Health technology is one of the focus domain at Samsung Electronics and working on delivering comprehensive features that give users opportunity to manage and achieve physical goals.
Joint with Google Health Connect solution that recognize more than 50 data types, such as exercise, blood pressure, sleep, heart rate etc.
Samsung R&D Institute – Ukraine (SRUKR) also brought a tangible contribution to enhance one of the existing features.
SRUKR developed a technology that predicts Heart Rate (HR) patterns for a future workout allowing the Galaxy Watch devices to realize a smart fitness coaching function and create personalized activity profiles.
Heart Rate prediction can be of particular importance is the new fitness trend, so called high-intensity interval training (HIIT), one of the most effective forms of full-body training, especially in terms of calories burning and weight loss. However, drawback of HIIT is that drastic HR changes may cause harmful effects on your health condition and cause pathologies in case of wrong trainings
We talked with one of our developers of the Heart Rate prediction technology, Illia Fedorin.
Q: Illia, can you tell us, if a wearable device can set individual training taking into account the biological and cardiovascular characteristics of a person?
A: Yes, modern fitness trackers constantly monitor HR.
Precise heart rate pattern analysis and heart rate forecasting can be intervals can be crucial in defining activity level for the next time. Nevertheless, modern wearable devices such as smartwatches or fitness belts during highly-intense exercises can be inaccurate. It happens because the heart rate sensors - photoplethysmogram (PPG) and electrocardiogram (ECG) - are exposed to the external noises.
Q: So, your body movement that you cannot hear, and external noises may affect the measurements?
A: Yes, exactly. Sensors used in the devices receive a raw signal together with a motion artifact that causes getting a not clear HR signal.
Q: How did you solve noises issue to get clear HR signal?
A: There were two basic steps:
1. Motion stable real-time measuring of HR and estimating HR quality.
2. Forecasting future HR dynamics based on known future workout or activity type.
To implement the idea, we used the Encoder-Decoder architecture.
For the Encoder module, utilized PPG sensor and 3-axis accelerometer (ACC) for a direct motion stable real-time HR measurement and a deep pre-trained neural network, based on combination of convolutional and recurrent layers that allows devices to catch clear HR signals.
For the Decoder module - hidden states of Encoder recurrent layers for HR pattern forecasting. Model training dataset includes about 300 logs from 98 different participants during high intensity interval training on a treadmill.
Architecture of the neural network for the HR dynamics forecasting (Details)
Q: One of the most important aspects of the HR signal measurement is high precision and ability to work in real time. What about latency in limited smartwatch resources?
A: Developed model gives an estimate of the first HR value after 3 seconds, and the delay of subsequent measurements is less than 1 second.
The mean absolute error (MAE) of the real time heart rate estimation Encoder part on the unseen test data was in average about 3 bpm, mean absolute percentage error (MAPE) is about 5%, which is comparable with a reference to medical ECG belt device performance.
The MAE for the heart rate forecasting Decoder part was in average about 5 bpm, MAPE is about 8%, which can be classified as good performance comparing with reference real-time HR estimation*.
Such a high accuracy rate allows the algorithm to forecast the person’s HR for further training intervals and set an optimal training program that takes into account individual biological abilities and prevents harmful cardio logical overloads, personalized coaching system.
Enhancing personal training and daily activity recognition by means of wearable devices (smartwatches and fitness trackers) is the future of wearable AI Healthcare domain.* Based on Association for the Advancement of Medical Instrumentation, American National Standards Institute