Table 4.
The detailed machine learning approaches.
| References | Predictor | Data preprocessing | Dataset split | CV | Machine learning approaches | Response | Evaluation | Performance |
Ground truth/
biomechanical model reference data |
|---|---|---|---|---|---|---|---|---|---|
| Stetter et al. (2020) | Tri-axial linear acceleration and tri-axial angular velocity | Filtered by a 4th order Butterworth filter with cut-off frequency of 15 Hz; IMUs signals were interpolated to keep the same sample frequency with knee joints moments data | / | LOSOCV | ANN (two hidden layers, 100 and 20 neurons) | External knee flexion and adduction moments | R2, rRMSE, RMSE | Knee flexion moment: Moderate running: R2 = 0.85, RMSE = 0.58 Nm/kg, rRMSE = 19.7%; fast running: R2 = 0.65, RMSE = 1.13 Nm/kg, rRMSE = 25.5%. Knee adduction moment: Moderate running: R2 = 0.4, RMSE = 0.37 Nm/kg, rRMSE = 34.4%; fast running: R2 = 0.21, RMSE = 0.8 Nm/kg, rRMSE = 33.8% | Kinematics and kinetics were collected by using a Vicon motion capture system and two AMTI plates simultaneously. Knee flexion and adduction moments were calculated via an inverse dynamic modeling |
| Stetter et al. (2019) | Tri-axial linear acceleration and tri-axial angular velocity | Filtered by a 4th order Butterworth filter with cut-off frequency of 15 Hz; IMUs signals were interpolated to keep the same sample frequency with knee joints forces data | Training/validation/ test: 0.70/0.15/0.15 |
LOSOCV | ANN (two hidden layers, 250 and 100 neurons) | Vertical, anterior-posterior, and medial-lateral knee joint forces | R2, rRMSE | Moderate running: mean R2 = 0.76, mean rRMSE = 25%; fast running: mean R2 = 0.73, mean rRMSE = 28.7% | Kinematics and kinetics were collected by using a Vicon motion capture system and two AMTI plates simultaneously. Knee joint forces were calculated via an inverse dynamic modeling |
| Hernandez et al. (2021) | Tri-axial linear acceleration and tri-axial angular velocity | Data were standardized using Z-score normalization | Training/ validation/test: 19/4/4 subjects |
Nested k-fold CV (user-independent approach) | DeepConvLSTM (two convolutional layers, two recurrent layers, sliding window: 100, step size:100) | Lumbar extension, bending, and rotation; hip flexion, adduction, and rotation (left and right); knee flexion (left and right); ankle dorsiflexion and inversion (left and right) | R2, ME, MAE | Mean R2 = 0.9 ± 0.16, mean MAE = 3.6 ± 2.1°, mean ME = 0.02 ± 3.75° | Marker-based Vicon motion capture system was utilized and inverse kinematics was conducted in OpenSim |
| Gholami et al. (2020) | Tri-axial linear acceleration | Filtered by a 4th order Butterworth low-pass filter with cut-off frequency of 6 Hz | Training/test: 0.80/0.20 | LOSOCV | CNN (kernel size = 3, stride = 1) | Hip, knee, and ankle angles | RMSE, NRMSE, R2 | Intra-participant model: R2 > 0.97, RMSE <3.4°, NRMSE <4.6%; inter-participant model: R2 > 0.78, RMSE <6.5°, NRMSE <11.1% | Marker-based Vicon motion capture system was utilized for collecting markers' trajectory and joint angles were calculated in Visual 3D (C-Motion inc.) |
| Wouda et al. (2018) | Relative orientation of the lower legs was input information in the first ANN; estimated joint angles and vertical accelerations were input in the ANN | Inertial data was down sampled to match the optical and vertical GRF data | Data of 10 and 14 km/h was used for training, running data at 12 km/h was used for test. | LOSOCV | ANN (two hidden layers, 250 and 100 neurons) | Vertical GRF and sagittal knee joint angles | R2, RMSE | Knee flexion/extension angles: RMSE <5°; vertical ground reaction force: RMSE <0.27 BW | Joint angles were collected with both Xsens MVN Link inertial and Vicon optical motion capture system; vertical ground reaction force was measured from an instrumented treadmill |
