Shoulder problems (pain and pathology) are highly prevalent in manual wheelchair users with spinal cord injury. These problems lead to limitations in activities of daily life (ADL), labor- and leisure participation, and increase the health care costs. Shoulder problems are often associated with the long-term reliance on the upper limbs, and the accompanying “shoulder load”. To make an estimation of daily shoulder load, it is crucial to know which ADL are performed and how these are executed in the free-living environment (in terms of magnitude, frequency, and duration). The aim of this study was to develop and validate methodology for the classification of wheelchair related shoulder loading ADL (SL-ADL) from wearable sensor data. Ten able bodied participants equipped with five Shimmer sensors on a wheelchair and upper extremity performed eight relevant SL-ADL. Deep learning networks using bidirectional long short-term memory networks were trained on sensor data (acceleration, gyroscope signals and EMG), using video annotated activities as the target. Overall, the trained algorithm performed well, with an accuracy of 98% and specificity of 99%. When reducing the input for training the network to data from only one sensor, the overall performance decreased to around 80% for all performance measures. The use of only forearm sensor data led to a better performance than the use of the upper arm sensor data. It can be concluded that a generalizable algorithm could be trained by a deep learning network to classify wheelchair related SL-ADL from the wearable sensor data.

Introduction Shoulder problems like pain or pathology are highly prevalent in manual wheelchair users (MWU) with spinal cord injury, lead to limitations in participation, a reduced quality of life and are often associated with “shoulder overload” (Mercer et al., 2006). Laboratory based experiments have examined shoulder load for a variety of wheelchair related activities (WRA; van Drongelen et al., 2005), but no research has been conducted that investigates the actual shoulder load profile of MWU’s in daily conditions. Such a profile would help in understanding the relation of shoulder overload and shoulder problems and could thereby support clinical decision making in the prevention of shoulder problems. Inspired by recent work on machine learning (ML) prediction of joint load from wearable sensors, the current project's ultimate aim is to develop a generalizable ML algorithm that can predict shoulder load for a variety of WRA. Methods 10 able bodied participants were trained before the actual measurement of the WRA of interest: wheelchair propulsion (WCP) at 0.56 and 1.1 m/s at 0%, 0.56 m/s at 6%, WCP in restricted space, short ramp 12% up and down, weight relieve lift, manual material handling with 2 kg and desk work. Participants were equipped with five Shimmer3 sensors on wheelchair, wheel, thorax, upper arm and forearm; EMG was collected from the biceps and medial deltoid muscles. An 8 camera Qualisys system was used to obtain kinematics conform ISB recommendations (Wu et al., 2005). A SmartWheel (Out-Front) for collecting propulsion kinetics replaced the original wheel of a standard Kuschal wheelchair. From laboratory kinematics and kinetics 3D shoulder joint reaction forces (JRF) were estimated with an OpenSim based musculoskeletal model (Wu et al., 2016) (MSM), which consequently served as target for the training of a variety of ML algorithms, using sensor data (acceleration, angular velocity, EMG) as input. Results & Discussion Starting with simple input (upper arm sensors only) and a linear neural network (NN, one input, one hidden, one output layer for total JRF; iteratively trained following a Leave One Subject Out approach, a R2 of around 60% was obtained between predicted JRF (NN) and target (MSM output), but results varied considerably over participants. In a next phase the complexity of the ML models was increased to deep learning models (recurrent NN) and more signals (e.g. forearm and thorax sensors) were added to the input, which, however, did not improve the overall performance considerably. Currently, it is explored whether using training the algorithms on individual datasets, for single tasks, can improve the performance. The explorative process will be presented and discussed in the light of the relevant results. References Mercer, J. L., Boninger, M., Koontz, A., Ren, D., Dyson-Hudson, T., & Cooper, R. (2006). Shoulder joint kinetics and pathology in manual wheelchair users. Clincal Biomechanics, 21(8), 781-789. https://doi.org/10.1016/j.clinbiomech.2006.04.010 van Drongelen, S., van der Woude, L. H., Janssen, T. W., Angenot, E. L., Chadwick, E. K., & Veeger, D. H. (2005). Glenohumeral contact forces and muscle forces evaluated in wheelchair-related activities of daily living in able-bodied subjects versus subjects with paraplegia and tetraplegia. Archives of Physical Medicine and Rehabilitation, 86(7), 1434-1440. https://doi.org/10.1016/j.apmr.2005.03.014 Wu, W., Lee, P. V. S., Bryant, A. L., Galea, M., & Ackland, D. C. (2016). Subject-specific musculoskeletal modeling in the evaluation of shoulder muscle and joint function. Journal of Biomechanics, 49(15), 3626-3634. https://doi.org/10.1016/j.jbiomech.2016.09.025 Wu, G., van der Helm, F. C., Veeger, H. E., Makhsous, M., Van Roy, P., Anglin, C., Nagels, J., Karduna, A. R., McQuade, K., Wang, X., Werner, F. W., Buchholz, B., & International Society of, B. (2005). ISB recommendation on definitions of joint coordinate systems of various joints for the reporting of human joint motion-Part II: Shoulder, elbow, wrist and hand. Journal of Biomechanics, 38(5), 981-992. https://doi.org/10.1016/j.jbiomech.2004.05.042

There is a high prevalence of shoulder problems in manual wheelchair users (MWUs) with a spinal cord injury. How shoulder load relates to shoulder problems remains unclear. This study aimed to develop a machine-learning-based methodology to estimate the shoulder load in wheelchair-related activities of daily living using wearable sensors. Ten able-bodied participants equipped with five inertial measurement units (IMU) on their thorax, right arm, and wheelchair performed activities exemplary of daily life of MWUs. Electromyography (EMG) was recorded from the long head of the biceps and medial part of the deltoid. A neural network was trained to predict the shoulder load based on IMU and EMG data. Different cross-validation strategies, sensor setups, and model architectures were examined. The predicted shoulder load was compared to the shoulder load determined with musculoskeletal modeling. A subject-specific biLSTM model trained on a sparse sensor setup yielded the most promising results (mean correlation coefficient = 0.74 ± 0.14, relative root-mean-squared error = 8.93% ± 2.49%). The shoulder-load profiles had a mean similarity of 0.84 ± 0.10 over all activities. This study demonstrates the feasibility of using wearable sensors and neural networks to estimate the shoulder load in wheelchair-related activities of daily living.