Wearable-Based Prediction of Heart Failure Decline Using Deep Learning

The landscape of heart failure management is evolving, with new technologies offering the potential for more proactive and personalized care. A study, known as TRUE-HF (Apple-CPET Ted Rogers Understanding Exacerbations of Heart Failure, NCT05008692), is investigating whether data collected from Apple Watches, combined with traditional clinical assessments, can improve the prediction of outcomes for individuals living with heart failure. This research, conducted by the University Health Network in Toronto, Canada, represents a significant step toward remote monitoring and early intervention in a condition that places a substantial burden on healthcare systems.

Heart failure is a complex syndrome, and predicting which patients will experience worsening symptoms – exacerbations – remains a clinical challenge. The TRUE-HF study aims to address this by leveraging the continuous data stream provided by wearable technology. The cardiopulmonary exercise test (CPET) is currently a key tool for assessing prognosis in heart failure patients, but it’s not practical for frequent, routine monitoring. Researchers are exploring whether the wealth of biometric data from an Apple Watch – including heart rate, activity levels, and even oxygen saturation – can supplement and potentially enhance the predictive power of CPET and other standard tests.

How the TRUE-HF Study Works

The TRUE-HF study enrolled outpatients with heart failure receiving care at the University Health Network. Participants, aged 18 and over, were provided with Apple Watches and guidance on their use. Crucially, the data gathered from the watches were not used to directly inform clinical decisions during the study period. rather, they were collected for research purposes. Participants also underwent comprehensive clinical evaluations, including CPET, bloodwork, and a six-minute walk test, both at the beginning and end of the study. Daily surveys were used to capture information on symptoms like shortness of breath, leg swelling, and medication changes, as well as any unplanned healthcare utilization – emergency room visits or hospital admissions.

A key aspect of the study design is the inclusion of an external validation cohort from the National Institutes of Health’s (NIH) All of Us Research Program. This allows researchers to test the generalizability of their findings beyond the initial study population. The All of Us cohort uses data from Fitbit devices, requiring a ‘knowledge distillation’ approach to adapt the TRUE-HF model to the different data streams. This involves using the more comprehensive TRUE-HF model as a ‘teacher’ to train a streamlined ‘student’ model that can work with the Fitbit data.

Data Collection and Analysis: Beyond Simple Step Counts

The TRUE-HF study isn’t simply counting steps. Researchers are collecting a range of data points from the Apple Watch, including step count, exercise time, distance traveled, heart rate, heart rate variability, and oxygen saturation. These variables are aggregated into 90-minute summaries to account for varying data resolutions. To ensure data quality, records with values more than three standard deviations from the population mean were removed. Missing data were addressed using a forward-filling imputation method, which preserves the temporal integrity of the data.

The study employs a sophisticated deep learning model to analyze this data. This model, described in detail in the supplementary information, incorporates patient-specific clinical information – such as age, sex, and medication dosages – and considers the temporal relationships between data points. It’s designed to predict changes in cardiopulmonary fitness over time, providing a near-continuous assessment of a patient’s condition. The model was trained on data from the first 154 patients in the TRUE-HF cohort and then tested on a held-out group of 63 patients to ensure its accuracy and prevent bias.

What Does This Mean for Heart Failure Patients?

Even as the TRUE-HF study is still ongoing, the initial findings suggest that wearable data holds promise for improving heart failure management. The ability to remotely monitor patients and detect subtle changes in their condition could allow for earlier interventions, potentially preventing hospitalizations and improving quality of life. However, it’s important to emphasize that this technology is not intended to replace traditional clinical care. Rather, it’s envisioned as a tool to augment and enhance the information available to clinicians.

The study’s focus on predicting declines in pVO2 (peak oxygen uptake), a measure of cardiopulmonary fitness, is particularly noteworthy. A 10% decline in pVO2 has been linked to worse outcomes in heart failure patients. The TRUE-HF model aims to identify these declines early, allowing clinicians to adjust treatment plans accordingly. The researchers are also evaluating the model’s ability to predict unplanned healthcare utilization, such as hospitalizations and emergency room visits.

Limitations and Future Directions

As with any research study, the TRUE-HF study has limitations. The sample size is relatively small, and the study population is limited to patients receiving care at a single center. The reliance on Apple Watch data may introduce bias, as not all patients have access to or are comfortable using this technology. The Apple Watch ECG app is also only validated for patients over 22 years of age, limiting its use in younger participants.

Looking ahead, the researchers plan to continue collecting data and refining their model. They are also exploring the potential of using this technology to personalize treatment plans and deliver targeted interventions. The knowledge-distillation approach used to adapt the model to Fitbit data suggests that this technology could be extended to other wearable devices as well, potentially broadening its reach and impact. The study’s data and model architecture are publicly available on GitHub, fostering transparency and collaboration within the research community.

What Comes Next: Refining Predictions and Expanding Access

The TRUE-HF team is currently focused on refining the predictive accuracy of their model and validating its performance in larger and more diverse populations. Further research is needed to determine the optimal way to integrate wearable data into clinical workflows and to assess the cost-effectiveness of this approach. Ongoing analyses will also explore the impact of removing structured exercise sessions from the data and the effect of different input window lengths on model accuracy. The goal is to develop a reliable and accessible tool that can empower both patients and clinicians in the fight against heart failure.