Improving Heart Rate Estimation on Consumer Grade Wrist-Worn Device Using Post-Calibration Approach

Tanut Choksatchawathi,Puntawat Ponglertnapakorn,Apiwat Ditthapron,Pitshaporn Leelaarporn,Thayakorn Wisutthisen,Maytus Piriyajitakonkij,Theerawit Wilaiprasitporn

DOI: https://doi.org/10.1109/JSEN.2020.2979191

2020-03-05

Abstract:The technological advancement in wireless health monitoring through the direct contact of the skin allows the development of light-weight wrist-worn wearable devices to be equipped with different sensors such as photoplethysmography (PPG) sensors. However, the motion artifact (MA) is possible to occur during daily activities. In this study, we attempted to perform a post-calibration of the heart rate (HR) estimation during the three possible states of average daily activity (resting, \textcolor{red}{laying down}, and intense treadmill activity states) in 29 participants (130 minutes/person) on four popular wearable devices: Fitbit Charge HR, Apple Watch Series 4, TicWatch Pro, and Empatica E4. In comparison to the standard measurement (HR$_\text{ECG}$), HR provided by Fitbit Charge HR (HR$_\text{Fitbit}$) yielded the highest error of $3.26 \pm 0.34$ bpm in resting, $2.33 \pm 0.23$ bpm in \textcolor{red}{laying down}, $9.53 \pm 1.47$ bpm in intense treadmill activity states, and $5.02 \pm 0.64$ bpm in all states combined among the four chosen devices. Following our improving HR estimation model with rolling windows as feature (HR$_\text{R}$), the mean absolute error (MAE) was significantly reduced by $33.44\%$ in resting, $15.88\%$ in \textcolor{red}{laying down}, $9.55\%$ in intense treadmill activity states, and $18.73\%$ in all states combined. This demonstrates the feasibility of our proposed methods in order to correct and provide HR monitoring post-calibrated with high accuracy, raising further awareness of individual fitness in the daily application.

Signal Processing

What problem does this paper attempt to address?

### What problem does this paper attempt to solve?

This paper aims to solve the error problem caused by motion artifacts (MA) when estimating heart rate (HR) using consumer - grade wrist - worn devices (such as Fitbit Charge HR, Apple Watch Series 4, TicWatch Pro and Empatica E4) in daily activities. Specifically, the research attempts to improve the heart rate estimation accuracy of these devices in three different activity states (resting, lying down and intense treadmill activity) through a post - calibration method.

#### Main problems:
1. **The influence of motion artifacts**: In daily activities, especially during intense exercise, due to the change in the distance between the device and the skin and other interfering factors, the heart rate data recorded by the photoplethysmography (PPG) sensor will be affected by motion artifacts, resulting in inaccurate heart rate estimation.
2. **Limitations of existing devices**: Commercial consumer - grade devices usually do not allow modification of the heart rate estimation algorithm or removal of the main components of motion artifacts. Therefore, an effective post - calibration method is required to improve the accuracy of heart rate estimation.

#### Research objectives:
- **Evaluate the performance of existing devices**: Compare the heart rate estimation errors of four popular wrist - worn devices in different activity states.
- **Develop a post - calibration model**: Propose a regression model based on rolling - window features to reduce errors in heart rate estimation.
- **Verify the improvement effect**: Verify the effectiveness of the proposed post - calibration method in different activity states through experiments and compare it with other methods.

#### Method overview:
- **Data collection**: Collect heart rate data from 29 participants (130 minutes per person) in three activity states (resting, lying down, intense treadmill activity).
- **Feature selection**: Select the most relevant features from multiple candidate features, including PPG heart rate, physical activity level (PAL), number of steps, gender, Pittsburgh Sleep Quality Index (PSQI) and body mass index (BMI).
- **Model construction**: Use six machine - learning algorithms (support vector regression SVR, random forest RF, Gaussian process GP, artificial neural network ANN, logistic regression LR and k - nearest neighbor regression kNN) to construct a regression model.
- **Result evaluation**: Evaluate the model performance through indicators such as mean squared error (MAE) and verify the effectiveness of the post - calibration method.

Through these methods, the research shows that the proposed post - calibration method can significantly reduce the error in heart rate estimation, thereby providing more accurate data support for individual health monitoring in daily applications.