Development of a deep learning algorithm to estimate knee adduction moment during gait using a single inertial measurement unit

doi:10.21203/rs.3.rs-3779001/v1

Download PDF

Method Article

Development of a deep learning algorithm to estimate knee adduction moment during gait using a single inertial measurement unit

https://doi.org/10.21203/rs.3.rs-3779001/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background: The external knee adduction moment (KAM) during gait is known as a digital biomarker for knee osteoarthritis (KOA). Three-dimensional motion capture system has been used to measure KAM, but it has limitations to measure KAM in cost, measuring time and space.

Purpose: To developed a simple and wearable system to estimate KAM, which can be used in daily practice. For this purpose, a deep learning algorithm was developed to estimate KAM of KOA using data from a single inertial measurement unit on the knee.

Study Design: Cross-Sectional Study. Level 3

Methods: Fifty female and ten male subjects who diagnosed as having KOA were included to develop an AI system. Simultaneous measurements of 5m walk using 3D motion capture and IMU attached on the knee were performed, and a deep learning algorithm was developed base on gait phase detection algorithm and 1D Convolutional Neural Network (CNN) model. Reliability of the AI system was evaluated by 5m gait data of another 13 KOA and 7 asymptomatic subjects.

Results: The mean absolute percent error (MAPE) and the mean absolute error (MAE) of estimated KAM peak of training data set and test data set were 13.0% and 0.0428 Nm.s/(BW*Ht), and 17.3% and 0.0505Nm.s/(BW*Ht) respectively, and those of estimated KAM impulse were 20.4% and 0.0214 Nm.s/(BW*Ht), and 21.4% and 0.0212 Nm.s/(BW*Ht) respectively. Peason’s correlation coefficients between estimated KAM (peak, impulse) and true KAM (peak, impulse) were 0.870 and 0.975 in train data set, and those in the test data set were 0.591 and 0.690. The intraclass correlation coefficients (ICCs) of estimated KAM impulse were 0.83, 0.94, 0.70 and 0.82, for ICCs (1.1), (1,k), (2,1) and (2,k) respectively.

Conclusions:

Using gait phase detection algorithm and 1D-CNN, we have successfully obtained the acceleration parameters (iKAM) from a single IMU data, and the acceleration parameters showed reasonable accuracy and reliability in estimating KAM. We require only single sensor on each knee which needs less calculation and hardware cost. Simple equipment of our system enables us to measure KAM during daily practice at clinic or hospital if there is free space to walk 5m.

Knee osteoarthritis

Gait analysis

Deep Learning

Inertial Measurement Unit

Knee osteoarthritis (KOA) is one of the major musculoskeletal disorders and is becoming a cause for increasing medical burden [1.]. The external knee adduction moment (KAM), which corresponds mechanical loads in the medial compartment of the knee joint during gait and it is known as important mechanical and clinical digital marker for KOA [2-4.]. Three-dimensional motion capture has been used to measure KAM, but it has limitations to evaluate the KOA patients in cost, measuring time and space, and it is still difficult to measure KAM during daily practice in clinical environment.

Recently, various wearable devices are used for health monitoring in older adults, to shift traditional medical routines to remote medicine, and to reduce an increasing medical burden for older people [5]. Improvements in measurement equipment and analysis technology have made it possible to easily measure quantitative objective measurements of various parameters that characterize gait, which can be used for sports, personal identification in security, medical care, and other applications. [6-8] Several novel systems such as Microsoft Kinect or inertial measurement units (IMUs) and pressure sensors have been developed, and wearable sensors combined with other sensors have been used to estimate knee joint moments instead of using the motion capture [8-10]. These systems have potentials to be used during clinical practice, but still need validations to use them outside the gait laboratory. A wearable system with IMU sensors and artificial neural network showed an excellent performance to estimate KAM [11], and has been used for gait retraining of the KOA in clinical trial. [12] However, there is a concern to estimate KAM with advanced KOA, since the limited numbers of advance KOA (3%) were involved to develop their system.

Lateral thrust is abnormal movement of the knee to lateral direction seen at the early stance phase and have been evaluated either visually and quantitively. [13-16] Castello et al. measured the lateral thrust by inertial sensors attached to mid-thigh, mid-shank and distal-shank, and reported that thigh adduction velocity was associated with KAM. [17] In our past study, we have attached the IMU on the tibial tubercle to collect the accelerations correspond to the lateral thrust, and found the medial-lateral acceleration range occurring at early stance phase correlated with the peak KAM. [18] Thus, the synchronicity and correlation between lateral thrust accelerations at the knee and first peak KAM is a possible rationale to use an IMU on the knee to estimate KAM.

The purpose of the present study was to develop a simple and wearable system to estimate KAM of KOA during gait, which can be used in daily practice in clinical environment. We hypothesized that development of a deep learning algorithm enables us to estimate KAM using data from a single IMU attached on the knee.

