1 Introduction

The human knee is the largest and perhaps the most complex joint in the body. The knee plays a significant role during gait. It supports body weight and deceleration during the stance phase (Nordin and Frankel 2001). It acts like a spring which is evident from the linear relationship of moment–angle during stance phase. The positive impedance characteristics of the knee assist it during stance phase (Cherry et al. 2006). It has been observed that gait characteristics vary between able-bodied subjects and trans-femoral (TF) amputees due to stability of the prosthetic knee (Silver-Thorn and Glaister 2009; Mohanty et al. 2020a, b). Knee buckling during stance results due to instability and leads to gait deviations with increased energy consumption (Silver-Thorn and Glaister 2009; Romo 2000). Patient’s abilities and functional goals are considered to determine suitable prosthetic knee for smooth and reliable gait for each individual (Romo 2000; Michael 1999), and therefore, gait analysis has been used to provide a valid tool to correlate with experimental results (Mohanty et al. 2020a; b). Prosthetic knee is stabilized by inherent mechanical stability of the mechanism itself and the voluntary stability by hip extensors of TF amputee (Radcliffe 1977, 1994; Oberg 1983). However, a smooth transition from stance to swing phase is more difficult function to replicate (Radcliffe 1977; De Vries 1995; Andrysek et al. 2004).

Based on control, lower limb prostheses can be categorized into passive, semi-active, or variable damping and powered or intelligent (Geng et al. 2010). Passive prostheses require positive power at the knee and ankle joints and present with asymmetrical gait with increased energy consumption (Eilenberg et al. 2010). In addition, it is not possible for the individual with an amputation to generate positive power at the knee and ankle joints of a passive prosthesis due to the fact that there is no direct control over these prosthetic parts. Semi-active or auto-adaptive (Sawers and Hafner 2013) prostheses modulate damping levels to improve the knee stability and cadence responsiveness but cannot produce positive power (Martinez-Villalpando and Herr 2009). Active or intelligent prostheses are coupled to human gait through electronic systems capable of user’s intent recognition and can produce necessary positive power, and, therefore, restore locomotion and other activities of daily living more efficiently and naturally (Liu et al. 2014, Torrealba et al. 2010). Robotic technology is applied predominantly for troubleshooting of current active prosthetic controllers (Hargrove et al. 2013a; b). They recognize user’s motion intentions and proceed to aid in movements with least perceptual disturbances. Thus, an ingenious and communal prosthetic controller must begin with knowledge of the human controller (Tucker et al. 2015). Figure 1 shows a generalized working framework for lower limb prosthesis and Fig. 2 shows the general architecture for control systems with gait phase identification.

Fig. 1
figure 1

Generalized control framework for lower limb prosthesis (Tucker et al. 2015)

Full size image
Fig. 2
figure 2

General architecture for control systems with gait phase identification (Varol et al. 2008)

Full size image

In spite of significant developments in active and semi-active knee devices, potential issues like development of control strategies, portable power supplies, lightweight actuators, and high-efficiency transmissions, etc. need improvements (Romo 2000). An overview of control strategies in lower limb prosthesis has been provided by several review articles (Fluit et al. 2020; Berry 2006). Authors have focused on control strategies with focus on ankle prostheses (Jiménez-Fabián and Verlinden 2012), actuator design issues (Pieringer et al. 2017; Windrich et al. 2016), and corresponding control strategies (Lara-Barrios et al. 2018; Ferreira et al. 2015), yet a clear and in-depth overview is lacking. Previous studies have failed to separate the control methods from hardware and implementation details. The various control approaches of prosthetic knee devices to enhance the locomotion of subjects with trans-femoral amputations are thoroughly addressed in this review. In contrast to the existing literature reviews, a stronger attention is placed on the control methods.

