Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Keywords: The paper is introduced with a brief survey about existing causes for Knee ailments followed by conventional
AI treatments for them. Studies show that ailments like Osteo Arthritis (OA) of the Knee and Knee related injuries
Arthritis cause chronic pain and stiffness to the knee joint. This affects the range of motion of the leg. The severity of this is
Automated system
highly dependent on the age and BMI (Body-mass Index) of the patient. Further, a contrast between conventional
K -means Clustering
Physiotherapy Machines (CPM) and the proposed model is established. The paper proposes an alternative to
Machine learning
One-dimensional convolutional neural network the existing CPMs. A cost-effective system capable of diagnosing the severity of the knee using machine learning
Physiotherapy models and provide appropriate Automated physiotherapy. Using gyroscopic data and a predefined questionnaire,
a 1D-CNN is trained. An accuracy of 90.21% was obtained from the machine learning model. The accuracy of
the proposed model exceeded the accuracy of some state-of-the-art algorithms in determining the severity of the
affected knee by utilizing gyroscopic parameters and with the least computational cost.
© 2019 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license(http://creativecommons.org/licenses/by-nc-
nd/4.0/)
Peer-review under responsibility of the scientific committee of the 8th International Conference on Through-
Life Engineering Service – TESConf 2019.
Introduction The greater the flexion angle of a knee, the healthier the knee. An
experimental study infers that patient suffering from knee ailments can
Knee ailments can potentially range from genetic disorders to ac- extend or flex their legs, ranging from 0 to 60 degrees. Such motion
cidental injuries. Initial diagnosis of any problem of the knee is solely produces an optimum result without damaging the tissues near the joint
based on the extent of the mobility of the knee. Age, Sex, BMI are further and over-stressing the muscles with excessive therapy [3]. Continuous
taken into account to obtain accurate inference in the process of diag- manual physiotherapy over 4 weeks has a positive impact on reducing
nosis. Recovery of the knee joint post-diagnosis predominantly depends pain, increasing range of motion and functionality of the knee, this helps
on systematic physiotherapy. us understand that creating an automated physiotherapy routine that
Age, joint injury, obesity, genetics, and anatomical factors that affect emulates the medically accepted passive physiotherapy will largely help
joint mechanics and therefore considered vital factors in determining the people affected be independent and control their therapy [4].
the severity of Osteo Arthritis or stiffness in the joint. The most com- Post-surgery patients and aged patients with the affected knee are
monly affected joint in the body due to Osteoarthritis (OA) is the knee, generally prescribed to undergo Passive physiotherapy. CPMs are used
in particular, the strength due to weight of the body, gait, and flexibil- to rehabilitate the range of motion for post-operated Knees by provid-
ity of the joint was reduced by 43.8% in people suffering from OA [1]. ing passive physiotherapy. These devices consist of large mechanical
Extensive research clarifies that knee rigidity is associated with a stereo- changeable parts to adjust for different patients. These usually make
typical knee-stiffening gait pattern which reports instability in motion them very bulky. Therefore, CPMs are generally are used in hospitals.
and balance which is proportional to age in 68% of the cases [2]. Furthermore, the CPMs used in the hospitals are quite expensive
Physiotherapy is broadly classified into Active Physiotherapy and which adds to the financial burden for patients to obtain the therapy.
passive physiotherapy. Active Physiotherapy is when the patient can The proposed system is comparatively cheaper and removes the require-
provide therapy by putting effort on their own to move their muscles. ment for the patient to travel to the hospital due to the device’s porta-
Passive physiotherapy is carried out manually with the appointed Phys- bility. As shown in Fig. 1, the device consists of minimal moving parts,
iotherapist or by utilizing a Continuous Passive Motion machine (CPM). making it simpler to use on a day-to-day basis. Using gyroscopic sen-
∗
Corresponding author.
