Derived spine loads in response to multiple risk factors.
Marras-WS; Karwowski-W; Zurada-JM; Davis-KG
Atlanta, GA: U.S. Department of Health and Human Services, Public Health Service, Centers for Disease Control and Prevention, National Institute for Occupational Safety and Health, R01-OH-007787, 2008 Jan; :1-65
Low back disorders are extensive and extremely expensive for industry. As such, there is a need for better techniques that can assess structural loading in the industrial setting. To date, investigation of the loading on the back has been with simple static models that neglected the complex coactivity patterns of the trunk musculature. Thus, the objective of the current project was to develop a neuron-fuzzy engine that could predict the muscle activation pattern for common lifting conditions. The predicted muscle activation data can then be input into the EMG-assisted model to predict the spine loads. The first specific aim was to develop an artificial neural network model in the form of a multi-stage hybrid neuro-fuzzy "engine"-HNFE for electromyography (EMG) signal estimation was built using kinematic, kinetic, anthropometric, and work condition variables as inputs including physical and psychosocial characteristics. A complex engine was developed using fuzzy average with fuzzy cluster distribution techniques in combination with neural network structure. In order to identify inputs that have significant influence on the output, a method using fuzzy average with fuzzy cluster distribution (FAFCD) was utilized. The FAFCD method allows for the reduction of a high dimensional input space so that more effective models for EMG estimation can be built. After key work condition variables affecting EMG in lifting tasks were found using this method, a novel structure of feed forward neural network was utilized to estimate the instantaneous EMG by evaluating the full lifting motion at one time rather than estimating one sampling point at a time. The complete neural network model accounts for both global and local features of the input data. The resulting neural network model has the capability of predicting muscle activity from the input variables: kinematic, kinetic, and anthropometric factors under a wide variety of lifting conditions. For all muscles, the overall average MAE is 7.56%; the overall average R-square is 0.5437. Thus, the model was designed to be accurately and robust with respect to predicting the muscle activity values under realistic lifting conditions. The second specific aim was to establish the large database that was utilized to develop the multi-stage hybrid neuro-fuzzy "engine". The entire data set is both extensive and robust with respect to the parameter space. While not all work conditions imaginable were incorporated into the training sets utilized to develop the HNFE, the boundaries of the data set encompasses the majority of lifting conditions found in the workplace. The conditions that were utilized to train the model serve as the boundary conditions. The model can then interpolate between the conditions to predict muscle activity of new conditions, reducing the necessity to collect an all encompassing data set. The structure of the model also allows the model to be continually trained when new data is acquired. The third specific aim was to develop software that will link the HFNE to a website to allow for users to predict the muscle activities based on their data that could be utilized for the prediction of spine loads with their own spine load model or with the OSU EMG-assisted spine load model. A software program has been linked to a website where users can predict muscle activity from the ten trunk muscles during specific work conditions as well as upload new data to train the fuzzy engine. The fourth specific aim was to link the HFNE to the OSU EMG-assisted model and compared the loads predicted from the original EMG values to the predicted HFNE EMG values. The model fidelity was actually improved with the predicted EMG as compared to the actual EMG with improved r-square and average absolute error values. Furthermore, the three-dimensional spine loads were almost identical for the predicted EMG as compared to the actual EMG (within 35 N in each plane). The compression forces predicted within 1% while shear forces were within 11 %. In all, the project develop a very robust model that can now predict muscle activity for lifting conditions in industry where previous techniques were significantly limited to lack of technology, need for expensive equipment, and interference from surrounding machinery. Thus, by integrating the HNFE into a state-of-the-art spine loading model, it will now be possible to accurately estimate the loads on the spine in real world lifting conditions.
Models; Computer-models; Mathematical-models; Biomechanical-modeling; Biomechanics; Musculoskeletal-system; Manual-lifting; Materials-handling; Manual-materials-handling
William S. Marras, Ph.D., Biodynamics Laboratory, The Ohio State University, 1971 Neil Avenue, Columbus OR 43210
Final Grant Report
NTIS Accession No.
National Institute for Occupational Safety and Health
Ohio State University