Machine Learning (ML) has been reported in the literature as the study of how computer algorithms (i.e., machines) can “learn” complex relationships or patterns from empirical data (Wang and Summers 2012) and, hence, produce (mathematical) models linking an even large number of covariates to some target variable of interest (Obermeyer and Emanuel 2016). Application of a machine learning technique and method to prosthetic knee systems has been explored which works with the help of input of different sensors and apply sophisticated algorithms to give a better gait pattern to an amputee. Currently, machine learning algorithms are designed to recognize or to predict a locomotion mode to automatically adapt the behavior of prosthetic knees. In spite of this, a clear review of commonly used machine learning approaches and control algorithms for recognition and prediction of prosthetic knee functions is not found in the literature to authors’ knowledge.

Ankle and foot prostheses are not taken into consideration for this review. The main research question that needs to be addressed in this review was to explore and study of what approaches are being used in the literature to control prosthetic knee devices for directly assisting gait in subjects with trans-femoral amputation? This contribution provides a review of main control strategies and machine learning techniques proposed for prosthetic knees in restoring mobility of individuals with trans-femoral amputations. It includes an overview and comparison of essential technical details with special attention being paid to the algorithms employed for motion intent recognition, different walking adaptation, gait phase identification, and generation of walking patterns for better understanding through a clear presentation. The paper is arranged as follows. Most common control strategies applied in powered lower limb prosthetic devices are discussed in Sect. 3. In Sect. 4, a brief comparison of the presented literature review is presented. The conclusions are presented in Sect. 5.

2 Methods

This section is organized as literature search, study selection, data extraction, data synthesis, and analysis to show the procedure involved in systematic review.

  1. A.

    Literature search

    This systematic review is reported following the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) guidelines. This review focuses on Machine Learning Techniques and Control Strategies used in prosthetic knees. Relevant articles were collected by searching databases like PubMed, SCOPUS, EMBASE, and the Cochrane Controlled Trials Register (Rehabilitation and Related Therapies) up to April 2021.

    The keywords such as amputation, trans-femoral, prosthetic knee, machine learning, control algorithms, and control strategies were used for literature search. All results were checked for any duplication. Moreover, the reference lists of all searched studies were screened with to reveal any additional eligible studies.

  2. B.

    Study selection

    The following topics were chosen for review:

    1. 1.

      Control strategies of prosthetic knee devices.

    2. 2.

      Applications of machine learning techniques and control algorithms of prosthetic knee mechanism.

      Inclusion criteria set for selection of articles are as follows:

    3. a.

      Control strategies intended for active and semi-active knees.

    4. b.

      Algorithms for motion intent recognition, different walking adaptation, gait phase identification, and generation of walking patterns (to determine the required joint kinematics for performance of activities).

    5. c.

      Control algorithms for knee and hip devices.

    6. d.

      Experimental and numerical results for understanding the behavior of the algorithms (an approach to algorithm design and analysis).

      Papers were excluded for application of endo-prosthesis and in case of non-fulfillment of inclusion criteria. Any published systematic reviews and meta-analyses were excluded.

      Two investigators (R.K.M. and S.S.) independent from one another, screened the title first, followed by abstracts identified in the database searches. R.K.M. was responsible for reviewing abstracts identified in PubMed, EMBASE, and S.S. reviewed abstracts from SCOPUS and the Cochrane Controlled Trials Register databases (Rehabilitation and Related Therapies). These two investigators then applied the inclusion and exclusion criteria to the abstracts. Abstracts which did not meet the inclusion criteria were excluded and the reason for exclusion was recorded. Duplicate articles were removed. The remaining full-length articles were then retrieved and reviewed by these two reviewers to further determine whether the study met inclusion or exclusion criteria. A senior investigator (R.C.M.) verified and made the final decision.

  3. C.

    Data extraction

    Articles were assessed for their relevance to the implementation of control strategies and machine learning algorithms in prosthetic knees. The evaluation of papers was based on:

  4. a.

    description of hypotheses and objectives, including study designs,

  5. b.

    sufficient description of control strategies and machine learning approaches to extract essential technical details; and

  6. c.

    report of results with enough details to correlate with conclusions.

