The knee is one of the most complex joint structures in the human body. It is composed of the tibiofemoral and patellofemoral joints^1-4 and its formation involves both bone components (femur, tibia, and patella) and soft tissue components (synovial membrane, joint capsule, bursae, retinaculum, meniscus, and ligaments).

The movements of the knee occur in the tibiofemoral joint and are mainly flexion and extension, but there may also be internal and external rotation to a lesser extent.^2,6,7 The range of motion for knee flexion is 130° to 140°; however, these values may increase or decrease depending on the position of the hip joint during knee movement.^2,8

Currently, there are different diagnostic tests and specific exploratory maneuvers to assess the anatomic and functional characteristics of the knee joint complex. These tools are based on tests and clinical signs, and require the expertise of the physical therapist for a correct execution and interpretation of the results, and for a proper assessment of the integrity of the cartilage, muscles, menisci, ligament stability, etc.⁹

The functioning of the knee can be affected by pathologies of traumatic, degenerative, genetic, neurological, or autoimmune origin,¹⁰ the first two being the most common types. Depending on the type of injury, different intervention protocols should be implemented, using different techniques aimed at proprioceptive re-education to encourage the execution of reflex activities and activate and strengthen muscle groups to stabilize the joint and improve its muscle elasticity and joint thickness. These techniques are also useful for gait training and re-education of the sporting gesture.

Rehabilitation processes are based on protocols and clinical practice guidelines with therapeutic objectives that seek to potentiate joint motion through anisometric contractions that modify the length of the muscle.^11-13 Also, to determine the progress of the interventions, multiple devices are available to measure variables such as angular position, angular velocities, force and torque in different joints of the body.^14-17 However, these equipment are expensive and rehabilitation centers cannot afford them and must perform therapies in the traditional manner. Therefore, it is common that physical therapists do not have access to quantitative data that help them determine patients' progress during the different phases of rehabilitation.

Specifically, joint torque measurement is used to objectively determine patient progress as rehabilitation progresses^15-17 and is used in therapeutic interventions for anterior cruciate ligament injuries,^11,17 postoperative meniscectomy rehabilitations,¹⁶ lumbar injuries,¹⁵ among others. To this end, devices such as the Contrex¹⁸ and Human Norm¹⁹ systems are available on the market to monitor torque and allow the visualization of graphs that evidence the progress of this variable but, as mentioned above, they can be expensive and, in the Colombian case, they cost at least 10 times more than surface electromyography (sEMG) signal processing equipment.^20,21

Isokinetic dynamometers are instruments that allow obtaining information on torque during knee flexion-extension movement and, this way, establish its angle and maximum peak, as well as muscle power, muscle balance, etc.; these results allow quantifying objectively the recovery of the patient.^16,22-27 It should be noted that, despite its usefulness, the periodic collection of these data is limited due to high technology costs and, therefore, institutions prefer to use isometric dynamometers that have a lower cost but only allow measurements in static positions. This considerably limits the collection of relevant information for the implementation of rehabilitation processes.

On the other hand, sEMG signals are used as an alternative to estimate joint movements and the amount of force needed to perform a motor task,²⁸ as well as to determine the state of the musculoskeletal or neuromuscular system,^29-31 as they provide valuable information on the timing and relative intensity of muscle activity.^32,33 These signals are measured with surface electrodes placed on the skin above the muscle group of interest.^28,29,34 Currently, there are several low-cost sEMG sensors, which represents an advantage over other devices such as isokinetic dynamometers.

Given this scenario, the objective of the present work was to conduct a literature review on the use of sEMG signals as an alternative to calculate knee joint torque to objectively measure patients' progress during the different stages of rehabilitation of injuries in this joint.

