Consumer Wearable Devices Show Promise in Clinical Heart Disease Assessment

Consumer wearable devices provide clinical information similar to hospital assessment of heart disease

University of BirminghamJul 15 2024 Monitoring of heart rate and physical activity using consumer wearable devices was found to have clinical value for comparing the response to two treatments for atrial fibrillation and heart failure. The study published in Nature Medicine examined if a commercially-available fitness tracker and smartphone could continuously monitor the response to medications, and provide clinical information similar to in-person hospital assessment. The wearable devices, consisting of a wrist band and connected smartphone, collected a vast amount of data on the response to two different medications prescribed as part of a clinical trial called RATE-AF, funded by the National Institute for Health and Care Research (NIHR). Led by researchers from the cardAIc group at the University of Birmingham, the team used artificial intelligence to help analyse over 140 million datapoints for heart rate in 53 individuals over 20 weeks. They found that digoxin and beta-blockers had a similar effect on heart rate, even after accounting for differences in physical activity. This was in contrast to previous studies that had only assessed the short-term impact of digoxin. A neural network that took account of missing information was developed to avoid an over-optimistic view of the wearable data stream. Using this approach, the team found that the wearables were equivalent to standard tests often used in hospitals and clinical trials that require staff time and resources. The average age of participants in the study was 76 years, highlighting possible future value regardless of age or experience with technology. People across the world are increasingly using wearable devices in their daily lives to help monitor their activity and health status. This study shows the potential to use this new technology to assess the response to treatment and make a positive contribution to the routine care of patients." Heart conditions such as atrial fibrillation and heart failure are expected to double in prevalence over the next few decades, leading to a large burden on patients as well as substantial healthcare cost. This study is an exciting showcase for how artificial intelligence can support new ways to help treat patients better." Professor Dipak Kotecha from the Institute of Cardiovascular SciencesUniversity of Birmingham and the lead author of the study The study was funded as part of the BigData@Heart consortium from the European Union's Innovative Medicines Initiative. The RATE-AF trial was funded by the UK National Institute for Health and Care Research. University of Birmingham Journal reference: Gill, S. K., et al. (2024). Consumer wearable devices for evaluation of heart rate control using digoxin versus beta-blockers: the RATE-AF randomized trial. Nature Medicine. doi.org/10.1038/s41591-024-03094-4.

Off-the-shelf wearable trackers provide clinically-useful information for patients with heart disease

A study published in Nature Medicine examined if a commercially-available fitness tracker and smartphone could continuously monitor the response to medications, and provide clinical information similar to in-person hospital assessment. The wearable devices, consisting of a wrist band and connected smartphone, collected a vast amount of data on the response to two different medications prescribed as part of a clinical trial called RATE-AF. Led by researchers from the cardiac group at the University of Birmingham, the team used artificial intelligence to help analyze over 140 million datapoints for heart rate in 53 individuals over 20 weeks. They found that digoxin and beta-blockers had a similar effect on heart rate, even after accounting for differences in physical activity. This was in contrast to previous studies that had only assessed the short-term impact of digoxin. A neural network that took account of missing information was developed to avoid an over-optimistic view of the wearable data stream. Using this approach, the team found that the wearables were equivalent to standard tests often used in hospitals and clinical trials that require staff time and resources. The average age of participants in the study was 76 years, highlighting possible future value regardless of age or experience with technology. Professor Dipak Kotecha from the Institute of Cardiovascular Sciences at the University of Birmingham and the lead author of the study said, "People across the world are increasingly using wearable devices in their daily lives to help monitor their activity and health status. This study shows the potential to use this new technology to assess the response to treatment and make a positive contribution to the routine care of patients." "Heart conditions such as atrial fibrillation and heart failure are expected to double in prevalence over the next few decades, leading to a large burden on patients as well as substantial health care cost. This study is an exciting showcase for how artificial intelligence can support new ways to help treat patients better."

Consumer wearable devices for evaluation of heart rate control using digoxin versus beta-blockers: the RATE-AF randomized trial - Nature Medicine