    The variables like control approach, control modes, gait control, hardware level, machine learning algorithm, and measured signals mechanism of prosthetic knee centered on restoration of mobility in amputees were drawn out from the included studies. This was done by several authors (R.K.M. and R.C.M.), independent from each other. All items were validated by all authors.

  7. D.

    Data synthesis and analysis

    Specific variables on control strategies and machine learning techniques were extracted and added to the summary table. These extracted features were used for the analysis. Summary table was used to group studies according to themes. During screening, grouping of articles was performed based on desired parameters. The results of this study were derived from a systematic review approach where all accessible articles were comprehensively analyzed and conclusions were made for best quantitative prediction.

3 Results

Flowchart showing results from the literature search is represented in Fig. 3. The search yielded abstracts from 278 published works. A total of 65 studies were included in this systematic review after applying inclusion and exclusion criteria.

  1. A.

    Control strategies

    The control strategies of prosthetic knee devices can be subdivided into three parts: high-, middle-, and low-level control. High-level control is accountable for recognizing the user’s intent of locomotion and supported signals from the user, environment, and the device. Mid-level controller receives these signals and translates the user’s intentions to a desired output state for the device. The directive from mid-level controller is delegated to the low-level, which represents the device-specific control loop that accomplishes the specified motion (Tucker et al. 2015). This is called the concept of the hierarchical behavior controller (Fukuda and Hasegawa 2004).

    The control approaches can be divided into four categories: echo control, finite-state impedance control, electromyography-based control, and Central Pattern Generator-based control.

    An echo control (Joshi et al. 2010; Grimes et al. 1977; Wu et al. 2011) synchronizes joint position of prosthesis based on the motions of the intact limb, but the implementation of response is not instantaneous.

    A finite-state impedance control (Eilenberg et al. 2010; Sup and Goldfarb 2008; Sup et al. 2009a, b; Liu et al. 2014; Lambrecht and Kazerooni 2009; Martinez-Villalpando and Herr 2009; Lawson et al. 2013; Sup et al. 2009) is the commonest and is based on generation of different joint torques around knee and ankle which are applied in each finite state of the locomotion. Calculation of impedance model of each joint was done using a virtual non-linear spring and damper.

    An electromyography-based control (Huang et al. 2009; Wu et al. 2011) uses EMG signals to compute the control inputs for the controller. An implicit agonist–antagonist linear muscle model is utilized to compute torque application at prosthetic knees. Thus, there is a proportional relationship between measured muscle signal and torque generation for control of prosthesis.

    A central pattern generator (Torrealba et al. 2012; Torrealba et al. 2010; Guo et al. 2010; Duysens and Forner-Cordero 2019) uses biologic neural grid. They are modulated by basic sensory signals which are being modeled and extensively studied. These controls can produce synchronized periodic patterns of activities. These systems have the advantages of smooth trajectory modulation, low estimated cost, and easy feedback integration.

  2. B.

    Machine learning techniques and control algorithms

    Machine learning uses various algorithms to realize gait harmony, movement analysis, and stumble control. Some commonly used techniques are: control logic, Intent detection algorithm, Genetic algorithm, mathematical logic-based classifier, Expectation maximization algorithm, and Impedance control algorithm.

Fig. 3
figure 3

PRISMA flowchart of the included studies

Full size image

Control algorithm (Awad and Abouhussein 2016; Tucker et al. 2015) is employed in which the controller supervises the joint position trajectory, however the output of those algorithms isn't always guaranteed.

Intent detection algorithm (Varol and Goldfarb 2007; Zhang and Huang 2013; Bhakta et al. 2020) is worn out two ways: unsupervised and supervised machine learning. Supervised learning uses the method of predicting a model on a trained range of inputs learning function to map the known output, which discovers the pattern of latest sets of information. The algorithm of unsupervised learning finds an answer to unknown or unlabelled data which does not require any reasonable supervision from humans.

Genetic algorithm (Martinez-Villalpando and Herr 2009; Zhang et al. 2019; Amador et al. 2012) may be a search-based optimization technique supported the principles of Genetics and survival of the fittest. It is frequently accustomed, find optimal or near-optimal solutions to difficult problems. The solutions are obtained from the big datasets which sensors have accumulated over a time span in response to situation-based demand from prosthesis.