E-mail address: sridharkashyap04@gmail.com (S. Kashyap).
https://doi.org/10.1016/j.gltp.2021.08.052
Received 20 June 2021; Accepted 5 July 2021
Available online 13 August 2021
2666-285X/© 2021 The Authors. Publishing Services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd. This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
S. Kashyap, V. Venkatesh, M.K. Pushpa et al. Global Transitions Proceedings 2 (2021) 484–491
485
S. Kashyap, V. Venkatesh, M.K. Pushpa et al. Global Transitions Proceedings 2 (2021) 484–491
of physiotherapy after knee surgery is observed. The study was done on i The thigh rests on plank A
50 patients who had undergone Anterior Cruciate Ligament Reconstruc- ii The calf rest on plank B
tion surgery and were subjected to physiotherapy for 12 months. It was
observed that the group that underwent the therapy had better recovery This wooden structure is further attached to a wooden chair. The rear
of the knee and lead a better quality of life. plane of plank B is attached to a linear actuator to allow flexion and ex-
[11] observes the rehabilitation of the knee after a meniscus repair. tension of the leg. (Fig. 1a). The Linear actuator has a full load capacity
Also, it was observed that patients that underwent physiotherapy after of 200N enabling it to withstand any sudden resistance by the user. It is
surgery had faster recovery of range of motion than the ones that didn’t. controlled by an ESP32 microcontroller through an IBT2 motor driver.
[12] conducted an extensive research experiment on how physio- An MPU-6050 gyroscope sensor is used to collect real-time gyroscopic
therapy would help relieve the pain caused by OA. The patient was data. The sensor is placed on the edge of Plank B, very close to the knee
a 52-year-old male with pains in both knee joints while walking and to collect accurate real-time Euler-angles made by the movement of the
standing. After undergoing physiotherapy for a certain period, it was leg (Fig. 1b).
observed that the pain felt by the patient had significantly reduced. The Bluetooth module in the ESP32 is utilized to control this design
by a PC through a Graphical User Interface (GUI). The GUI is used to
Continuous Passive Motion Machines execute two phases of the system:
Methodology
Experimental Setup
486
S. Kashyap, V. Venkatesh, M.K. Pushpa et al. Global Transitions Proceedings 2 (2021) 484–491
During this process, the MPU6050 gyroscope continuously reads the Sample Resizing: Each sample of the training data must be of a fixed
real-time variation in angles along the direction of the leg movement. length. To ensure that, sample resizing is done. Sample with the largest
Further, the read sensor data is stored in a buffer array, in the ESP32 array length is determined from the training data. Every other sample is
memory. This data is sent back to the PC once the 3 cycles are finished. resized to the largest sample size. This is done by repeating the sequence
Fig. 3 represents the 1-dimensional array of gyroscopic data during of each sample till the length of each sample is equal to the length of
the Calibration Phase. The X-axis denotes the count of the output sensor the largest sample. Fig. 5 depicts before and after sample resizing
value and Y-axis denotes the change in Angle. From a set position, as
the leg extends, the change in angle observed here is from 0° to -70°. The Proposed Algorithm
During Retraction, the angle returns to 0° from -70°. (The negative sign
here is due to the orientation of the sensor attached). Once gyroscopic data is obtained from the Calibration Phase the fol-
Data Pre-processing lowing steps are this algorithm to estimate appropriate physiotherapy
The received 1-D gyroscopic array from the microcontroller under- routine for the user
goes two pre-processing phases:
I Estimation of Knee quality Metric (KQM) Using 1-D CNN by feeding
i Sample Thresholding the gyroscopic Data into the network.
ii Sample Resizing II Determining appropriate cluster based on obtained KQM from step
1, Age, and BMI.
Sample Thresholding: Since the sensor collects data in real-time, the III Determining the corresponding Physiotherapy routine based on ob-
size of different samples collected varies. A large number of static sen- tained Cluster and execute automated Physiotherapy
sor values are read during the calibration phase. After completion of one
The Flow of the algorithm is shown in Fig. 6
Calibration cycle, the user might take time to initiate the next cycle. Be-
Training Data for the 1D CNN MODEL: Experimental data was col-
tween these two cycles, a time delay is observed. During this time delay,
lected by simulating different possible cases to train the 1-D CNN model.
the sensor starts reading a static angle value as there is no movement
A total of 84 samples of different permutations were created which con-
of the leg in this time delay. This results in storing static values in the
sisted of a 1-D gyroscopic array ranging from 0° to -70° and other fea-
buffer array.