A literature review was conducted based on the Cochrane Collaboration handbook.³⁵ The search was performed on the Engineering Village, IEEE Xplore, ScienceDirect, Web of Science, Scopus and PubMed databases using the following search strategy: years of publication: 1986 to 2018; type of publications: article and conference proceedings; language: English and Spanish; search equation: ("torque measurement" OR "torque estimation" OR "estimation of torque") AND (EMG OR sEMG OR electromyography OR electromyographic) AND Knee.

This review was based on the algorithms that have been developed to estimate knee joint torque through sEMG signals. It also considered how these algorithms can be used as an alternative to quantify the progress of patients during rehabilitation. To determine the search equation, MeSH terms that met the description required by the authors were established.

Publications in which sEMG signals were used to calculate knee joint torque were included. State-of-the-art reviews and references where torque was not measured using sEMG signals or which estimated torque in joints other than the knee were excluded. For information analysis, the current commercial value of isometric and isokinetic dynamometers in Colombia was considered.

355 records were retrieved, of which 47 were eliminated because they were duplicated. Exclusion and inclusion criteria were applied to the remaining 308 records, which led to eliminate 290 of them. Therefore, 18 publications were finally included (Figure 1).

Figure 1 Source: Own elaboration.

A total of 18 publications that describe, in a comparative way, how to estimate knee joint torque from sEMG signals were retrieved; the most relevant aspects are presented in Table 1. All the articles found were published in English and were original research works published in indexed journals and in memoirs of events.

Table 1

Author/Year	Torque estimation strategy	Surface electromyography signal processing	Muscles used	Type of contraction	Number of people studied
Hahn 35 2007	Neural networks	Full wave rectification and 5Hz low-pass filter	Vastus lateralis and biceps femoris	Isokinetic, eccentric, and concentric	20
Anwar et al31 2017	Neural networks, fuzzy logic	Quadratic mean	Rectus femoris and vastus medialis	Isokinetic	1
Anwar &AI- Jumaily38 2017	Support vector machine	Mean frequency, median frequency, total transformed spectral power, and wavelet	Rectus femoris, vastus medialis, vastus lateralis, biceps femoris, semitendinosus, semimembranosus	Isometric	5
Nurhanim et al.39 2017	Particle swarm optimization	Quadratic mean	Vastus lateralis	Isokinetic	1
Peng et al.40 2015	Neural networks	Full wave rectification and 2Hz low-pass filter	Rectus femoris, vastus lateralis, vastus medialis, biceps femoris, and semitendinosus	Eccentric and concentric	1
Menegaldo et al.41 2014	Hill's muscle model	Rectification and bandpass filter	Rectus femoris, vastus medialis, and vastus lateralis	Isometric	1
Tsutsui et al.42 2005	Neural networks	Rectification and moving average	Rectus femoris and biceps femoris	Isometric	1
Simon et al. 43 1995	Polynomial model	Rectification and low-pass filter	Rectus femoris, vastus lateralis, vastus medialis, semitendinosus, and biceps femoris	Isokinetic	5
Heine et al.44 2018	Hill's muscle model	Rectification and 6Hz low-pass filter	Vastus medialis, vastus lateralis, vastus medialis, and rectus femoris	Isokinetic	1
Ardestani et al.45 2014	Wavelet neural networks	Quadratic mean and 1Hz low-pass filter	Semimembranosus, biceps femoris, vastus intermedius, vastus lateralis, and rectus femoris	Gait	4
Anwar & Anam45 2016	Neural networks and machine learning	Mean frequency, median frequency, average power, total power, power spectral density, spectral momentum, and power spectral ratio	Rectus femoris, vastus medialis, vastus lateralis, biceps femoris, semitendinosus, semimembranosus	Isometric	5
Peng et al.47 2015	Musculoskeletal model and optimization with genetic algorithms	2Hz low-pass filter	Quadriceps, hamstrings, and gastrocnemius	Eccentric and concentric	1
Bai et al.48 2013	Continuous wavelet transform	Mean frequency	Quadriceps and hamstrings	Eccentric and concentric	10
Simon et al.49 1994	Pattern comparison	Rectification and low-pass filter	Rectus femoris, vastus lateralis, vastus medialis, biceps femoris, and semitendinosus	Isokinetic	5
Amarantini & Martin50 2004	Optimization	Full wave rectification	Rectus femoris, vastus medialis, biceps femoris, and gastrocnemius	Site walks	9
Anwar & Al- Dmour51 2017	Adaptive neural networks and fuzzy logic	-	Quadriceps	Isokinetic	1
Liu et al.52 2017	Hill's muscle model	Full wave rectification, low-pass filter	Rectus femoris, vastus lateralis and semitendinosus	Eccentric and concentric	1
Shabani & Mahjoob 53 2016	Hill's muscle model	Full wave rectification, 200Hz low pass filter	Rectus femoris, vastus medialis, vastus lateralis, semimembranosus, semitendinosus, and biceps femoris	Eccentric and concentric	1

