Wearable Data Predicts Heart Failure Outcomes with Modern AI Model
A new artificial intelligence model, TRUE-HF, demonstrates the potential of wearable data – specifically from Apple Watches – to predict cardiopulmonary fitness and changes in heart failure patients, offering a pathway towards more personalized and proactive care. The research, detailed in a recent study, validates the relationship between wearable data and key indicators of heart failure progression, including declines in cardiopulmonary fitness and unplanned healthcare utilization.
The TRUE-HF study enrolled outpatients with heart failure, providing them with Apple Watches and guidance on their use. Researchers collected data on various metrics, including heart rate, activity levels and sleep patterns, alongside traditional clinical measurements like cardiopulmonary exercise tests (CPET) and bloodwork. This data was used to train a deep learning model capable of predicting an individual patient’s cardiopulmonary fitness and changes over time.
To address limitations in data availability, researchers also leveraged data from the National Institutes of Health’s All of Us Research Program, a large biomedical data resource. A knowledge-distillation approach was employed to adapt the TRUE-HF model to the All of Us dataset, which primarily contained heart rate and step count data from Fitbit devices. This involved a “teacher-assistant” model to facilitate knowledge transfer and maintain performance.
The TRUE-HF model processes 30 days of wearable data, incorporating patient-specific clinical information like age, sex, and weight. It utilizes a contextualized deep learning approach to analyze temporal trends and make daily predictions. The model was trained on data from 154 patients and then validated on a held-out test set of 63 patients.
Results showed a strong correlation between the model’s predictions and clinically measured CPET values. The model also demonstrated accuracy in detecting declines in cardiopulmonary fitness, with a ≥10% reduction in pVO2 associated with worse outcomes in heart failure patients. The model’s predictions of declining fitness were associated with unplanned healthcare utilization, such as hospitalizations and emergency room visits.
Researchers also investigated the impact of removing structured exercise sessions and varying the input data window length, finding that the model remained robust. Saliency analyses were conducted to understand feature importance, providing insights into which data points contributed most to the model’s predictions.
The study highlights the potential of wearable technology and AI to transform heart failure management, enabling more proactive and personalized interventions. The All of Us Research Program played a crucial role in validating these findings on a larger, more diverse population.