A Pilot Study on Falling-Risk Detection Method Based on Postural Perturbation Evoked Potential Features

Shenglong Jiang, Hongzhi Qi, Jie Zhang, Shufeng Zhang, Rui Xu, Yuan Liu, Lin Meng, Dong Ming2019 · DOI

TL;DR

This paper introduces a novel method for detecting falling risk in human-robot systems by analyzing brain activity (EEG) in response to unexpected postural perturbations. By focusing on "postural perturbation evoked potentials," the researchers achieved high detection rates with short latencies, demonstrating the potential of EEG as a real-time indicator of impending falls, even across different perturbation types.

What Problem Does This Paper Address?

In human-robot hybrid systems, errors in the robot's motor execution can lead to a fall risk that is difficult for the system to detect on its own. While humans are adept at sensing these risks through their central nervous system, current falling-risk detection methods primarily rely on external sensors like computer vision or kinematic sensors. These methods often lag behind the actual fall event. This paper addresses the gap by proposing a brain-computer interface (BCI) approach that leverages the human's internal neural response to detect falling risk before a fall occurs, aiming for a more proactive and integrated safety system. The authors specifically target motor-dysfunctional patients, for whom falling is a significant concern and data collection is challenging.

Methods

The core of this study lies in analyzing electroencephalogram (EEG) signals to detect falling risk. The researchers simulated unpredictable postural perturbations by rapidly deflating an airbag, causing subjects to lose balance.

  • EEG Data Acquisition: 60-channel EEG was recorded at 1000 Hz, with a 0-200 Hz bandpass filter.
  • Artifact Removal: Measures were taken to minimize motion artifacts, event expectation, and sound stimulation, including bandages, filtering, unpredictable event timing, and noise silencing.
  • Feature Extraction: EEG signals were filtered (0.1-10 Hz), down-sampled to 50 Hz, and segmented into epochs around the perturbation onset. The xDAWN algorithm was employed to enhance the signal-to-noise ratio of postural perturbation evoked potentials.
  • Pattern Recognition: Bayesian Linear Discriminant Analysis (BLDA) was used as the classifier, with regularization to prevent overfitting. The classifier threshold was subject-specific.
  • Cross-Task Recognition: To assess generalization, a classifier trained on one type of perturbation (e.g., left-side) was tested on the other (e.g., right-side) to simulate detecting unknown events.

Datasets

The study utilized data collected from 15 healthy subjects (23-28 years old, 50-75 kg). Each subject underwent 50 simulated falling-risk events. The data was split into training and testing sets (45/5 trials per subject). Data augmentation techniques were used for the training set. The study also evaluated the model's generalization performance by training on one type of perturbation (e.g., left-side) and testing on the other (right-side), effectively creating a cross-task dataset for evaluation.

Key Insights and Ideas

  • Neural Signatures of Falling Risk: Unexpected postural perturbations elicit a distinct and rapid neural response, specifically a strong negative peak (N1 potential) in the EEG signal at frontal-central channels (around FCZ) within 100 ms of the event. This neural signature is significantly more pronounced than those observed for typical sensory or auditory evoked potentials.
  • EEG as a Proactive Detector: The identified neural features are not just indicators of a fall happening, but can predict the risk of a fall with sufficient lead time to intervene. The short latency of these evoked potentials is crucial for real-time detection.
  • Robustness Across Perturbation Types: The study demonstrated that a classifier trained on one type of postural perturbation (e.g., left-side) could effectively detect another, unlearned type (e.g., right-side) with high accuracy. This "cross-task recognition" capability is vital for real-world applications where fall events are unpredictable and varied.
  • Potential for Personalized and Generic Systems: The method shows promise for both subject-specific models (high detection rates) and a more generalized model that can adapt to different individuals and perturbation scenarios, which is particularly important for patients with motor impairments.

Key Results

  • High Detection Rate: The falling-risk detection method achieved a high averaged detection rate of 98.67%.
  • Short Detection Latency: The detection latency, defined as the time to reach a detection rate beyond 90%, was 334 ms on average.
  • Low False Alarm Rate: The averaged false alarm rate in the rest state (non-perturbation) was 4.69%, crucial for avoiding unnecessary interventions.
  • Cross-Task Detection Performance: When trained on typical perturbation events, the classifier achieved a 94.2% detection rate for unlearned atypical perturbation events, with a low averaged false alarm rate (4.7% and 8.6%).

Limitations and Future Work

  • Pilot Study on Healthy Subjects: The study was conducted on healthy subjects. While promising, further validation is needed on populations with motor impairments (e.g., SCI patients) who are at higher risk of falls.
  • Simulated Environment: The falling-risk events were simulated in a controlled laboratory setting. Real-world environments present more complex and varied challenges.
  • Generalization to Real-World Falls: While cross-task recognition was demonstrated, extending this to detect naturally occurring falls in diverse daily activities remains a challenge.
  • Feature Selection and Portability: The authors suggest that further research into feature selection methods could reduce EEG channel numbers, improving device portability and processing speed for practical applications.
  • Integration with Intervention: Future work could explore integrating this detection system with robotic interventions or assistive devices to actively prevent falls.

Summary updated

BibTeX
@misc{jiang2019pilot,
  author = {Shenglong Jiang and Hongzhi Qi and Jie Zhang and Shufeng Zhang and Rui Xu and Yuan Liu and Lin Meng and Dong Ming},
  title = {{A Pilot Study on Falling-Risk Detection Method Based on Postural Perturbation Evoked Potential Features}},
  year = {2019},
  doi = {10.3390/s19245554},
}
Figure