tures specific to the user were collected through a set of questions that
To overcome this, a limit angle is set. Values that are smaller than
were asked to the user in the GUI. The following features were collected
this limit angle are only stored in the array and values above the limit
for the 84 different samples in the training data:
angles are discarded. This simple logic ensures no static readings are
stored in the array. Fig. 4 depicts before and after sample thresholding. i Age,
487
S. Kashyap, V. Venkatesh, M.K. Pushpa et al. Global Transitions Proceedings 2 (2021) 484–491
Table 1
Sample training data
Table 2
One-Dimensional CNN model Parameters
ii Sex,
iii Height in meters,
Where x is the output obtained when kernel w convolves with the
iv Weight in kilograms,
input s. l is the current layer and k is the current neuron of the input sig-
v If the user was recently injured in the knee?
nal. Here, s is the obtained gyroscopic data obtained after the calibration
vi If the user underwent surgery on the knee?
phase.
vii 1-D gyroscopic array collected from the MPU6050 sensor
i The output obtained after performing the convolution pass is fed into
From Height and weight, Body to mass index was calculated from the fully connected layers of the neural network where the forward
the following formula: propagation is governed by the eq. 3
BMI=Weight/Height2 (1)
The training data set were labeled as ‘p’ to denote a sample as af- ℎ𝑜𝑢𝑡 = ℎ𝑖−1 𝑜𝑢𝑡 ⋅ 𝑊𝑖 + 𝐵𝑖 (3)
fected knee, and ‘n’ to denote a sample as a normal knee as shown in Here i defines the current layer of the neural network and B is the
Table 1. Bias of the Network. W is the Weights governing between layer i and
STEP 1 : 1D – CONVOLUTIONAL NEURAL NETWORK i-1. h is the activated neurons of layer i.
A convolutional Neural network uses a trainable filter that performs
convolution on the input layer of the network. Features learned by i The output layer of the 1D CNN consists of a single neuron which is
a CNN are more robust compared to a Multi-layered perceptron and activated by the sigmoid activation function given in Eq. 4
chances of model overfitting are significantly reduced. The convolu- 1
𝑌 = (4)
tional neural network tends to be robust when a change in the input data 1 + 𝑒−𝑥
is observed, like padding, skewing, scaling, etc. This variant of a con-
Where Y is the output of the 1D-CNN and x is the input to the final
volutional neural network requires a smaller number of perimeters and
layer of the network.
outputs higher accuracy on 1-Dimensional time-variant signals. They
are well suited for applications in mobile phones or microcontrollers. i Mean Squared Error is calculated between the obtained output from
Due to convolution, the parameters within the 1-dimensional filters are the forward propagation and the True label of the sample. The Mean
tightly packed and therefore learn the pattern of the 1-D inputs [22-24]. squared error is calculated from Eq. 5 as shown
In the proposed system, the quality percentage of the knee, termed (𝑚 )
1 ∑( )
as the Knee Quality Metric (KQM) was quantified by utilizing a 1-D Con- 𝑀𝑆𝐸 = 𝑌̂𝑖 − 𝑦𝑖 (5)
volutional Neural network model. This model was trained on the afore- 𝑚 − 1 𝑖=0
mentioned training data using Keras This 1-D CNN model was trained ii Post obtaining the MSE, backward propagation is carried out to train
as a linear regression model by using Linear activations for the hidden the 1D – CNN using Keras.
layer and Sigmoid activation for the output layer. The input layer of the
1-D CNN model is the obtained 1-dimensional Gyroscopic array from The model outputs a value between 0 to 1 which essentially regards
the sensor. The output layer consists of a single node representing the to the KQM. A value closer to 0 can be inferred as a good knee and vice
KQM. Fig. 7 shows the 1-D CNN network. versa. The following tabular column (Table 2) specifies the parameters
These steps were followed to train the neural network Model: of the model.
