Powered lower-limb assistive devices, such as powered prostheses and exoskeletons, have significant potentials for helping regain functional gait for people with mobility impairments from a variety of causes, including resulting from a spinal cord injury, stroke, or limb amputation.

Generally, these devices are expected to generate joint trajectories to support movements that are based on user intentions. Therefore, a proper understanding of joint kinematics is important for proper control of these devices. For example, prostheses control the powered knee and ankle to restore mobility by analyzing historical information from prior gait strides (Hargrove et al., 2015). Actuating timing of an ankle exoskeleton was controlled for the next stride based on the sensor signal from the previous stride; optimal timing reduced muscle activities and metabolic costs (Galle et al., 2017).

Walking can be represented as a gait cycle. Each gait cycle begins at a heel strike (0%) and ends at the next heel strike of the same leg (100%). The gait cycle can be broken down into a set of periods (e.g., stance and swing), with gait phase representing a specific timing in the gait cycle. Because gait cycle depends on joint kinematics and kinetics, gait phase prediction plays an important role in observing the human state and controlling powered assistive devices. Additionally, the device can be quantitatively evaluated for normal gait generation by analyzing the gait phase.

A variety of methods have been proposed to predict gait phase. Mechanical sensors (Kotiadis et al., 2010; Agostini et al., 2013; Maqbool et al., 2016; Han et al., 2019) (e.g., foot switches embedded in a shoe and inertial sensors on the shank) have been used to determine the gait phase directly. Using a finite-state machine (Lawson et al., 2014) is a popular method for most passive prostheses (Fluit et al., 2019); a state machine sub-divides the gait into a set of discrete phases and defines the transition conditions. A set of surface electromyography sensors on the lower limbs (Yao et al., 2021) were used to predict the discretized gait phase using a neural network having two hidden layers. Since discretization of a gait limits the information during locomotion, continuous gait phase prediction modes have alternatively been proposed. For example, a human-inspired phase variable (Quintero et al., 2017) (e.g., the thigh angle) that uniquely represents the gait cycle in a continuous and monotonically increasing manner was used to predict the gait phase. A discrete wavelet transform-based method (Livolsi et al., 2021) was proposed to predict the continuous gait phase for treadmill walking using hip encoders.

Recently, deep neural network-based (DNN-based) gait phase prediction methods have been proposed. Deep neural networks are highly capable of determining the non-linear relationship between the input and output mapping. The discretized gait phase was predicted using electromyography signals based on a temporal convolutional network (Chen et al., 2022). A recurrent neural network with a shank-mounted inertial measurement unit (IMU) was used to predict the gait phase for controlling an ankle exoskeleton (Seo et al., 2019). In addition, a long short-term memory (LSTM)-based network was proposed to predict the gait phase using a set of wearable sensors, including IMUs and a force sensor at the heel (Lee et al., 2021).

The end-to-end learning capability of DNNs enables gait phase prediction, with multiple ambulation modes showing different gait trajectories. A convolutional neural network-based gait phase estimator (Kang et al., 2021) was proposed to modulate hip exoskeleton assistance for multiple ambulation modes. Gait phases for three different ambulation modes were predicted using a single IMU placed at the shank (Weigand et al., 2020).

Although several studies have successfully demonstrated gait phase prediction, a remaining issue is the lack of efficient adaptation of gait phase prediction methods on a specific system as conventional methods are generally device-specific. In other words, conventional methods were validated using their own hardware. Therefore, a gait phase prediction model trained using a specific system cannot be easily adapted to other systems. Furthermore, training or validating a prediction model for a specific system is time-consuming and burdensome, particularly when targeting mobility-limited patients.

To address this issue, in this study, we propose prediction of continuous gait phase using a DNN. The primary objective was to eliminate the need for device-specific training of gait phase prediction models by using publicly available benchmark datasets (Camargo et al., 2021; Reznick et al., 2021) containing locomotion data of non-disabled individuals. The vertical load (i.e., the weight of users on force plates) and lower limb joint angles [i.e., thigh (induced by hip joint movement), knee, and ankle] were selected as the input data for the proposed gait phase prediction model because these data are commonly available in powered lower-limb assistive devices. Additionally, these data from non-disabled individuals can be collected using various sensor systems.