| Derie et al. (2020) | Auto-generated statistical features of 3D acceleration waveform; trial-specific features; subject-describing features | Filtered by a 2nd order band-pass Butterworth filter with cut-off frequencies of 0.8 and 45 Hz | / | LOSOCV; LOTOCV | EN, LASSO, XGB | VILR | MAE, R2, ROC | Subject-dependent XGB model: MAE = 5.39 ± 2.04 BW/s, R2 = 0.95; Subject-independent XGB model: MAE = 12.41 ± 7.90 BW/s; R2 = 0.77 | GRF were measured by two built-in force platforms (2 and 1.2 m, AMTI) |
| Liu et al. (2020) | Tri-axial accelerometer and gyroscope data | The number of data points per sample and mean, standard deviation, median, maximum, and minimum of the acceleration and angular velocity data were extracted from each step and anthropometric features for RunNet-MLP | Training/test: 0.80/0.20 | LOSOCV | Biomechanical parameter: RunNet-CNN (6 layers), RunNet-MLP (3 layers), and GBDT; running performance level: RunNet-MLP | Runners' performance level (novice, recreational and competitive), VALR, peak braking force and propulsion force, stride length, and running speed | Accuracy, confusion matrix, R2 | Runners' performance level: an overall accuracy of 97.1%; biomechanical parameters: RunNet-CNN: R2 > 0.9 | Biomechanical parameters were measured from an instrumental treadmill |
| Rapp et al. (2021) | Tri-axial accelerometer and gyroscope data | Synthetic accelerometry and gyroscope data were generated by taking numerical derivatives and adding Gaussian noise | Training/validation/test: 0.80/0.10/0.10 | / | Conv1D, LSTM | Flexion/extension, abduction/adduction, internal/external rotation of hip, knee, and ankle | RMSE | Mean RMSE of flexion/extension <1.27 ± 0.38°, Mean RMSE of abduction/adduction <2.52 ± 0.98°, Mean RMSE of internal/external rotation <3.34 ± 1.02° | Marker-based Vicon motion capture system used for collecting markers' trajectory and joint angles were calculated with custom software (Running Injury Clinic Inc.) |
| Ngoh et al. (2018) | Acceleration along x-axis | Acceleration was filtered using 2nd Butterworth low-pass filter with cut-off frequency of 10 Hz | Training/validation/test: 280 trials for training, 120 trials for validation and testing; Remain 230 data for accuracy evaluation | / | ANN (two hidden layers, 10 and 100 neurons) | Vertical GRF | R2, RMSE | RMSE <0.017 BW, R2 > 0.99 | Vertical GRF was measured from an instrumented treadmill |
| Young et al. (2020) | Tri-axial accelerometer and gyroscope data | Degree of pronation (neural, slight, and severe) and foot strike type (heel, midfoot, and forefoot) were measured or calculated from raw data | Training/test: 0.75/0.25 | / | Ensemble deep learning model (a MLP classifier, a GB classifier, and a custom-train ANN model) | Recommending running shoes type | Accuracy | Accuracy = 94.6% | / |
| Robberechts et al. (2021) | Filtered acceleration, Jerk, roll, pitch, acceleration right x peak min | Filtered by a 2nd order band-pass Butterworth filter with cut-off frequencies of 0.8 and 45 Hz | The perceptron model: training/test: 83/10 subjects; The RNN model: training/validation/test: 73/10/10 subjects | 5-fold CV, LOSOCV | The averaged structured perceptron algorithm; RNN (two bidirectional long short-term memory layers, 50 hidden neurons, dropout 20% after each layers) | Gait event detection (initial contact and toe off), stance time | MRE, MAE, ROC | The perceptron model: IC: MAE = 2.00 ± 2.89, TO: MAE = 9.00 ± 8.18, ST: MAE = 10.00 ± 8.73; The RNN model: IC: MAE = 2.00 ± 3.29, TO: MAE = 4.00 ± 4.52, ST: MAE = 6.50 ± 5.74 | Gait event were detected by two built-in force platforms (2 and 1.2 m, AMTI) |