STEP 2 : K-MEANS CLUSTERING ALGORITHM : K-Means is an un-
i Forward Propagation of 1-D CNN is given in Eq. 2. supervised learning algorithm. The algorithm is used to cluster or create
partitions in the data based on similarity in features of the samples. This
𝑁𝑙−1
∑ ( ) algorithm is capable of accepting input and clustering without a label
𝑥𝑘 𝑙 = 𝑏𝑘 𝑙 + 𝑐𝑜𝑛𝑣1𝐷 𝑤𝑖𝑘 𝑙−1 , 𝑠𝑖 𝑙−1 (2)
set [25]. K-Means Clustering method is performed as follows:
𝑖=1
488
S. Kashyap, V. Venkatesh, M.K. Pushpa et al. Global Transitions Proceedings 2 (2021) 484–491
Table 3
Inference from the obtained clusters
Cluster Lower Age limit Upper Age Limit Lower BMI limit Upper BMI limit Minimum Threshold Angle
KQM value
0 28 54 23.72 30.12 0.81 -26
1 66 83 21.22 25.29 0.06 No Therapy
2 47 88 24.8 27.6 0.18 -54
3 42 75 21.83 27.21 0.73 -30
4 25 46 23.11 28.04 0.05 No Therapy
5 70 85 22.27 29.10 0.88 -20
Table 4
Fig. 8. Clustered Data Points Using K-Means Algorithm. Results obtained from users with a healthy knee
489
S. Kashyap, V. Venkatesh, M.K. Pushpa et al. Global Transitions Proceedings 2 (2021) 484–491
Table 5 The need for X-Ray imaging or MRI imaging is not required compared
Results obtained from users with Affected Knee to other methods to determine Knee Quality.
Name Age BMI Predicted KQM Threshold angle
User 1 65 21.6 0.830914855 -24.14141707 Discussion
User 2 53 30.1 0.844910085 -23.49595626
User 3 34 27.2 0.837528229 -23.8364079 The conceived design for an automated knee physiotherapy machine
User 4 21 23 0.7021234 -30.0812862
was implemented successfully. The ergonomic design along with the ac-
User 5 32 26.5 0.76 -27.41201416
curate machine learning algorithm provides a much more elegant and
robust solution compared to conventional means of physiotherapy. The
cost for a session was significantly reduced and the portability of the ap-
paratus makes it easily accessible to all demographics. The proposed al-
Affected Knee: gorithm uses the capability of a 1D convolutional neural network solely
The trend of data in the Table 5 speculates that an affected knee has a based on the obtained pattern of angles by the movement of the leg.
comparatively high KQM index (>0.70) as the average range of motion The algorithm is capable of differentiating different patterns of the gy-
of the knee angle ranges from 20-30 degrees (negative sign is due to roscopic data to detect the severity of the affected knee pain. This al-
the direction of motion to plane). Figure 5a. substantiates the condition gorithm reduces the computational cost and obtains similar or higher
of the knee as the patient cannot flex it beyond 30 degrees due to the results based on flexion and extension of the leg compared to state-of-
severity. the-art Deep learning systems. Determination of the severity of the Knee
The proposed algorithm was selected such that the highest accuracy joint is also done using CNN which requires MRI imaging or X-Ray Imag-
is obtained on the simplest parameter. The system determines the pat- ing. Another approach utilized in determining the severity of the Knee
tern of the recorded gyroscopic data based on the movement of the leg using Machine learning is by considering the Gait parameters of users.
and using 1D CNN followed by K A brief comparison with the existing state-of-the-art technology to
Means clustering, a metric to determine the severity of the knee is ob- detect the severity of the knee is shown in Table 6 compared to the
tained. With very less layers of the neural network, the system is compu- proposed algorithm’s effectiveness.
tationally inexpensive. This algorithm competes with the existing Deep The major limitation of the proposed model is limited data. In the
Learning Technologies proposed by [19-21] with a simpler architecture. coming future, more data will be collected to further augment the model.
Table 6
Comparison with existing State of the Art algorithms
H. Chang et al. Siamese Convolutional Neural MRI Images 86% -Requires MRI imaging.
Network
Soon Bin Kwon et al. a. Inception Res-Net CNN a. Radiological Imaging a.86.6% - Computationally expensive
deep NN with SVM b. Gait Pattern b.88.6%
Proposed Model 1-D CNN Gait Pattern – gyroscopic data 90.21% -comparatively less training
data
490
S. Kashyap, V. Venkatesh, M.K. Pushpa et al. Global Transitions Proceedings 2 (2021) 484–491
Further clinical trials are required to estimate the effectiveness of [14] N. Wirries, M. Ezechieli, K. Stimpel, M. Skutek, Impact of continuous passive motion
the device as compared to its traditional. Soon, the GUI is planned to on rehabilitation following total knee arthroplasty, Physiother. Res. Int. 25 (2020)
e1869.
be replaced by a mobile application, and data communication can be [15] D. Wang, J. Li, C. Li, An adaptive haptic interaction architecture for knee rehabil-
established through Wi-Fi as the algorithm requires less computational itation robot, in: 2009 International Conference on Mechatronics and Automation,
power to process data. 2009, pp. 84–89, doi:10.1109/ICMA.2009.5246430.