Source: Own elaboration.

Table 2 classifies the records included according to the year of publication. It shows that most articles were published in 2017, with 27.78%.

Table 2

Year	1994	1995	2004	2005	2007	2013	2014	2015	2016	2017	2018
Articles	1	1	1	1	1	1	2	2	2	5	1

Source: Own elaboration.

The Journal of Biomechanics was the source from (2 references). The remaining journals and conferences which the largest number of publications was retrieved only contributed one article each (Table 3).

Table 3

Name of journal or conference	Number of articles	References
Journal of Biomechanics	2	36,50
Procedia Computer Science	1	37
2016 International Conference on Systems in Medicine and Biology (ICSMB)	1	38
2016 2nd IEEE International Symposium on Robotics and Manufacturing Automation (ROMA)	1	39
2015 International Joint Conference on Neural Networks (IJCNN)	1	40
Biomedical engineering online	1	41
Optomechatronic Sensors and Instrumentation	1	42
Proceedings of 17th International Conference of the Engineering in Medicine and Biology Society	1	43
Medical Engineering & Physics	1	44
Expert Systems with Applications	1	45
2016 6th IEEE International Conference on Biomedical Robotics and Biomechatronics (BioRob)	1	46
2015 IEEE International Conference on Robotics and Biomimetics (ROBIO)	1	47
2013 IEEE 13th International Conference on Rehabilitation Robotics (ICORR)	1	48
Proceedings of 16th Annual International Conference of the IEEE Engineering in Medicine and Biology Society	1	49
2017 IEEE Symposium Series on Computational Intelligence (SSCI)	1	51
2017 IEEE International Conference on Cyborg and Bionic Systems (CBS)	1	52
2016 4th International Conference on Robotics and Mechatronics (ICROM)	1	53

Source: Own elaboration.

Of the 18 papers included, 14 used algorithms to calculate knee joint torque during the execution of motor tasks involving movement^{36,38-41,43,45,47-53} and 4 used them for motor tasks without joint movement.^37,42,44,46

Each algorithm has different characteristics, such as the type of sEMG signal processing, the muscles used, the strategy implemented for the development of the algorithm, and the number of people studied. Some of the strategies used are neural networks, fuzzy logic, Hill's muscle model, support vector machines, particle swarm optimization, polynomial models, wavelet neural networks, and wavelet transform. Similarly, the algorithms differ in the types of contractions (concentric, eccentric, isokinetic, or isometric) used during sEMG signal detection.

Moreover, 13 of the articles reviewed used black-box techniques to estimate knee joint torque, while 5 did so using white-box models, specifically Hill's muscle models. Of the 13 investigations that opted for black-box models, 7 used neural networks; 3, regression and optimization-based models; 1, continuous wavelet transform; 1, vector support machines; and 1, pattern matching.

With this in mind, the authors of the present research describe below a work developed using neuronal networks, one using a regression model (black-box model), and another using a musculoskeletal model (white-box model).