| Zrenner et al. (2018) | Tri-axial accelerometer and gyroscope data | The IMUs data in each stride was zero padded to 200 samples | / | LOSOCV | CNN (two convolutional layers, two max pooling layers, one flattening layer, two fully-connected layers, and one 30% dropout layer) | Stride length and velocity; distance of running (3.2 km) | ME, MAE, MAPE | Running stride length: ME = 2.5 ± 20.1 cm, MAE = 15.3 cm, MAPE = 5.9%; velocity: ME = 0.055 ± 0.285 m/s, MAE = 0.216 m/s, MAPE = 5.9%; distance of running: MAE = 194.5 m | The Vicon motion capture system was used as the gold standard for velocity and stride length; total distance of field running was recorded using GPS by a smartphone (Galaxy S8, Samsung Inc.) |
| Komaris et al. (2019) | Tri-axial linear acceleration | Data were standardized using Z-score normalization | Training/validation/ test: 0.60/0.20/0.20 |
LOSOCV | ANN (one hidden layer with 10 neurons) | Vertical, anterior-posterior, and medial-lateral GRF | RMSE for force-time waveform evaluation; ME for peak force evaluation | RMSE: Vertical GRF: 0.134 ± 0.027 BW, anteroposterior GRF: 0.041 ± 0.007 BW, and mediolateral GRF: 0.042 ± 0.006 BW | GRF was measured using an instrumented dual-belt treadmill (Bertec Corp.) |
| Tan et al. (2019) | Tri-axial linear acceleration and composite accelerations over three timesteps | Data were scaled to a range of 0–5 using Min-Max scaling | Training/validation/test: 0.47/0.23/0.30 | / | LSTM (five layers, 44 hidden neurons in each layer) | Gait event detection (heel strike and toe off) | F1, Precision, Recall, and MAE | F1: heel strike: treadmill run = 0.92, indoor run = 0.96, outdoor run = 0.92; toe off: treadmill run = 0.77, indoor run = 0.86, outdoor run=0.81 | / |
| Watari et al. (2018a) | Tri-axial pelvic acceleration, patient reported outcome measures and demographic variables | Raw data were standardized to a mean of 0 and a standard deviation of 1, dimensionality reduction was performed with PCA | / | 10-fold CV | PCA (for feature extraction), SVM | Classifying patellofemoral pain cohort | Accuracy, precision, recall, F1-score, MCC, confusion matrix | Accuracy: 85.4%, precision: 90.0%, recall 96.4%, F1-score: 0.93, MCC: 0.69 | / |
| Watari et al. (2018b) | Tri-axial pelvic acceleration | Dimensionality reduction was performed with PCA, each step was normalized to 100 points and standardized to zero mean and unit variance | / | / | PCA (for feature extraction), HCA | Clustering patellofemoral pain patients into homogeneous subgroups | / | Two subgroups were identified for female runners | / |
| Ahamed et al. (2019) | Pelvic drop, vertical oscillation of the pelvis, ground contact time, braking, pelvic rotation, and cadence | / | Subject-specific approach | LOSOCV | RF | Classifying inclination conditions (downhill, level, and uphill) and determining the importance of each variable | Accuracy | Subject-specific approach: mean accuracy = 86.29%; LOSOCV approach: mean accuracy = 76.17% | / |
| Ahamed et al. (2018) | Pelvic drop, vertical oscillation of the pelvis, ground contact time, braking, pelvic rotation, and cadence | Biomechanical variables were averaged for each ten-strides | Training/test: 0.70/0.30 | One-against-another | RF (the number of trees: 100) | Classifying changes in subject-specific running gait patterns based on the environmental weather conditions and ranking the importance of biomechanical variables | Accuracy | Partitioning datasets: accuracy = 95.42%; One-against-another: accuracy = 87.18% | / |
| Clermont et al. (2019a) | Cadence, braking, vertical oscillation of pelvis, pelvic rotation, pelvic drop, and ground contact time | Biomechanical variables were averaged for each ten-strides | / | / | K-means clustering | Clustering running patterns throughout the marathon based on running gait alternations | / | Runners were clustered into two subgroups | / |