[16] T.K. Yoo, S.K. Kim, S.B. Choi, D.Y. Kim, DW. Kim, Interpretation of movement during
stair ascent for predicting severity and prognosis of knee osteoarthritis in elderly
References women using support vector machine, Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.
(2013) 192–196 2013doi: 10.1109/EMBC.2013.6609470. PMID: 24109657.
[1] A.Sh.ane Anderson, Rich.ard F. Loeser, why is osteoarthritis an age-related disease? [17] B. Norman, V. Pedoia, A. Noworolski, T.M. Link, S. Majumdar, Applying Densely
Best Practice & Research Clinical Rheumatology 24 (1) (2010) 15–26 VolumeIs- Connected Convolutional Neural Networks for Staging Osteoarthritis Severity
suePagesISSN 1521-6942. from Plain Radiographs, J. Digit. Imaging 32 (3) (2019) 471–477 Jundoi:
[2] J.A. Gustafson, M.E. Robinson, G.K. Fitzgerald, S. Tashman, S. Farrokhi, Knee 10.1007/s10278-018-0098-3. PMID: 30306418; PMCID: PMC6499841.
motion variability in patients with knee osteoarthritis: The effect of self-re- [18] G.H. Chang, D.T. Felson, S. Qiu, A. Guermazi, T.D. Capellini, V.B. Kolachalama, As-
ported instability, Clin. Biomech. (Bristol,. Avon) 30 (5) (2015) Jun475-80doi: sessment of knee pain from MR imaging using a convolutional Siamese network, Eur.
10.1016/j.clinbiomech.2015.03.007. Epub 2015 Mar 10. PMID: 25796536; PMCID: Radiol. 30 (6) (2020) 3538–3548 Jundoi: 10.1007/s00330-020-06658-3. Epub 2020
PMC4428966. Feb 13. Erratum in: Eur Radiol. 2020 Dec;30(12):6968. PMID: 32055951; PMCID:
[3] J.R. Brinkmann, J. Perry, Rate and range of knee motion during ambulation PMC7786238.
in healthy and arthritic subjects, Phys. Ther. 65 (7) (1985) 1055–1060 Juldoi: [19] S.B. Kwon, H. Han, M.C. Lee, H.C. Kim, Y. Ku, D.H. Ro, Machine Learning-
10.1093/ptj/65.7.1055. PMID: 4011684. Based Automatic Classification of Knee Osteoarthritis Severity Using Gait Data
[4] N. Arshad, M.J. Ijaz Gondal, M.M. Qamar, A. Basharat, Manual therapy can be a and Radiographic Images, IEEE Access 8 (2020) 120597–120603, doi:10.1109/AC-
potential therapy in knee osteoarthritis, Saudi J Sports Med 18 (2018) 18–21. CESS.2020.3006335.
[5] U.M. Kujala, M. Kvist, K. Osterman, Knee injuries in athletes. Review of exertion [20] T.D. Ngo, T.T. Bui, T.M. Pham, H.T. Thai, G.L. Nguyen, T.N. Nguyen, Image deconvo-
injuries and retrospective study of outpatient sports clinic material, Sports Med. lution for optical small satellite with deep learning and real-time GPU acceleration,
(1986) Nov-Dec. J. Real-Time Image Processing (2021) 1–14.
[6] S.L. Michlovitz, B.A. Harris, M.P. Watkins, Therapy interventions for improving joint [21] S. Vadivel, S. Konda, K.R. Balmuri, A. Stateczny, B.D. Parameshachari, Dynamic
range of motion: A systematic review, J. Hand Ther. 17 (2) (2004) 118–131 Apr-Jun- Route Discovery Using Modified Grasshopper Optimization Algorithm in Wireless
doi: 10.1197/j.jht.2004.02.002. PMID: 15162100. Ad-Hoc Visible Light Communication Network, Electronics 10 (10) (2021) 1176.
