Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Orthopedic and Rehabilitation Engineering

Orthopedic and Rehabilitation Engineering - Poster Session B

Poster K10 - Developing a Preliminary Fall Classification Algorithm (FCA) for Perturbations While Wearing a Robotic Exoskeleton

Thursday, October 24, 2024

2:45 PM - 3:45 PM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Benjamin D. Tondre (he/him/his)

Student/Intern
University of Florida
Tampa, Florida, United States

Co-Author(s)

MF

Marshaun Fitzpatrick

Graduate Assistant
Advanced Platform Technology Center, United States
Musa Audu, PhD

Research Engineer and Investigator
Advanced Platform Technology Center, United States

Primary Investigator(s)

Sandra Hnat, DEng (she/her/hers)

Biomedical Engineer (Research)
Louis Stokes Cleveland VA Medical Center, United States

Last Author(s)

RT

Ronald Triolo

Prof Biomedical Eng CWRU; Executive Dir. VA APTC;VA Senior Research Career Scientist; PI at ReLLiNC
Case Western Reserve University, VA Center for Advanced Platform Technology (APTC), United States

Introduction: Commercial powered exoskeletons are wearable walking robots with motorized joints that improve mobility for individuals with lower-limb paralysis by controlling joint movement in predefined kinematic trajectories. While these devices have proven to show reduced muscle atrophy and stable walking, commercial models lack controllers to prevent or correct postural instability or falls. Current fall-prevention strategies involve locking or damping joints when trunk flexion or extension exceeds a predefined threshold as detected by a pelvic tilt sensor. However, without intervention from a trained caregiver or spotter, falls could still occur, thereby confining exoskeleton use to controlled environments and restricting their usability in everyday, unmonitored settings.
To remove this limitation and improve the overall usability and safety of exoskeletons, it is critical to first detect and classify fall direction as quickly as possible to allow for the exoskeleton to reconfigure itself to prevent the fall from occurring. This project involves the creation and validation of a Fall Classification Algorithm (FCA) trained on kinematic data collected from inertial measurement units (IMUs) during neurotypical standing treadmill perturbation experiments while participants wore a custom Motor-Assisted Hybrid Neuroprosthesis (MAHNP). Participant signed a written consent form approved by the Institutional Review Board of the Louis Stokes Cleveland Veteran Affairs Medical Center (#16027-H18).

Materials and

Methods: We collected data from one neurotypical participant (28, 1.7526m,81.647kg,male) wore the MAHNP set in passive mode to allow the joints to move freely. The participant wore a network of 6 external IMUs (Adafruit, BNO085): placed on the left/right thigh, left/right shank on the exoskeleton, as well as on the participant’s upper and lower torso. The subject stood on a split-belt instrumented treadmill (Fit 5, Bertec Corporation, Columbus, OH, USA) and wore a safety harness that did not restrict movements but ensured safety by preventing falls. The IMU data was recorded via a Speedgoat IO505 Serial I/O Module (Speedgoat RealTime Target Machine Speedgoat, Köniz, Switzerland) at 200 Hz interfaced with a MATLAB Simulink controller (Mathworks, Natick, Massachusetts, USA). The experiment protocol is shown in Figure 1.

A block diagram of our approach for training the FCA using the Optimized Efficient Linear algorithm is shown in Figure 2. IMU trial data was sampled at four different durations, with each duration generating a distinct feature table containing 240 features. These feature tables (FT) were used to train FCA models using 50 k-fold cross-validation in the MATLAB Classification App. FCA models were then tested and ranked based on their Validation Accuracy (VA) and Test Accuracy (TA).

Results, Conclusions, and Discussions:

Results:
Confusion Matrices (Fig 2a - Fig 2d) showing True Positive Rate (TPR) and False Negative Rate (FNR) for Optimized Efficient Linear (OEL) FCAs trained with the 200ms FT.
Conclusions and

Discussion:
We trained a Fall Classification Algorithm (FCA) to predict fall direction from IMU data collected during exoskeletal standing perturbation experiments. Our results indicate that our FCA can make highly accurate predictions based on IMU kinematic data within at least a half-second time window from which the perturbation is experienced, even showing an overall test accuracy of 95.15% with data from only the 50ms duration. Furthermore, the time window durations of 200ms and 350ms yielded the most accurate and robust models, suggesting that the optimal duration for data analysis and model training likely lies between these values. While this may not directly correlate with prediction speed, it highly correlates with prediction accuracy during a fall.

The participant’s reactive movements to the prescribed perturbations were consistent and repeatable, mirrored between the left and right legs. The perturbation direction corresponded directly to the initial fall direction; for example, a Mid-Stance Forward Right (MSFR) perturbation resulted in a forward right fall. This consistency is promising for the future development of the balance controller, as these reactions likely correlate to the same movements and positions needed to move the exoskeleton to recover during a fall.

Future work involves real-time implementation and validation testing of the FCA via MATLAB Simulink and Speedgoat controller. Additionally, a complete dataset for all neurotypical participants will be necessary to further improve the robustness of the FCA. Finally, additional controller logic must determine when to activate the FCA based on a threshold value in application. Lastly, an idle standing or walking condition may be implemented into a future iteration of the FCA for a more sophisticated controller.

Acknowledgements (Optional): This work has been supported by the Wen H. Ko Summer Internship Program at the Advanced Platform Technology (APT) Center and the facilities at the Motion Study Laboratory of the Louis Stokes Cleveland VA Medical Center.