Comparison of Predictors to Estimate Lumbar Muscle Activity using a Single Inertial Sensor during Patient Handling

Abstract

Monitoring of lumbar muscle activity during patient handling is important for the preventing lower back pain among caregivers. Electromyography is used for the measurement of muscle activity; however, EMG has some uncomfortable. Therefore, our previous study proposed measurement method for lumbar muscle activity using wearable sensors and found that the combination of wearable sensors and machine learning can be applied to estimate muscle activity. However, optimal machine learning-based predictor has not been explored. In addition, our previous method requires multiple sensors. Thus, this study explores optimal predictors to estimate lumbar muscle activity during patient handling by a single inertial sensor. The proposed method estimates lumbar muscle activity by acceleration and angular velocity on a trunk and machine learning-based predictor. In the evaluation, five machine learning-based predictors (artificial neural network, gaussian process, k-nearest neighbor, linear regression, and support vector machine) estimated lumbar muscle activity during patient handling. Mean absolute error values of lumbar muscle activity between each predictor and EMG were obtained as actual values. Results show that the proposed method using k-nearest neighbor could estimate muscle activity with minimal error. These results suggest that the proposed method using k-nearest neighbor as an optimal predictor can measure lumbar muscle activity.