Data Distribution Dynamics in Real-World WiFi-Based Patient Activity Monitoring for Home Healthcare

Mahathir Monjur, Jia Liu, Jingye Xu, Yuntong Zhang, Xiaomeng Wang, Chengdong Li, Hyejin Park, Wei Wang, Karl Shieh, Sirajum Munir, Jing Wang, Lixin Song, Shahriar Nirjon

Abstract

This paper examines the application of WiFi signals for real-world monitoring of daily activities in home healthcare scenarios. While the state-of-the-art of WiFi-based activity recognition is promising in lab environments, challenges arise in real-world settings due to environmental, subject, and system configuration variables, affecting accuracy and adaptability. The research involved deploying systems in various settings and analyzing data shifts. It aims to guide realistic development of robust, context-aware WiFi sensing systems for elderly care. The findings suggest a shift in WiFi-based activity sensing, bridging the gap between academic research and practical applications, enhancing life quality through technology.

Introduction

WiFi-based activity recognition represents a significant opportunity for non-intrusive home healthcare. By utilizing Channel State Information (CSI) to detect movement, systems can monitor patients without the privacy concerns of cameras or the compliance issues associated with wearable devices. However, a substantial performance gap exists between controlled laboratory environments and "in the wild" deployments. This gap is primarily driven by dataset shift—variations in signal patterns caused by different room layouts, furniture, and individual physiological differences. This research quantifies these shifts across eight real-world environments and evaluates strategies to mitigate performance degradation.

The AURA System and Methodology

To address the challenges of real-world deployment, the researchers developed AURA, a system designed for resource-constrained edge hardware like the NVIDIA Jetson Nano. The architecture is optimized for both spatial and temporal feature extraction:

  • Hybrid Architecture: The model utilizes three convolutional layers to extract spatial features from WiFi sub-channels, followed by three stacked bidirectional Long Short-Term Memory (Bi-LSTM) layers to capture the temporal progression of activities.
  • Domain Adaptation: To combat the "cliff" in accuracy when moving to new environments, two primary strategies were implemented:
  • Gradient Reversal Layer (GRL): An adversarial training component that forces the model to learn features that are invariant to the specific environment, focusing instead on the underlying activity signatures.
  • k-shot Adaptation: A fine-tuning approach where the pre-trained model is updated using a very small sample size (e.g., 5 samples per activity) from the target environment.

Data Collection and Experimental Setup

The study utilized a multi-modal dataset to ensure high-fidelity ground truth. Data was collected from 16 participants across 8 distinct environments, including kitchens, living rooms, and offices. The dataset comprises approximately 7.4 million WiFi CSI packets and 10 hours of synchronized depth camera footage (Intel RealSense D435) used for precise labeling. This robust dataset allowed the team to visualize data clustering using t-SNE embeddings, which confirmed that different environments create distinct "data islands" that typically confuse standard classifiers.

Analysis of Dataset Shifts

The research identified three primary dimensions of data distribution dynamics:

  1. Environmental Shift: This is the most significant hurdle. Moving the system to a new room can cause accuracy to drop from 82% to as low as 42%. Complex environments with high interference (like cluttered living rooms) pose the greatest challenge.
  2. Subject Shift: Differences in body mass and height affect WiFi reflections. However, the performance drop when encountering a new person in a familiar room (12-20%) is significantly less severe than the drop caused by a change in environment.
  3. Temporal Stability: The study found that if the environment and subject remain constant, the WiFi signature of an activity is remarkably stable over time, suggesting that once a system is calibrated for a specific home, it does not require frequent retraining.

Results and Discussion

The baseline model achieved 82% accuracy in seen environments. While this crashed to 42% in unseen environments, the application of GRL boosted performance to 63%. When combining GRL with k-shot adaptation, the performance gap for unseen individuals was narrowed to a 2-10% disparity.

Furthermore, the model is highly efficient, with a footprint of only 6.76 MB and 1.77 million parameters. This makes it viable for deployment on consumer-grade IoT hubs or home routers, bringing sophisticated healthcare monitoring closer to practical reality.

Conclusion and Future Work

While domain adaptation techniques like GRL and k-shot learning significantly improve model robustness, a 10-30% accuracy gap remains between seen and unseen environments. Future research will explore hyper-network-based approaches to parameterize the environment itself, allowing the model to dynamically subtract environmental "noise." Additionally, improving the detection of subtle movements—such as drinking or small arm exercises—remains a priority for enhancing the granularity of home healthcare monitoring.

Summary updated

Abstract

This paper examines the application of WiFi signals for real-world monitoring of daily activities in home healthcare scenarios. While the state-of-the-art of WiFi-based activity recognition is promising in lab environments, challenges arise in real-world settings due to environmental, subject, and system configuration variables, affecting accuracy and adaptability. The research involved deploying systems in various settings and analyzing data shifts. It aims to guide realistic development of robust, context-aware WiFi sensing systems for elderly care. The findings suggest a shift in WiFi-based activity sensing, bridging the gap between academic research and practical applications, enhancing life quality through technology.
BibTeX
@inproceedings{monjur2024data,
  author = {Mahathir Monjur and Jia Liu and Jingye Xu and Yuntong Zhang and Xiaomeng Wang and Chengdong Li and Hyejin Park and Wei Wang and Karl Shieh and Sirajum Munir and Jing Wang and Lixin Song and Shahriar Nirjon},
  title = {{Data Distribution Dynamics in Real-World WiFi-Based Patient Activity Monitoring for Home Healthcare}},
  booktitle = {ICHI},
  year = {2024},
  url = {https://openreview.net/forum?id=HqxzqNghCH},
}
Figure