In summary, a wrist-worn consumer-grade wearable device and smartphone were successfully deployed within a randomized controlled trial of older, multimorbid patients to evaluate continuous, ambulatory heart rate and physical activity. Including an average of two to three million data points per patient, including at rest and exertion, digoxin and beta-blocker therapy had similar effects on heart rate measured over a 20-week period. A neural network model of wearable sensor data showed similar performance for predicting future health status as conventional measures used in clinical trials. The RATE-AF trial was a prospective, randomized, open-label, blinded end-point trial that compared the use of low-dose digoxin versus beta-blockers for long-term heart rate control. Recruitment took place across primary care sites and three hospitals in the West Midlands region of England between 2016 and 2018. Inclusion criteria were: (1) age 60 years or older; (2) permanent AF in need of rate control; and (3) symptoms of heart failure, with breathlessness equivalent to NYHA class II or above. Exclusion criteria were limited so that the trial population reflected routine clinical practice (see published protocol for full list of selection criteria). The trial was co-designed by a patient and public involvement (PPI) team, with the aim of improving quality of life for patients with AF. Ethical approval was obtained from the East Midlands-Derby Research Ethics Committee (16/EM/0178), the Health Research Authority (IRAS 191437) and the Medicines and Healthcare Products Regulatory Agency. The trial was publicly funded by the UK National Institute for Health and Care Research (CDF-2015-08-074) and registered with clinicaltrials.gov (NCT02391337) and clinicaltrialsregister.eu (2015-005043-13). Each participant was randomized to either digoxin 62.5-250 µg or bisoprolol 1.25-10 mg once daily in a 1:1 ratio at their baseline visit. Randomization was completed using a computer-generated minimization algorithm to ensure treatment arms were balanced for gender and AF symptoms, based on the mEHRA classification. The trial was embedded into usual care within the National Health Service (NHS), with participants attending formal visits at baseline, 6 months and 12 months. Endpoints acquired were patient-reported quality of life, NT-proBNP, symptoms and functional capacity using mEHRA and NYHA class, 6MW distance and time, heart rate (pulse examination), 12-lead ECG, LVEF using cardiac ultrasound and assessment of adverse events. Funding for the wearables substudy was obtained after the main trial had commenced from the European Union Innovative Medicines Initiative BigData@Heart program (grant no. 116074). The study was supported by the Application of Artificial Intelligence to Routine Healthcare Data to Benefit Patients with Cardiovascular Disease (cardAIc) team at the University of Birmingham and University Hospitals Birmingham NHS Foundation Trust. An amendment was made to the trial protocol and subsequently approved by the Research Ethics Committee. One of the original stated aims of the substudy was to correlate wearable sensor data with patient quality of life using the Short Form (36) Health Survey (SF-36). Following work led by the PPI team that showed SF-36 to be as suboptimal measure of assessment, this was subsequently changed to NYHA class in the statistical analysis plan completed before data analysis (Supplementary Note). Participants with at least 2 months remaining in the RATE-AF trial were considered eligible for inclusion in the substudy. All participants were provided with a specific patient information leaflet written by the PPI team, and were asked to sign an optional form to indicate informed consent. As an exploratory analysis, no sample size calculation was performed in advance of recruitment. Heart rate sensor data from the first 5 weeks in the first ten participants was used to estimate the minimum number of participants needed. The average weekly heart rate, s.d. and correlation of repeated measures from this data indicated that a sample size of 40 participants would provide 90% power over 20 weeks to detect a 1/3 s.d. difference in heart rate (2 bpm) between digoxin and beta-blockers (control 72 bpm; s.d. 6 bpm; repeated measures correlation 0.91; two-sided alpha 0.05). A minimum target of 50 enrolled participants would account for death and loss to follow-up during the substudy. Consenting individuals were given a Samsung A6 Android smartphone (with a prepaid mobile data contract) and wrist-worn Fitbit Charge 2 wearable device for passive monitoring. There were no exclusions related to age or previous proficiency with information technology. Applications were preinstalled and set up for remote data collection, providing active monitoring and an educational resource for patients, including the European Society of Cardiology smartphone application (app) specifically designed for patients with AF. Participants were shown how to charge and carry out basic functions on each device, and how to use the apps. They were instructed to carry the phone throughout the day and to wear the wrist device continuously, only removing it for showering, bathing, swimming or charging. After the set-up appointment, in-person or telephone follow-up was provided after the first week, after 4 weeks and ad hoc to maintain engagement and address any concerns or technical issues raised by participants. Data collected via the device and smartphone was encrypted and uploaded to a secure server, temporarily cached on the smartphone until an appropriate Wi-Fi or mobile data connection was available. The collection of wearable data streams was automated using the RADAR-base platform, funded by the European Union Innovative Medicines Initiative RADAR-CNS (grant no. 115902). This platform allows for secure streaming of data from wearables, apps and devices to a central location. For this study, the RADAR-base platform was installed on a virtual machine hosted by Amazon Web Services in the Europe (London) region and was maintained by the Hyve (IT company, Netherlands). By applying for Fitbit developer application, the RADAR-base platform automatically collected data from registered participants, who were also able to see their own heart rate and step counts. For clinical data storage, secure electronic case report forms were generated using the Research Electronic Data Capture (REDCap) system hosted by the University of Birmingham, and the main trial case report forms hosted by the Birmingham Clinical Trials Unit. Data were analyzed by intention-to-treat according to the randomized allocation (digoxin versus beta-blockers), with no imputation for any missing data. Continuous measurements of heart rate and step count were pooled at 1-min