In conclusion, the proposed method (i) is versatile owing to using sensor data available in generic powered lower-limb assistive devices; (ii) facilitates developing a single unified gait phase prediction model trained on multiple datasets with diverse sensor characteristics; (iii) enables quantitative evaluation of a controller in terms of whether it can generate resembling gait phases of able-bodied individuals.

To validate feasibility, level-ground walking (LGW) and ascending stairs (SA) among various ambulation modes were considered for gait phase prediction because they have distinct differences in gait pattern (Kim et al., 2022).

For demonstration, the proposed method was applied to a prosthetic leg system. The gait phases of a state machine-controlled powered prosthetic leg for transfemoral amputees (Simon et al., 2014) were predicted and evaluated to verify that the proposed method can be used for gait evaluation and has the potential to be used in a system not included in the training datasets.

We hypothesized that the proposed gait phase prediction model could (i) predict the gait phase for a prosthetic leg using a model trained on benchmark dataset, and (ii) it can evaluate the generated gait using the device.

The proposed method uses a DNN to predict the continuous gait phase based on the history of the vertical load and joint angles (thigh, knee, and ankle). The gait phase prediction model was trained using benchmark datasets, and its prediction performance across benchmark datasets was evaluated. Then, we collected prosthetic leg data to demonstrate the trained model; four participants with transfemoral amputation then performed in-laboratory LGW and SA using state machine-based impedance controllers at a self-selected speed. Then, the gait phase was predicted from this recorded data. The details of this process are provided below.

The proposed method can learn the continuous gait phase from sensor data (i.e., the vertical load and lower limb joint angles) commonly available in powered lower-limb assistive devices. The model was trained using benchmark datasets and applied to a powered prosthetic leg with a different sensor configuration without tuning. Sensor data, such as vertical loads and joint angles, may depend on sensor characteristics. Therefore, a model trained on a specific sensor dataset may have poor prediction performance when testing the model on a different sensor dataset. For example, in the case of dataset A, the joint angles were simultaneously collected using a motion capture system and goniometers. However, the joint trajectories (Figure 5) and the performance of the models (Figures 3, 4) based on them differed. These results demonstrate that a gait phase prediction model trained using a specific system may not be easily adapted to other systems. These limitations suggest that a device-agnostic approach is needed. The end-to-end capabilities of a deep neural network enable the integration of data from different datasets and the development of a unified device-agnostic model. As demonstrated in this work, applying a model trained from benchmark datasets to powered lower-limb assistive devices without additional tuning will help to validate hardware or software easily.

Eliminating discontinuity is a crucial factor in the DNN-based gait phase prediction. We tried to use the 1-D variables in the Cartesian coordinate system rather than the 2-D variables in the polar coordinate system as the output of the network. However, it degraded prediction performance compared to the 2-D variables with most errors made around 0% or 100%. An example case is described in the Supplementary Document S1.

Prediction performance may depend on the configuration of the DNN. In this study, 250 ms of sensor data history were used to formulate predictions. This parameter was chosen based on our previous work (Kim and Hargrove, 2022), in which 250 ms of vertical load and thigh angle history was best to predict desired impedance parameters.

We scaled input and output data for the proposed DNN using Equations (2) and (4). This scaling is not mandatory and there may be better scaling methods. The reason for the scaling is that the OSL, which is the prosthetic leg we use, transmits and receives data in the scaled values. We wanted to minimize additional process for data conversion by following the data scaling of the OSL for our future work, real-time evaluation gait phase and updating controllers.

In this study, we heuristically determined the network architecture hyperparameters including the number the layers, number of units in the layers, and activation functions. We adapted the number of units in dense layers from our previous work for controlling prosthetic legs using DNN (Kim and Hargrove, 2022). In the case of the activation function, sigmoid was better than other activation functions, such as tanh and ReLU. We speculate that the sigmoid function is suitable for our small range near 0.5 of input and output to the DNN. In addition, additional layers may improve the performance; however, execution time for prediction will be increased as a trade-off.

The LGW and SA models showed poor inter-mode prediction performance due to differences in gait trajectories. However, the LGW models showed slightly better inter-mode prediction performance than the SA models. This difference in performance may be a result of the amount of training data, as LGW datasets have more data than SA datasets; the models based on SA datasets had fewer variations and may have overfitted. This speculation is also supported by the fact that dataset A showed better performance than dataset B; dataset A has nearly 20 times more data than dataset B (see Table 1).

