Table 5.
Model selection and IMU sensor position for estimating joint force/torque.
| Reference | Number of Sensors | Position | Frequency | Activation | Tasks | Optimal Model | Accuracy |
|---|---|---|---|---|---|---|---|
| (Stetter et al., 2019) [62] | 2 | Right thigh and shank | 1500 Hz | Walking, jumping | KJF | ANN | Pearson correlation coefficients: vertical KJF: 0.60–0.94, P KJF: 0.64–0.90, M-L KJF: 0.25–0.60. |
| (Lim et al., 2020) [16] | 1 | CoM | 100 Hz | Walking | Joint torques | ANN | The approximate errors: hip joint torques: 16.7 Nm, knee joint torques: 11.4 Nm, ankle joint torques: 15.3 Nm. |
| (Jiang et al., 2019) [65] | 2 | Shank, foot | 100 Hz | Walking | Ankle joint power | RF | Intra-subject test: R = 0.98, Inter-subject test: R = 0.92. |
| (Derie et al., 2020) [66] | 2 | Antero-medial side of both tibias | 1000 Hz | Running | Maximal vertical loading rate | XGB | Subject-dependent: mean absolute percentage error: 6.08%, Subject-independent: mean absolute percentage error: 11.09%. |
| (Lee and Park, 2020) [9] | 1 | Sacrum | 148 Hz | Walking | Joint torques | ANN | NRMSE: joint torques: 11.4–24.1% |
| (Stetter et al., 2020) [4] | 2 | Right thigh and shank | 1500 Hz | Walking, running, 45° cutting maneuver |
KFM, KAM | ANN | KFM: R = 0.74 ± 0.36, KAM: R = 0.39 ± 0.32. |
| (De Brabandere et al., 2020) [5] | 1 | Left hip | 50 Hz | Walking, walking upstairs/downstairs, sitting down and standing up, forward lunge and side lunging, standing on one leg, squatting on one leg | Hip moment | Regularized linear regression | Mean absolute error: left hip: 29%, right hip: 36%. |
| (Dorschky et al., 2020) [10] | 4 | Lower back, the right thigh, shank and foot | 1000 Hz | Walking, running | Joint moments | CNN | Pearson correlation coefficients: hip moment: 0.94, knee moment:0.975, ankle moment: 0.981. |
| (Mundt et al., 2020) [11] | 5 | Pelvis, thigh, shank | Virtual IMU data | Walking | Joint moments | MLP | The mean correlation of the models: r-kinetic-measured: 0.95, r-kinetic-combined: 0.95. |
| (Barua et al., 2021) [12] | 2 | Foot, shank | 100 Hz | Walking | Ankle joint power | LSTM | R > 81.25% |
| CNN | R > 83.09% | ||||||
| CNN-LSTM | R > 83.19% | ||||||
| (Chaaban et al., 2021) [60] | 4 acc, 4 gre |
Thigh, shank | 1125 Hz | Jumping | Knee extension moment, sagittal plane knee power absorption | Linear regression | RMSE: knee extension moment: 0.028 ± 0.0002 BW·HT, sagittal plane knee power: 0.27 ± 0.003 BW·HT. |
| 4 acc | RMSE: knee extension moment: 0.031 ± 0.0002 BW·HT sagittal plane knee power: 0.32 ± 0.003 BW·HT |
||||||
| (Mundt et al., 2021) [13] | 5 | Pelvis, thigh, shank | 100 Hz | Walking | Joint moments | MLP, LSTM, CNN |
Mean model correlation coefficients: joint moment > 0.939. |
| (Molinaro et al., 2022) [61] | 3 | Trunk, thigh, and hip | Virtual IMU data | Walking | Hip moment | TCN | Average RMSE: steady-state ambulation: 0.131 ± 0.018 Nm/kg, mode transitions: 0.152 ± 0.027 Nm/kg. |
| (Hossain et al., 2023) [63] | 3 | Thigh, shank, and foot | 100 Hz | Tread-mill walking, level-ground walking, ramp ascent/descent, and stair ascent/descent |
Hip, knee, and ankle joint moment, 3D GRFs | Hybrid model based on 1D, 2D convolutional, GRU, and dense layers with the application of bagging techniques | PCC: 0.923 ± 0.030 |
| 8 | Trunk, pelvis, and both thighs, shanks | 100 Hz | Walking | KFM, KAM, and 3D GRFs | PCC: 0.884 ± 0.029 |
KJF: Knee joint force, KFM: Knee flexion moment, KAM: Knee adduction moment, NRMSE: Normalized root mean square error, RMSE: Root mean square error, TCN: Temporal convolutional network.