Fuzzy logic (Alzaydi et al. 2011; Hong-liu et al. 2008) is an approach to variable processing that permits for multiple values to be processed through the identical variable. It strives to unravel problems with an open, non-specific range of knowledge to produce possible groups of precise conclusions. It works on the logic of grouping amputees, and hence, the input file is classed according to logic created within the algorithm.

Expectation maximization algorithm (Fessler 1994; Nandi 2008; Varol et al. 2008; Varol et al. 2009) is employed in an energetic knee where the body of data is somewhat congregated. It maximizes the probable data space function which is not measured instead of maximizing the function of the unfinished or measured data with improved efficiency and accuracy.

Impedance control algorithm (Varol and Goldfarb 2007; Herr and Wilkenfeld 2003; Duysens and Forner-Cordero 2019) is the most commonly used control strategy (El-Sayed et al. 2014) within which the torque generated is tailored to the produced knee angle. It ensures that the knee joint produces sufficient torque that is worthy for every phase of gait (Martinez-Villalpando and Herr 2009). Table 1 summarizes the control strategies and machine learning techniques used in prosthetic knee control.

Table 1 Summary of control strategies and machine learning techniques used in prosthetic knee control
Full size table

4 Discussion

The aim of this review paper was to overview the control strategies and machine learning techniques for prosthetic knee applications for restoring and improving gait performance. An attempt was made to enlighten on the basic concepts used in knee control including control approaches and modes, hardware platforms, machine learning algorithms, measured signals, and type of prosthesis.

Advancements in electronics and continuous research work have led to the development of cutting-edge to intelligent knee prostheses to improve the quality of life of amputees. The most common control strategies used by these kinds of prostheses are finite-state ones which are also noted as soft control. There are a group of pre-declared rules combined with information about specific criterion associated to natural and prosthetic gait available in a system directory. Control signals are sent to the prosthetic actuator for the necessary output action. Sensors placed at different tactical points on the body surface provide input signals which are processed through a processor. The processor compares these input values with the information in the directory, and attempts to recognize the current instant and activity of involvement during gait cycle. Each of these instants is correlated with a state, which is again linked with a pre-declared rule. This finally transforms into a control action through a prosthetic actuator. However, these controls are more complex, less robust, and lack real-time implementation of desired functions (Zlatnik et al. 2002). CPGs have been intended to spot users’ purpose during gait cycle and generate an algorithm that replicates bipedal locomotion (Gupta and Anand 2005). In particular, researchers have always been fascinated by electro-myographic signals. It is documented that the information transferred by these signals is useful to identify user intention and subsequently controls prosthesis. However, practical execution of this technology possesses some challenges like obtaining precise and desired signals, processing, and identifying the distinction between muscular activity and muscle fatigue (Park and Meek 1993). The kinematic and kinetic signals taken from several positions around the knee on amputated and sound limb of unilateral patients are used as soft signals for control of prosthesis. Continuity of works (Herr and Wilkenfeld 2003; Bar et al. 1983; Aeyels et al. 1992) in this direction has been able to replace a natural anatomic knee with a bionic one. This allows the amputee to achieve a suitable gait and human operation of it. Robotic systems are in operation with notions like CPGs, for producing joint tracking and harmonizing them to breed the various gait modes (Brambilla et al. 2006; Muthuswamy 2005; Billard and Ijspeert 2000). Currently, the cybernetic era of prostheses can apply bio-stimulated conceptions to get knee joint trajectory, walking modes, and overall performance in smooth and reliable way.