| Dixon et al. (2019) | Tri-axial linear acceleration | The first 2s of each trial were excluded, then the data were scaled from 0 to 1 according to the minimum and maximum value in the set of available trials for each subject; statistical, autocorrelation, sample entropy, smoothness, body load, and wavelet-derived energy features were extracted for the GB model | Training/test: 90%/10% | / | GB and CNN (two convolutional layers, one max pooling layer, two convolutional layers, one global average pooling layer and one drop out layer with probability of 0.5) | Classifying three different surfaces (concrete road, synthetic track, woodchip trail) | Accuracy, precision, recall, F1-score, confusion matrix | Accuracy: GB: concrete:93.7 ± 2.8, synthetic: 92.2 ± 2.1, woodchip: 95.7 ± 2.4, average: 93.9 ± 1.9; CNN: concrete:95.9 ± 4.0, synthetic: 94.7 ± 3.3, woodchip: 97.6 ± 1.2, average: 96.1 ± 2.6 | / |
| Johnson et al. (2021) | Tri-axial linear acceleration and time | 4D acceleration inputs were flattened into 2D images by representing the five sensors' locations on the horizontal axis, stance-normalized time frames upwards on the vertical axis | / | / | Two CNN models CaffeNet and ResNet-50 | GRF | R2, rRMSE | For moderate speed running of the left stance limb using CaffeNet, Vertical GRF: R2 = 0.97, rRMSE = 13.93%; For slow speed running of the left stance limb using ResNet-50, anterior-posterior GRF: r = 0.96, rRMSE = 17.06%; | Kinematics and kinetics were recorded and calculated with Vicon optical motion capture system and AMTI force plate |
| Tan et al. (2020) | Tri-axial accelerometer and gyroscope | Min-max normalization was used to normalize each IMU channel | / | LOSOCV | CNN (3 hidden layers with 50, 50, and 10 neurons, respectively) | VALR | R2, MAE, NRMSE | R2 = 0.94 ± 0.03, MAE = 13.8 ± 5.8 BW/s, NRMSE = 9.7 ± 3.6% | GRF data were collected using an instrumented dual-belt treadmill (Bertec Corp.) |
| Koska and Maiwald (2020) | Sagittal plane (gyroscope) data | Filtered by a 4th order low-pass Butterworth filter with cut-off frequencies of 20 Hz, data were normalized between 0 and 1 | 10, 20, 50, and 100% dataset were used for training model, respectively | / | KNN | Classifying the subjective perception of running shoe comfort (comfortable and uncomfortable) | CCR | Mean CCR = 0.92 | / |
| Matijevich et al. (2020) | Foot and shank minimum and maximum angles and angles at midstance | Feature were normalized to z-scores prior to model training | / | LOSOCV | LASSO | Peak force on the tibial bone | MAPE | Foot: MAPE = 7.9 ± 2.3%, shank: MAPE = 8.0 ± 2.9% | Kinetics was collected on a force-instrumented treadmill (Bertec Corp.) |
CV, cross-validation; LOSOCV, leave-one-subject-out cross-validation; LOTOCV, leave-one-trial-out cross-validation; RMSE, root-mean-squared error; rRMSE, relative root-mean-squared error; NRMSE, normalized root-mean-square error; R2, Pearson's correlation coefficient; ME, mean error; MAE, mean absolute error; MAPE, mean absolute percentage error; MRE, mean relative error; CCR, correct classification rate; ROC, receiver operating characteristic curves; MCC, Matthews correlation coefficient; ANN, artificial neural network; DeepConvLSTM, deep learning framework of the convolutional and LSTM recurrent neural networks; CNN, convolutional neural network; EN, linear regression with elastic net regularization; LASSO, linear regression with least absolute shrinkage and selection operator regularization; GB, gradient boosting; XGB, gradient boosted regression tree; GBDT, gradient boosting decision tree; MLP, multilayer perceptron; Conv1D, 1D convolutional neural network; LSTM, lone short-time memory; RNN, recurrent neural network; PCA, principle component analysis; SVM, support vector machine; HCA, hierarchical cluster analysis; RF, random forest; KNN, k-nearest neighbors algorithm; GRF, ground reaction force; VILR, maximal vertical instantaneous loading rate; VALR, average vertical loading rate; IMUs, inertial measurement units; BW, body weight.