intervals to form time-series data (heart rate averaged over each minute; step counts summed over each minute), with the primary analysis over a prespecified period of the first 20 weeks of device use. The results were summarized and presented as a number, percentage, mean and s.d. or standard error of the mean, or median with i.q.r. The Kruskal-Wallis nonparametric test or a t-test were used to determine differences between the two treatment arms depending on normality, and Spearman's test was used to quantify correlations. To account for multiple repeated measurements of heart rate over time in individual participants, generalized linear models were generated using a random-effects estimator and exchangeable correlation matrix. A post-hoc subgroup analysis according to activity levels was based on US Centers for Disease Control recommended activity levels (minimum 150 min per week aerobic activity equivalent to 15,000 steps per week, and health benefits goal of 300 min per week aerobic activity equivalent to 30,000 steps per week). Statistical analyses were performed using Stata v.17 (StataCorp LP), with a two-tailed P-value <0.05 denoting statistical significance. Machine learning algorithms were generated to explore whether continuous sensor data were comparable with conventional periodic trial measurements at the closest RATE-AF trial appointment, developed according to our previously published AI framework. Unlabeled wearable sensor data from staggered 4-h periods were used to develop a self-supervising CNN (Extended Data Fig. 2). The self-supervising model was motivated by the principle that important information is carried not only in the heart rate and step count channels, but also in the temporal interaction between those channels. To learn this interaction, an auxiliary dataset was synthesized from a training set of the original sensor data where the heart rates and step counts of each sample were scrambled across patients and dissociated. For example, a multichannel sample might include the heart rate time-series of patient A, but the step count time-series of patient B. The auxiliary dataset was combined with the original data to create a classification problem: to discriminate whether a given sample came from the original or scrambled data. Because the original data have temporal interdependencies between the channels, and the scrambled data do not, it is believed that learning this objective is equivalent to learning the relationship between those sensor channels. Heart rate measures were standardized to z-scores; step counts were normalized to the range [0,1] because of frequent and meaningful measurements of zero (inactivity). Both were defined with respect to each participant's individual statistics -- a heart rate z-score of 0 indicates the mean average heart rate for that patient. Small amounts of missing data were present throughout the recordings: these may have been short periods in which participants were not wearing their devices, or where data was not received because of connectivity issues. This missing data was neither dropped nor imputed, but used as a third time-series channel alongside heart rate and step count. This allowed the model to learn the significance of missing data instead of making assumptions about its distribution. Multichannel time-series data were the input for a one-dimensional convolutional layer of 8 filters and a kernel size of 21 (minutes), followed by a one-dimensional max pooling layer of size and stride 2. After pooling are two further convolutional layers with 20 and 32 filters, each with kernel size 21. Finally, one-dimensional global average pooling was performed to reduce the data representation to a vector of length 32. During training, dropout (with probability = 0.5) was applied to this layer to improve regularization. Finally, the prediction is made by a fully connected layer comprising a single sigmoid unit. Every layer but the last used rectified linear unit activation, and the network was trained to minimize binary cross-entropy using an Adam optimizer with a learning rate of 3 × 10 and L2 regularization with weight 1 × 10 applied to each nonbias parameter. After training the network to convergence, the dropout and classification layers were removed from the model, and the output of the final convolutional layer was used as a 32-dimensional embedding vector representing the time-series data used as input. Because the objective of this model was to predict the patient's future NYHA class, it was evaluated by embedding data from each patient's first week, and using that embedding to predict their NYHA class at the end of trial as the outcome of interest. This self-supervised model was trained using all data for each patient other than in this first week, while also holding out a subset of patients as a validation set to monitor under- or overfitting during model training. The hold-out set comprised 20% of the patient group, repeated across five iterations with k-fold cross-validation. For this exploratory analysis, participants were only included if they had available nonmissing time windows in the first and subsequent 19 weeks, and reached the final follow-up assessment for NYHA class. Models were compared for prediction of NYHA class at the end of the trial (5 months later): a conventional logistic regression model including ECG heart rate and 6MW test results (distance traveled, time taken and participant speed), and the wearable sensor model using CNN latent time-series embeddings from wearable sensor data as input features with L2 norm regularization (as used in ridge regression). We prespecified evaluation of models using the F1 score -- a metric combining precision and recall that ranges from 0 to 1, with 1 indicating perfect accuracy for classification. For each model, label smoothing was used over the NYHA class targets as a further method of regularization. The 95% CI was estimated by bootstrap resampling. During the peer review process, a post-hoc analysis was added to calculate the area under the receiver operator characteristic curve for each model, which provides an aggregate measure of classification performance with values ranging from 0 to 1 (higher indicates better performance). Machine learning analyses were performed using Python (Python Software Foundation) with scikit-learn, and TensorFlow (Google Brain). None of the organizations providing funding had any role in the design or conduct of the study (including collection, analysis and interpretation of the data) or any involvement in preparation, review or approval of the manuscript. The study is reported according to the Minimum Information about Clinical Artificial Intelligence Modeling (MI-CLAIM) checklist. Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.