Over the past year, Cardiogram and UC San Francisco (UCSF) have presented a series of findings on how well consumer wearables like the Apple Watch and Android Wear can detect medical conditions in their users, including diabetes as well as hypertension and sleep apnea.
Now, the startup is reaching a new milestone, this morning publishing the first large-N peer-reviewed clinical study showing that the Apple Watch and other wearables can detect atrial fibrillation with a high degree of accuracy.
The study, published in JAMA Cardiology, included 9,750 participants who used Cardiogram while enrolled in UCSF’s Health eHeart Study. The company collected more than one hundred million heart rate and step counts from users, and that data was fed into a deep learning model to determine whether a particular user had atrial fibrillation. Results from the study show that the condition can be detected at 97% accuracy (c statistic), with a sensitivity (true positive rate) of 98%, and a specificity (true negative rate) of 90%. The study is a continuation of earlier work that Cardiogram had previously presented.
One of the major aspects of the study that Cardiogram is highlighting is that their deep learning model, named DeepHeart, required significantly less training data than comparable models targeting medical conditions. Only 6,338 electrocardiograms (ECGs) were required to build the model, which was 8 layers. This is an important development, since ECGs are both expensive and time consuming to perform at scale. The company has published the methodology of their deep learning model on arXiv.
Discussing the study, Brandon Ballinger, a co-founder of Cardiogram, explained to me that “This is super important. Every healthcare company needs to be built on a foundation of hard, clear evidence.” Ballinger noted that medical journal articles like the one published today are the only mechanism for building trust among health care professionals. “So we are super excited to reach this milestone.”
One caveat of the study is that it focused on patients with a known risk of atrial fibrillation, and further research needs to be conducted to determine how well the company’s deep learning model can prospectively detect the condition in patients with no treatment history. A second exploratory analysis on self-reported patients had an accuracy (c-statistic) of 71%.
Cardiogram will continue to develop more studies going forward. “Just like Google invests in search quality, we are always going to be investing in clinical research,” Ballinger said. He said that the company is developing random control trials — the gold standard in healthcare clinical studies — as well as launching an economic analysis to evaluate whether consumer wearables may improve the cost structure of health care diagnostics.
A broader challenge is what to do with these results. Ballinger said that “consumer wearables can be used for accurate detection of these conditions, but we need to figure out the workflow.” If Cardiogram detects atrial fibrillation for instance, what should happen next for the patient? Should they go to a cardiologist and get follow-up tests, should they be sent an at-home detection kit? At scale, those decisions will have staggering consequences for patient outcomes as well as health care costs, and more work has to be done to properly and rigorously develop these workflows.
Cardiogram, which was founded by Ballinger and Jonathan Hsieh in 2016, has raised $2 million in venture capital from A16Z’s Bio Fund. The app works both on Apple Watch as well as Android Wear watches with a heart rate sensor such as the Huawei Watch. The lead authors of the study were UCSF physicians Gregory Marcus, who is Director of Clinical Research in the Division of Cardiology, José Sanchez, and Geoff Tison.