Estimation of muscle forces based on a multi-DOF EMG-driven neuromusculoskeletal modeling approach: Impact of different EMG normalization strategies

Abstract

Introduction:Normalization of EMG signals amplitude is a critical step within EMG-driven neuromusculoskeletal (NMS) models [1].Maximal voluntary contraction (MVC) is themost commonnormalizationmethodusedwhen recurring to single-DOFNMSmodels for muscle forces estimation. Applications about multiple DOF are, instead, still largely unexplored. In the work of Barret et al. [2], EMG data were normalized to the maximum values obtained duringwalking trials (WTN–WalkingTrialsNormalization). Feasibility of such approach would be interesting in a clinical context, where patients may be not able to perform MVC. Therefore, the WTN method was compared to the one based on MVC trials (MVCTN). This study aims at determining the reliability of the WTN method for normalizing gait EMGs, by evaluating its impact on the calibration and open-loop predictions of a multi-DOF EMG-driven NMS model [3]. Methods: The workflow included four steps. 1. Collecting movement data. Two healthy subjects (mean age: 44.0±22.6, height: 174.5±7.8 cm, mass: 75.0±24.0 kg) were enrolled with informed consent. 3D marker trajectories, ground reaction forces and EMG signals from 15 lower limb muscles were simultaneously recorded. The dataset included a static pose, 10 walking trials and MVCs. 2. Processing experimental data. Motion data were processed using MOtoNMS (https://simtk.org/home/motonms/) and EMG envelopes were normalized applying the WTN and MVCTN strategies. 3.Musculoskeletal simulation. The OpenSim software (https://simtk.org/home/opensim)wasused to scale ageneric musculoskeletal model and obtain joint angles, joint moments, musculotendon lengths and moment arms. 4. Calibrating and executing the multi-DOF EMG-driven NMS model. The CEINMS software (https://simtk.org/home/ceinms) was used as implementation of the EMG-driven NMSmodel [3]. Flexion-extension of the hip, knee and ankle joints were considered. The model was calibrated and then executed on different walking trials. The whole procedure was repeated twice, using as input the EMG envelopes normalized with the two strategies. Data analysis included comparison of joint moments about the 3DOF, as predictedwith theNMSmodel versus inverse dynamics, and mean muscle forces. Results: Significant differences in muscle forces were obtained for muscles with different activation levels in the two cases (Fig. 1). This discrepancy between the two methods occurred despite comparable results being obtained in joint moments estimation (hip: R2 =0.84, %MAE=9.39 (MVCTN) R2 =0.83, %MAE=9.54 (WTN); knee: R2 =0.91, %MAE=7.81 (MVCTN) R2 =0.87, %MAE=8.18 (WTN); ankle: R2 =0.95, %MAE=6.09 (MVCTN) R2 =0.94, %MAE=6.24 (WTN)). Discussion: This work shows how different EMG normalization techniques can lead todifferent results for amulti-DOFEMG-driven NMS model. We hypothesized that the overall higher muscle activation level resulting from theWTN normalization with respect to theMVCTNmethod,would lead todifferences inmuscle forces estimation, despite agreement in thepredictions of jointmoments. Our results confirm this hypothesis, suggesting that the WTN method Fig. 1. Comparison of mean muscle force estimated for the rectus femoris muscle with CEINMS, using MVC (blue) and walking (red) trials for normalization of input EMG signals.