Research Article: Low-back electromyography (EMG) data-driven load classification for dynamic lifting tasks

Date Published: February 15, 2018

Publisher: Public Library of Science

Author(s): Deema Totah, Lauro Ojeda, Daniel D. Johnson, Deanna Gates, Emily Mower Provost, Kira Barton, Zhan Li.

http://doi.org/10.1371/journal.pone.0192938

Abstract

Numerous devices have been designed to support the back during lifting tasks. To improve the utility of such devices, this research explores the use of preparatory muscle activity to classify muscle loading and initiate appropriate device activation. The goal of this study was to determine the earliest time window that enabled accurate load classification during a dynamic lifting task.

Nine subjects performed thirty symmetrical lifts, split evenly across three weight conditions (no-weight, 10-lbs and 24-lbs), while low-back muscle activity data was collected. Seven descriptive statistics features were extracted from 100 ms windows of data. A multinomial logistic regression (MLR) classifier was trained and tested, employing leave-one subject out cross-validation, to classify lifted load values. Dimensionality reduction was achieved through feature cross-correlation analysis and greedy feedforward selection. The time of full load support by the subject was defined as load-onset.

Regions of highest average classification accuracy started at 200 ms before until 200 ms after load-onset with average accuracies ranging from 80% (±10%) to 81% (±7%). The average recall for each class ranged from 69–92%.

These inter-subject classification results indicate that preparatory muscle activity can be leveraged to identify the intent to lift a weight up to 100 ms prior to load-onset. The high accuracies shown indicate the potential to utilize intent classification for assistive device applications.

Active assistive devices, e.g. exoskeletons, could prevent back injury by off-loading low-back muscles. Early intent classification allows more time for actuators to respond and integrate seamlessly with the user.

Partial Text

Back pain and injury are highly prevalent, often leading to missed work days and potentially debilitating problems [1, 2]. Lifting belts and passive lumbar supports commonly provide static methods for preventing such injuries in the workplace. Although these braces serve as postural reminders and limit the range of motion, they cannot dynamically off-load the lower back, thus limiting their effectiveness [3, 4].

Nine healthy subjects (4 female, 5 male; age: 24 (±2) yrs; weight: 148 (±22) lbs; height: 170 (±11) cm), with no history of chronic pain, participated in this study. The study was approved by the University of Michigan’s Institutional Review Board (IRB # HUM00075027) and all subjects provided written informed consent prior to their participation.

During the course of a lift, muscle activity increased as subjects bent forward to pick up the weights. The EMG signals indicated a spike in activity just before load-onset as subjects prepared to lift the weight off the table. As shown in Fig 3, the amplitude of the spike increased as the lifted load increased, with the ‘no-weight’ lifts showing significantly smaller spikes than the other two load categories. The data around the spike, i.e. near load-onset, resulted in the highest classification accuracy. The region with the highest average testing classification performance (see Fig 4) ranged from 200 ms before load-onset to 200 ms after load-onset, with average classification accuracy ranging from 80% (±10%) to 81% (±7%).

The earliest window when loads can be classified with 80% accuracy was −200 to −100 ms prior to load-onset, which corresponds with the electromechanical delay of the low-back erector spinae muscles at the L5 region [24]. This classification performance during pre-onset was not statistically significantly different from post-onset performance. Thus, it can be argued that loads can be classified 100 ms prior to full load support with as high of an accuracy as after load-onset. As mentioned in the Introduction, the earlier a load classification can be made, the more time a controller and actuator have to respond. Once the load is classified, a control command must be generated and sent to an actuator, which will provide the appropriate assistive torque to off-load the user’s muscles. With load classification completed 100 ms prior to load-onset, an actuator with a 100 ms or less response time would ensure that an assistive torque can be initiated prior to full load support by the user. A few common actuator response times are listed in Table 1, ranging from 37.5 to 315 ms. The 100 ms head start afforded by the early classification shown in this study gives us a significant advantage over manual force input or post-onset detection methods, which indicates greater potential for device actuation prior to significant user loading. The use of preparatory muscle activity pattern recognition brings us a step closer towards seamless integration of device and user for effective injury prevention.

This study found that an EMG signal can be used to classify lifted loads, as early as 100 ms prior to load-onset, i.e. when the load is fully supported by the subject. This finding encourages the use of myoelectric-based control strategies for assistive devices that aid weighted lifting. The results demonstrate that early load classification is possible, which will allow the controller more time to respond and meet optimal controller delay specifications for seamless integration of intent-recognition and actuated assistance. The classification generalized well among the healthy subject population, achieving average accuracies greater than 80% using inter-subject cross-validation and testing. Nevertheless, the generalizability of the results is limited by the small sample size and lack of diversity of the subject pool. Future work should include a more heterogeneous participant group (in terms of age and health/fitness level) to allow further extension of the results. Moreover, this initial study demonstrated the ability to classify only three discrete loads in the 0 to 24-lbs range. We expect these findings will extend toward classification of other loads, particularly as larger changes in loads might be more readily identifiable as differences in EMG patterns are expected to be more pronounced. Future work should include a more thorough exploration of optimal features, time windows, and classification schemes to ensure optimal classification accuracy, greater granularity in loading, as well as employing sensor fusion and Bayesian modeling. The classifier algorithm can be implemented online and has the potential for incorporation into real-time control of active assistive devices.

 

Source:

http://doi.org/10.1371/journal.pone.0192938

 

0 0 vote
Article Rating
Subscribe
Notify of
guest
0 Comments
Inline Feedbacks
View all comments