There have been various studies regarding control algorithm in prosthetic knee. There have been practical implementations of various machine learning techniques for automatic inception of control rules on human motions (Jonic et al. 1999) and exploration of inherent propulsion of human locomotion (Popovic et al. 2004). Most commonly used machine learning algorithms used in prosthetic knees include SVM and neural network-based control algorithm (ANN and CNN). There exists higher classification accuracy in SVM as it can be combined with other pattern classification methods to target different objectives (Labarrière et al. 2020). The ANN features significantly lower classification error than LDA (p < 0.05) (Woodward et al. 2016). CNN is often used to avoid manual feature selection and reported with accuracy above 89%. This has been reported that an adaptive algorithm performs significantly better than a non-adaptive algorithm and possesses an encouraging solution to achieve long-term locomotion mode classification (Liu et al. 2017). Kalanovic et al. (2000) proposes a feedback error learning (FEL) neural network approach for control structure of a powered prosthesis. This approach identifies the inverse dynamics of straight-forward single joint movements of an arbitrary trans-femoral prosthesis, which may be wont to track an arbitrary trajectory or a particular walking pattern. However, to achieve the convergence of neural network weights, learning rate must be adjusted, because it is extremely sensitive and no known method aside from trial and error which will guarantee the weights to converge to the simplest value. Another study (Herr and Wilkenfeld 2003) proposes a user-adaptive control for a variable-damping electronic knee. This approach utilizes data from sensors located on knee axis to adapt damping values of knee for matching the amputee's gait. The results show that the proposed open-loop controller performs better to match with biological gait compared to the mechanical passive knee. A study combining the four-bar link mechanism that usually employed in passive knees with MR damper is presented in Xie et al. (2010). The modeling and control proposed are a parametric approach, and thus contains many parameters to be defined. The results showed that the intelligent control proposed during this study is ready to follow the gait tracking of the healthy side of amputee leg in spite of particular delay. MR damper has been widely utilized in various applications. In robotics field, the study investigated by Garcia et al. (2011) proposed a mixture of MR damper and series elastic actuation for locomotion control for all-terrain robot. There are two control schemes utilized in this study, i.e., direct joint force control employing a PID control to come up with current command to the amplifier module and a cascade controller within the amplifier module. The results show that the proposed combined actuator and control can achieve a natural looking motion and can also reduce 20% of power in braking knee mechanism.

A variable stiffness control as proposed by Wentink et al. (2013) has been investigated in a modeling study of prosthetic knees to revive knee buckling during stance. Rotational stiffness is controlled to forestall excessive knee flexion, which is vital to supply a traditional gait. Torque generation in stance phase around the knee ensures normal gait trajectory and avoids risk of falling in amputees. EMG-based modeling approaches have been investigated that used joint kinematics and EMG data as input to the model (Lloyd and Besier 2003; Kwon et al. 2012; Schauer 2017). Swing phase control structure as proposed by Ekkachai and Nilkhamhang (2016) consists of a neural network predictive control with particle swarm optimization, and also a non-parametric feed forward neural network swing phase model. It utilizes knee angle data and voltage commands as the input to the controller. The performance of this controller was measured by normalized root mean squared error with validated data from an experiment. The results show that this controller performs better than the user-adaptive control found in Herr and Wilkenfeld (2003). Moreover, huge differences in outcome parameters of these reviewed studies investigating similar control strategies and machine learning techniques did not allow a meta-analysis to be performed.

5 Conclusion

This literature review has focused on control strategies and machine learning techniques applied to knee prostheses. Advancements in electronics, fluid mechanics, and mechanical knowledge provide the real-time functioning of active or powered prosthesis to attain a natural gait. The control approaches cannot be generalized, and may vary depending on variation of characteristics of disabilities, available prosthetic devices, organization of executed control schemes, and necessity of prosthesis to identify the gait episode or the user’s intent. Machine learning approach has been used to automatically optimize the high-dimension control parameters of the advanced knee prosthesis. Multiple algorithms have been proposed and used to control prosthetic knees for specific tasks and best suits the needs of amputees. The correct combination of control strategies and machine learning techniques will enable the improvement of prosthetic performance and enhance the quality of life of amputees. Future work on quantitative indicators of execution of the suggested algorithms, error measurements in joint trajectories, and functional outcome measure under realistic conditions are warranted with a comprehensive perspective to focus and resolve challenges of current prosthetic knees.