In addition, one noticeable prediction result is that the models trained using AgRightLGW showed better performance on SA datasets than other LGW datasets. We speculated that the large variation in the joint angles from the goniometers improved the robustness of the model. Thus, model training using a variety of datasets and different sensor configurations may improve the overall performance of gait phase prediction for untrained modes.

In general, the state machine parameters for the stance phase should be tuned to generate a more natural gait similar to non-disabled individuals. Specifically, the ankle parameters for SA should be improved. In details, the ankle trajectory (blue line, Figure 6B) is too flat compared to the non-amputee trajectory (Figure 5B). A controller should provide more dynamic movement in the mid-to-late stance phase. In conclusion, the model provided information on how to improve the controller in addition to the similarity between gait data from the benchmark datasets.

The proposed method has several limitations. First, it was limited to level-ground walking and ascending stairs. Additional ambulation modes should be considered, including turning and descending stairs and ramps. We believe use of additional benchmark data can overcome this limitation. For example, dataset A contains data for turning on the ground. Training with these data will prove the robustness of the phase prediction model. Furthermore, to ensure that the method can be applied in practice, the model must be able to correctly identify the ambulation mode from the sensor data (i.e., ambulation mode classification). Second, although improving the performance of the controller for a powered prosthetic leg may be possible, the focus of this study was on evaluating the pre-developed controller. For instance, control parameters can be updated in real-time to reduce differences in gait phase and trajectory. However, updating the controller is beyond the scope of this study and will be performed in future work.

Real-time gait phase prediction may significantly improve the gait by incorporating various control methods. For example, walking at various speeds and slopes was facilitated using virtual constraints based on a human-inspired phase variable (Quintero et al., 2018) for controlling a prosthetic leg.

To investigate feasibility, we deployed the proposed model on a smartphone running Android 12 (Galaxy Z Flip 3, Samsung). The Android parsed sensor data every 5 ms from the OSL. Execution time for the deployed network was approximately 1.2 ms. In conclusion, model predicted the gait phase using a 250 ms history of sensor data at intervals of 5 ms, and this prediction took approximately 1.2 ms. This indicates that the model has the potential to predict in real-time. The Supplementary Video S2 contains real-time gait prediction based on the sensor data from the OSL, for which an intact limb user performed LGW on the OSL using a bypass adaptor. Performance was reasonable, but a significant error was observed when the user turned in place to change the body direction, as turning was not in the training datasets. This result shows the limitation of the proposed method: poor performance when an untrained mode was performed. In future work, we will continue to improve the robustness of untrained ambulation modes on various terrains. Then, the model will be used in the real-time update of control parameters to generate a normal gait similar to that in training datasets.

The proposed DNN model outperformed general machine learning-based algorithms (see the Supplementary Document S2). However, obviously, the performance of machine learning algorithms depends on model parameters as DNN models do; a state-of-the-art machine learning method may outperform the proposed DNN model. Despite this limitation, this work demonstrated the feasibility of device-agnostic gait phase prediction and evaluation could be possible by analyzing the history of the sensor data commonly available on powered lower-limb assistive devices.

In this study, we propose a gait phase prediction method using a DNN. The predictions are based on a set of sensor data commonly available on powered lower-limb assistive devices. The model was trained using benchmark datasets containing gait data of intact-limb individuals and applied to evaluate the performance of a powered prosthetic leg. The model successfully predicted the gait phase of individuals with transfemoral amputation using a powered lower-limb prosthesis and provided information on the differences between intact limb individuals from the benchmark datasets.

Hypothesis (i) was partially confirmed because performance may depend on the characteristics of the sensors and the amount of data used for training. In addition, the proposed model did not require tuning to evaluate the gait of transfemoral amputee patients wearing a powered prosthetic leg. Furthermore, it provided information on the differences between the actual gait and testing data, confirming Hypothesis (ii).

The original contributions presented in the study are included in the article/Supplementary material, further inquiries can be directed to the corresponding author/s.

The studies involving human participants were reviewed and approved by Northwestern University Institutional Review Board. The patients/participants provided their written informed consent to participate in this study.

MK designed the study, performed data collection, analyzed data, and wrote the first draft. LH supervised and administered the project and revised the manuscript. Both authors contributed to the article and approved the submitted version.