TLDR: A new research paper demonstrates, for the first time, robust recognition of six distinct human activities (sitting, standing, walking, skipping, running, climbing stairs) using only electrocardiogram (ECG) data. Three deep learning models were developed, with a CNNTransformer hybrid achieving the best accuracy of 72% for unseen subjects. This breakthrough paves the way for simplified wearable health devices that can simultaneously monitor cardiac activity and recognize physical movements without needing additional motion sensors, offering significant potential for personalized healthcare.
Human activity recognition (HAR) is a vital technology for applications like early health intervention and advanced health analytics. Traditionally, HAR systems have relied on inertial measurement units (IMUs), which are effective but often come with drawbacks such as high resource consumption and the need for frequent calibration. While electrocardiogram (ECG)-based methods have been explored, they typically served as supplementary tools to IMUs or were limited to very broad classifications, like distinguishing between active and inactive states, or simply detecting falls.
A groundbreaking study titled Human Activity Recognition Based on Electrocardiogram Data Only by Sina Montazeri, Waltenegus Dargie, Yunhe Feng, and Kewei Sha, marks a significant advancement in this field. For the first time, researchers have demonstrated robust recognition of six distinct physical activities using only ECG data, moving beyond the scope of previous work that often required additional motion sensors.
The research team designed and evaluated three innovative deep learning models tailored for ECG-only activity classification. These include a CNN classifier enhanced with Squeeze-and-Excitation blocks for refining feature importance, a ResNet classifier utilizing dilated convolutions to capture temporal dependencies across multiple scales, and a novel CNNTransformer hybrid. The hybrid model ingeniously combines the strengths of convolutional networks for local feature extraction with the attention mechanisms of Transformers for modeling long-range temporal relationships in cardiac signals.
The methodology involved collecting data from 54 subjects performing six activities: sitting, standing, walking, skipping, running, and climbing stairs. The ECG data underwent meticulous preprocessing, including filtering, downsampling, and Empirical Mode Decomposition (EMD) to extract rich features from various frequency components of the physiological signals. To rigorously test the models’ ability to generalize, the data was split subject-wise, meaning some subjects were entirely unseen during training.
The results were compelling. All three models achieved over 94% accuracy for subjects included in the training population. However, the true test of generalization came with unseen subjects. Here, the CNNTransformer hybrid emerged as the top performer, achieving an impressive 72% accuracy. The ResNet followed with 67%, and the CNN with 61%. This significant lead of the CNNTransformer highlights its superior capability in adapting to the physiological variability between different individuals.
Further investigation into the impact of training set size revealed that model performance consistently improved with more diverse training data, without reaching a saturation point even with the full dataset. This suggests that larger and more varied training populations would likely yield even better results. A critical threshold for achieving clinically acceptable performance was identified around 49% of the available training subjects, below which models exhibited unstable and unreliable behavior.
Also Read:
- PPGFlowECG: Transforming PPG Signals into Diagnostic-Grade ECG
- SelfMIS: Enhancing Heart Attack Detection from Single-Lead ECGs Through Latent Space Alignment
Implications for Future Wearables and Healthcare
The success of ECG-only activity recognition has profound implications for the next generation of wearable health devices. By eliminating the need for additional motion sensors, devices can be simplified, reducing complexity, power consumption, and cost. This breakthrough enables simultaneous cardiac monitoring and activity recognition from a single sensor, offering enhanced clinical value for patients, especially those with cardiovascular conditions who require continuous ECG surveillance.
For instance, cardiac rehabilitation programs could greatly benefit from integrated monitoring of both cardiac status and physical activity levels. Remote patient monitoring systems could provide more comprehensive physiological evaluations, leading to better-informed clinical interpretations of cardiac events within the context of a patient’s daily activities.
While this study represents a transformative advance, the authors acknowledge limitations, including a predominantly young and healthy subject population and controlled experimental conditions. Future research will focus on advanced motion artifact removal, identifying persistent cardiac features across diverse demographics, and conducting long-term validation in real-world environments with varied patient populations. Personalized models using transfer learning and integrating cardiac physiology domain knowledge are also promising avenues for further improvement.
In conclusion, this research establishes a new paradigm for wearable health monitoring, demonstrating the feasibility and robustness of classifying multiple distinct physical activities using only ECG data. It paves the way for simpler, more efficient, and more integrated health monitoring solutions that promise substantial benefits for personalized healthcare and remote patient management.