First, Han³⁶ estimated the knee joint torque of 20 individuals using a three-layer feed-forward artificial neural network in two stages. In the first stage, they were asked to perform the maximum voluntary contraction; in the second stage, they were asked to perform exercises at 30% and 60% within the whole range of motion of the knee, exercising eccentric and concentric contraction. In both stages, the measurements of the sEMG signals and joint torque were recorded. It should be noted that in the neural network model, the second layer contained a variable number of "hidden" units ^{5,10,15,20,25,30} that represented the portion of the network learning process in which most of the processing solution occurred. Also, age, sex, height, body mass, the envelopes of the sEMG signals of the agonist (vastus lateralis) and antagonist (biceps femoris) muscles, which were obtained from full wave rectification using a 5 Hz low-pass filter, the joint angle and joint speed were considered as predictive variables of net torque. The study concluded that artificial neural network models achieved a more accurate torque estimate (R=96) compared to stepwise regression models (R=0.76), that the accuracy of the model increased considerably when the number of "hidden" units increased from 5 to 10, that accuracy improved progressively as more hidden units were added, and that, according to the results obtained, it is possible to say that the performance of the model could be the best if 15 or more "hidden" units were used, achieving 100% convergence and 88% to 90% accuracy.

On the other hand, Simon et al.⁴³ analyzed in 5 test subjects the relationship between the sEMG signals of the rectus femoris, vastus lateralis, vastus medialis, semitendinosus and biceps femoris muscles, as well as knee joint torque during flexion and extension. In this work, the authors designed a regression model as a function of angular position and velocity, the previous values of the torque and the rectified and smoothed sEMG signals. Based on this, they determined the coefficients using the least squares method from the information of 3 of the test subjects; the information of the 2 remaining subjects was used to validate the model. The authors obtained acceptable results in the validation subjects, where the R^A2 values were 0.98 and 0.96 for extension, and 0.92 and 0.73 for flexion.

Finally, Peng et al.⁴⁷ designed a model that consists of two main modules. Firstly, a muscle-tendon model calculates muscle force through the dynamics of muscle contraction; secondly, the values of these forces are entered into a musculoskeletal model to estimate joint torque. This model requires knowing details related to the muscles, such as length, force-length relationship, force-velocity relationship, among others; to validate it, the researchers used the mean squared error and the correlation coefficient, obtaining 3.65Nm in the first one and 0.96 points in the second one when they delayed the signal in 100ms. The results were considered logical due to the nature of the sEMG signal, which occurs 10-100ms before joint movement. It should be noted that this type of model allows us to know the individual contribution of each of the muscles studied, which can optimize patients' rehabilitation plans.

The algorithms of knee joint torque estimation found in this research have been developed by means of diverse techniques and their main objective is to estimate torque using the electrophysiological signals of the muscles of this joint. Unlike other algorithms that do not use muscle signals, they allow physical therapists to obtain additional and relevant information, such as muscle activation during rehabilitation processes.

The present literature review led to find an important number of publications that document the calculation of knee joint torque from sEMG signals during the execution of anisometric exercises. However, no publications were identified in which the torque calculated from sEMG signals was used in rehabilitation processes as such, so it is necessary to carry out research on this topic, which promises interesting applications in physical therapy.

Results regarding the measurement of knee joint torque from sEMG signals are an application of the biomedical signal processing theory and, therefore, an alternative route to traditional work in biomechanics and rehabilitation, which usually involves the application of mechanical laws.

In practice, measuring joint torque dynamically using sEMG signals represents an easily accessible and low-cost alternative to the use of isokinetic dynamometers for the patient's rehabilitation process. This alternative is also an option that expands the possibilities of monitoring and assessment.

Another advantage of measuring knee joint torque using sEMG signals is that they are always available for processing and, therefore, the physical therapist permanently has data on muscle activation, which are not provided by other joint torque measurement technologies. However, sEMG signals also have limitations because they require professionals to have basic knowledge of the capture technique to obtain good-quality results.