[7] S.P. Messier, R.F. Loeser, J.L. Hoover, E.L. Semble, C.M. Wise, Osteoarthritis of the [22] Se.rkan Kiranyaz, Onu.r Avci, Osa.ma Abdeljaber, Turk.er Ince, Monce.f Gabbouj,
knee: effects on gait, strength, and flexibility, Arch. Phys. Med. Rehabil. 73 (1) Dani.el J. Inman, 1D convolutional neural networks and applications: A survey,
(1992) 29–36 JanErratum in: Arch Phys Med Rehabil 1992 Mar;73(3):252. PMID: Mech. Syst. Signal Processing 151 (2021) Volume107398, ISSN 0888-3270.
1729969. [23] V.T. Pham, T.N. Nguyen, B.H. Liu, T. Lin, Minimizing latency for multiple-type data
[8] Ju.n Iwamoto, et al., Effectiveness of Exercise for osteoarthritis of the knee: aggregation in wireless sensor networks, in: 2021 IEEE Wireless Communications
A review of the literature, World journal of orthopedics (2011) 37–42 vol. and Networking Conference (WCNC), IEEE, 2021, pp. 1–6.
2,5doi:10.5312/wjo. v2.i5.37. [24] M.K. Kumar, B.D. Parameshachari, S. Prabu, S. liberata Ullo, Comparative Analysis
[9] Wall A., Board T. (2014) The Biological Effect of Continuous Passive Motion on the to Identify Efficient Technique for Interfacing BCI System, IOP Conference Series:
Healing of Full-Thickness Defects in Articular Cartilage. An Experimental Investiga- Materials Science and Engineering, 925, IOP Publishing, 2020.
tion in the Rabbit. In: Banaszkiewicz P., Kader D. (eds) Classic Papers in Orthopedics. [25] P. Berkhin, A Survey of Clustering Data Mining Techniques, in: J. Kogan, C. Nicholas,
Springer, London. M. Teboulle (Eds.), Grouping Multidimensional Data, Springer, Berlin, Heidelberg,
[10] K. Przybylak, M. Sibiński, M. Domżalski, et al., Supervised physiotherapy 2006.
leads to a better return to physical activity after anterior cruciate ligament [26] T.N. Nguyen, S. Zeadally, A. Vuduthala, Cyber-physical cloud manufacturing sys-
reconstruction, J. Sports Med. Phys. Fitness 59 (9) (2019) 1551–1557 Sep, tems with digital-twins, IEEE Internet Comput. (2021).
doi:10.23736/s0022-4707.18.08692-9. [27] G.B. Rajendran, U.M. Kumarasamy, C. Zarro, P.B. Divakarachari, S.L. Ullo, Land-use
[11] M.E. Şimşek, M.İ.S. Kapıcıoğlu, Physiotherapy in Orthopedic Knee Injuries: Rehabil- and land-cover classification using a human group-based particle swarm optimiza-
itation Program Following Treatment of Meniscus Repair, in: M. Bozkurt, Aç.ar H.İ. tion algorithm with an LSTM Classifier on hybrid pre-processing remote-sensing im-
(Eds.), Clinical Anatomy of the Knee, Springer, Cham, 2021. ages, Remote Sensing 12 (24) (2020) 4135.
[12] Poul.ami Adwa.it Bhandakkar, et al., Impact of Physiotherapy Rehabilitation on Pa- [28] L. Tan, K. Yu, F. Ming, X. Cheng, G. Srivastava, Secure and Resilient Artificial Intelli-
tients with Bilateral Osteoarthritis Knee Pain–A Case Report, Journal of Evolution gence of Things: a HoneyNet Approach for Threat Detection and Situational Aware-
of Medical and Dental Sciences (2020) vol. 9, no. 322316+Gale OneFile: Health and ness, IEEE Consumer Electronics Magazine, 2021.
Medicine.
[13] La.cy White, CPM Machine Protocol for Knee Replacements, MSN Capstone Projects
(2020) Paper 63.
491