Project involves developing an algorithm to remove noise from heart rate data collected from wrist sensors, in order to better estimate stress.
In collaboration with Northeastern University Health Behavior Informatics Lab https://hbi.ccs.neu.edu/
git clone https://github.com/tjeng/HRV_analysis.git
Stress has devastating effects on health, resulting in headaches, stomach upset, high blood pressure, chest pains, heart problems, depression, and anxiety. Chronic, untreated stress, could eventually lead to lifetime emotional disorder, and such cases happen more than 50% of the time (https://www.webmd.com/balance/guide/causes-of-stress#2). Stress could be overcome by deep breathing techniques, which is often overlooked because stress occurs unconsciously. By making people aware of stress through vibration on wrist sensors and providing relaxation techniques, chronic stress could be reduced. However, heart rate signals from wrist sensors contain noise generated from everyday movement. To estimate stress accurately, noise needs to be removed, which is the focus of the project.
For full details, refer to the paper (Kos et al., 2017) https://ieeexplore.ieee.org/document/8037141
Data on interbeat intervals (also known as RR intervals), denoting time between successive heart beats, are collected from wrist sensors (Empathica 4, E4, and Microsoft Band 2, MB), and ECG, used as the gold standard to compare measurements, from 9 healthy participants.
- Remove short interval and add it to the subsequent interval
- Impute long interval based on moving average within a certain window size
- Implement Singular Spectrum Analysis (SSA), a dimensional reduction technique, to smooth irregularities in heart rate signals and reconstruct function of heart rate over time
For full details, refer to the poster (Kos et al., 2017) https://www.researchgate.net/publication/320911791_The_Accuracy_of_Monitoring_Stress_from_Wearable_Devices_Background_and_Objective_References
After implementing the data cleaning algorithm, the RR interval of data from wrist sensors was correlated with that of ECG. The cleaned data had 10% higher correlation with ECG than the raw data, indicating that the algorithm improves heart rate estimates by 10%.
Future work would focus on stress detection algorithms based on cleaned heart rate signals.