythm if they also demonstrated atrial tachycardia or atrial flutter (AFL)—a condition where the atria of the heart beat much faster than the ventricles.
“These observations are not surprising, as smartwatch automated detection algorithms are based solely on cycle variability. Ideally, an algorithm would better discriminate between PVCs and AF. Any algorithm limited to the analysis of cycle variability will have poor accuracy in detecting AF/AFL,” explained Strik. “Machine learning approaches may increase smartwatch AF detection accuracy in these patients.”
Two members of the Canadian team—Andrés F. Miranda-Arboleda, MD, and Adrian Baranchuk, MD, of Kingston Health Science in Ontario—assert that while the Apple Watch leaves much to be desired, the study still provides valuable information about its capability and potential pitfalls.
“[This study] is of remarkable importance because it allowed us to learn [that] the performance of the Apple Watch, in the diagnosis of AF, is significantly affected by the presence of underlying ECG abnormalities. In a certain manner, the smartwatch algorithms for the detection of AF in patients with cardiovascular conditions are not yet smart enough. But they may soon be,” said Miranda-Arboleda and Baranchuk in a complementary editorial (DOI: 10.1016/j.cjca.2022.09.007).
Working Smarter with Machine Learning
As technology advances, algorithms become more refined, and machine learning models make better predictions, the combination of smartwatches and artificial intelligence may soon offer continuous and highly accurate monitoring similar to hospital-administered ECGs.
Scientists are already using machine learning models—primarily deep neural networks (DNNs)—to detect atrial fibrillation in stroke patients (International Journal of Cardiology, DOI: 10.1016/j.ijcard.2021.11.005). In another study, DNNs successfully delineated between AF and AFL (PLoS One, DOI: 10.1371/journal.pone.0261571), a feat that smartwatches alone have yet to accomplish.
Though machine learning has proven extremely useful under key circumstances, this newfound technology comes with a warning (Clinical Research in Cardiology, DOI: 10.1007/s00392-022-02012-3). Some researchers worry that while neural networks can generate accurate AF diagnoses, their “black box” nature may not explain how a diagnosis was achieved. In other words, their lack of transparency may make them ineligible for widespread clinical use.