This proof-of-concept study found that smartwatch photoplethysmography coupled with a deep neural network can passively detect AF but with some loss of sensitivity and specificity against a criterion-standard ECG. Further studies will help identify the optimal role for smartwatch-guided rhythm assessment.

We train and validate a semi-supervised, multi-task LSTM on 57,675 person-weeks of data from off-the-shelf wearable heart rate sensors, showing high accuracy at detecting multiple medical conditions, including diabetes (0.8451), high cholesterol (0.7441), high blood pressure (0.8086), and sleep apnea (0.8298). We compare two semi-supervised training methods, semi-supervised sequence learning and heuristic pretraining, and show they outperform hand-engineered biomarkers from the medical literature. We believe our work suggests a new approach to patient risk stratification based on cardiovascular risk scores derived from popular wearables such as Fitbit, Apple Watch, or Android Wear.

Introduction: We aimed to evaluate whether a novel deep neural network (DNN) can predict cardiovascular risk factors from off-the-shelf wearables with a photoplethysmographic (PPG) heart rate sensor and accelerometer. Longitudinal heart rate variability and activity patterns have previously been associated with incident hypertension, diabetes, and sleep apnea, conditions which are frequently undiagnosed. Methods: Health eHeart, an IRB-approved UCSF study, enrolled 6,115 active users of the Cardiogram app for Apple Watch. Heart rate and step counts were collected for a period of 1 to 53 weeks (mean=8.9). Data from 70% of participants (33,628 person-weeks of data) was used to train a semi-supervised, multi-task DNN with both convolutional and recurrent layers to simultaneously predict prevalent hypertension, sleep apnea, and diabetes. Test performance characteristics were estimated using the remaining 30% of participants. Results: Mean age was 42.3 ± 12.1, 69% male. 2,230 (36.5%) of participants had hypertension, 1,016 (16.6%) had sleep apnea, and 462 (7.6%) had diabetes. In the validation set, the DNN outperformed a baseline logistic regression model incorporating age, sex, and beta blocker use, predicting prevalent hypertension with a c-statistic of 0.819 (95% CI 0.76-0.88; with an optimal operating point yielding 84.8% sensitivity and 63.6% specificity) vs a baseline c-statistic of 0.682 (95% CI 0.60-0.76), and prevalent sleep apnea with a c-statistic of 0.902 (95% CI 0.85-0.95; with an optimal operating point yielding 90.4% sensitivity and 59.8% specificity) vs a baseline c-statistic of 0.459 (95% CI 0.39-0.53). Results were not statistically significant for diabetes. Conclusions: Our DNN demonstrates surprisingly good prediction of hypertension and sleep apnea given that its only inputs are heart rate and step count. Whether such DNNs can provide durable and portable predictions for these conditions in other study samples is worth pursuing.

No abstract

Introduction: While mHealth platforms can enable rapid participant recruitment, the first 5 ResearchKit apps retained less than 10% of daily participants after the first 90 days, whereas well-optimized consumer mobile apps like Instagram and Twitter retain 31% and 48%, respectively. We proposed to apply design principles from the world of consumer internet to achieve high engagement and retention in a mHealth study. Methods: We enrolled 14,011 users of Cardiogram for Apple Watch app into the Health eHeart Study, an IRB-approved study at UCSF. We applied 3 key design principles to drive engagement and retention. First, give valuable insights back to user—e.g., notify users of abnormal heart rate spikes, and show when exercise caused a lower resting heart rate trend—using Gottman’s ratio: for every negative insight, give 5 positive ones. Second, minimize latency so users’ data updates many times per day. Third, use simple user interfaces that easily visualize trends and provide insights in small digestible formats. Results: Mean age was 42.3 ± 12.1, 31% were women. Seven days after app install, 64% of participants were active; 63% were active after 30 days, and 54% after 90 days. Retention was consistent across age groups—day 90 retention was 52% for 20-40 y.o., 55% for 40-60, and 49% for above 60. Refreshing data as frequently as possible had the highest impact on user engagement—in A/B testing, where heart rate visualizations were updated less frequently, we saw 20.9% drop in daily active users (51951 to 41094) within 7 days. The ratio of daily to monthly active users, a key measure of engagement, is 69.8% in our study, while average mHealth app is 8%. Conclusions: By applying design techniques from consumer mobile apps, we achieved day 90 retention 5x higher than the best ResearchKit app, showing that mHealth studies can retain large cohorts of participants and collect unique ambulatory health data with high engagement, improving the impact of mobile health interventions.