Effectiveness of a Batteryless and Wireless Wearable Sensor System for Identifying Bed and Chair Exits in Healthy Older People

Roberto Luis Shinmoto Torres, Renuka Visvanathan, Stephen Hoskins, Anton van den Hengel, Damith Chinthana Ranasinghe

TL;DR

This paper introduces a batteryless, wireless wearable sensor system using RFID technology to detect bed and chair exits in older adults in real-time. The system achieves high recall and precision, offering a promising, less intrusive alternative to existing fall prevention technologies.

What Problem Does This Paper Address?

Falls are a significant concern, especially for older adults, leading to injuries, increased hospital stays, and substantial costs. Current hospital systems often rely on localized bed and chair sensors that suffer from high false alarm rates, leading to "alarm fatigue" among caregivers. While wearable sensors offer a more personalized approach, existing systems are often bulky, battery-dependent, and uncomfortable for older individuals. This paper aims to address the need for an effective, unobtrusive, and real-time system for identifying critical activities like bed and chair exits, which are precursors to falls.

Methods

The core of the system is the Wearable Wireless Identification and Sensing Platform (W²ISP), a batteryless, passive RFID tag incorporating an accelerometer and a microcontroller. This device is worn over clothing, typically on the torso. Data from the W²ISP, including acceleration and Received Signal Strength Indicator (RSSI) from RFID readers, is fed into a multi-stage recognition approach:

  1. Feature Extraction: This stage extracts various features from the sensor data, including time-domain acceleration, RSSI, and contextual information derived from a sliding window.
  2. Activity Prediction: A dynamically weighted conditional random fields (dWCRF) model is employed. This is a probabilistic approach that accounts for class imbalance (e.g., longer durations of lying vs. short ambulation periods) and learns class-specific weights to improve prediction accuracy. It generates marginal probabilities for different activities (sitting on bed, sitting on chair, lying, ambulating).
  3. Activity Recognition: A score function is applied to the marginal probabilities over a 1-second window to smooth out noisy predictions and assign a dominant activity. This is followed by a finite state machine that triggers an alert upon detecting a bed or chair exit based on transitions between predicted activities.

The study also explores two different room configurations (Room 1 and Room 2) with varying RFID antenna placements to assess the impact of deployment on performance.

Datasets

The study utilized data collected from 14 healthy older participants (aged 65+) performing scripted daily living activities in two simulated hospital room settings (Room 1 and Room 2). The data was collected during supervised trials. The authors have made this dataset publicly available for future research.

Key Insights and Ideas

  • Batteryless and Wireless is Feasible: The W²ISP demonstrates that a batteryless and wireless wearable sensor can collect meaningful data for activity recognition, addressing comfort and maintenance concerns.
  • RFID as a Complementary Sensor: Combining accelerometer data with RSSI from RFID provides richer information, particularly for distinguishing postures that might appear similar in acceleration alone (e.g., sitting vs. standing).
  • dWCRF for Class Imbalance: The use of dWCRF is crucial for handling the inherent class imbalance in activity recognition data, where some activities are significantly longer or shorter than others.
  • Room Configuration Matters: The placement of RFID antennas significantly impacts performance, with a smaller deployment (Room 2) showing better results for bed exit recognition. This suggests that optimized antenna placement can lead to more efficient and effective systems.
  • Score Function for Robustness: The score function effectively smooths out noisy predictions from the activity predictor, reducing false alarms and improving the overall reliability of the system.

Key Results

The system achieved promising results in identifying bed and chair exits:

  • Recall (Sensitivity): Above 93% for Room 2 (bed exits) and 94% for Room 1 (chair exits).
  • Precision: Above 78% for Room 2 (bed exits) and 67% for Room 1 (chair exits).
  • F-score: Above 84% for Room 2 (bed exits) and 77% for Room 1 (chair exits).

Specifically, Room 2 demonstrated statistically significant higher performance for bed exit recognition (p < 0.001) with higher recall and precision, despite having fewer antennas. This suggests a more practical and cost-effective deployment.

The study also reported low average delays in recognition, with medians of 1.20 s (Room 1) and 1.25 s (Room 2) for bed exits, and 1.13 s (Room 1) and 0.00 s (Room 2) for chair exits.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Antenna Positioning: The performance is sensitive to antenna placement, and further optimization is needed, especially for chair exit recognition.
  • Sensor Placement: Placing the sensor on the shoulder is suggested as a future direction to reduce occlusions and improve data collection.
  • Trial Duration: The study used short trials; longer-term trials, including day and night, are needed to represent a full hospital day and evaluate long-term acceptability.
  • Larger and Hospitalized Cohort: Future studies should involve a larger sample of hospitalized older participants to validate the approach and assess its real-world applicability.
  • Machine Learning Enhancements: Investigating additional data sources and machine learning methods to better discriminate similar postures is recommended.
  • Textile Integration: Developing inconspicuous, textile-integrated sensors is a key area for future work to improve user comfort and acceptance.

Summary updated

BibTeX
@inproceedings{torres2016effectiveness,
  author = {Roberto Luis Shinmoto Torres and Renuka Visvanathan and Stephen Hoskins and Anton van den Hengel and Damith Chinthana Ranasinghe},
  title = {{Effectiveness of a Batteryless and Wireless Wearable Sensor System for Identifying Bed and Chair Exits in Healthy Older People}},
  booktitle = {Sensors},
  year = {2016},
  url = {https://openreview.net/forum?id=Rh7lVhlo8XR},
}
Figure