Participants

This cross-sectional study was conducted in accordance with the Declaration of Helsinki. Each participant was given a written informed consent and the study protocol was approved by our ethical committee (#20190294).

To develop the AI system and to validate its reliability, we had two different gait studies and enrolled different subjects for each gait study. In the first gait study, simultaneous measurements of 3D motion capture and IMU were performed. In the second gait study, gait analysis using IMU and AI system was performed.

AI development (Gait study 1)

We recruited patients who visited our outpatient clinic from March 2017 to February 2022 and were diagnosed with KOA by an orthopedic surgeon who has more than 30 years clinical experience in treatment of KOA. Patients with any symptoms in either the hip or the ankle joint were excluded from the study. Patients with any disorders that affect gait activity such as rheumatoid arthritis and lumbar spinal stenosis were also excluded. A total of 46 KOA patients and 14 asymptomatic subjects (50 females and 10 males) were enrolled in the present study. Mean age was 62.9 (12.8) years old and mean Body Mass Index (BMI) was 22.3 (3.3) kg/m. In this study, the radiographic findings of knee joint were evaluated on the basis of the Kellgren-Lawrence (KL) classification [19]. The patients with KL grade 0 showed no symptoms, and KL grade 1 showed symptoms such as pain or stiffness in the knee joint and tenderness or crepitus at the medial joint line without obvious joint space narrowing or osteophyte on radiographs, and the patients with KL grade 2 or higher showed obvious radiological changes and were defined as medial knee OA [19,20]. A total of 115 knees were involved and 5 knees were excluded because of having total knee replacement previously, and 28,27,26,26 and 8 knees were allocated to the KL grade 0,1,2,3, and 4 respectively on plain radiographs. (Table 1)

Reliability of the AI system (Gait study 2)

A total of 40 knee joints; 24 knees of 12 OA patients (8 females 4 males) and 16 knees of 8 asymptomatic subjects (7 females 1 males) were enrolled to evaluate the intra- and inter-rater reliabilities of the AI system. Mean age was 61.3 (10.1) years old and mean Body Mass Index (BMI) was 23.2 (4.3) kg/m2. A total of 15,9,12,3 and 1 knees were allocated to the KL grade0,1,2,3 and 4, respectively.

Testing Procedure

AI development (Gait study 1)

To generate deep learning protocol, gait data from simultaneous measurements of 3D motion capture system and IMU were obtained. A 3D motion capture system consisting of eight cameras (200 frames/second; Oqus, Qualisys, Sweden) and two force plates (frequency 2000 Hz; AM6110, Bertec, Colum-bus, OH, USA) were used. The force plates were synchronized to the sampling rate of the camera (200 Hz). In addition, two IMU sensors were fixed to the bilateral tibial tubercle with a strapped band. The IMU sensors used in this study was a commercial IMU (TSND151, ATR-Promotions, Kyoto, Japan) at the same sampling rate (200 Hz). The data from the IMU (3-axis accelerations, 3-axis angular accelerations, and 3-axis gyroscopes) were recorded on a laptop computer using a commercial soft-ware (SensorController, ATR-Promotions, Kyoto, Japan). After several practice trials, subjects performed 6 to 10 trials of a 5m walk at a self-selected speed in the gait laboratory. (Figure 1)　 Marker movements were recorded with Qualisys Track Manager software (version 2.7). Visual 3D (C-motion, Rockville, MD, USA) was used for knee joint kinematic and kinetic calculations. The first peak value of KAM in the stance phase was defined as KAM peak, and KAM impulse, the timed integral of all the KAMs was calculated for stance duration.

Figure 1 Experimental set up for simultaneous measurement of 3D motion capture and IMUs

Figure 2 shows the overview of our algorithm. Our algorithm comprises two main components. The input consists of a six-dimensional sequence formed by acceleration and angular velocity, which is preprocessed through a band-pass filter allowing only components ranging from 1Hz to 16Hz to pass. Subsequently, gait phase detection is conducted as illustrated on the left side of Figure 2. The process of gait phase detection unfolds as follows: (1) detection of local maxima and minima; (2) pairing consecutive local minima and maxima (indicated by the red squares in Figure 2). This pattern marks the final stage of the swing phase in each gait cycle and is utilized to define the gait cycle, denoted by the red numbers in Figure 2. The gait cycles detected in the area subsequently defined as Regions of Interest (ROI), and the original input sequence is cropped to exclusively incorporate these identified periods. [10] The cropped sequence is then fed into a one-dimensional Convolutional Neural Network (CNN) model, which ultimately outputs the predicted values after passing through a Generalized Mean (GeM) pooling [21] layer. The KAM peak prediction model adopts a broad kernel size of 31, resulting in a total parameter count of 3986. The impulse prediction model adopts a wider kernel size of 45, culminating in a total parameter count of 17314.

Overall, Total of 3 accelerations and 3 angular velocities about 3 orthogonal axis from IMU sensers during the stance phase were matched with KAM data (peak and impulse) obtained by 3D motion capture simultaneously. Gait data of 121 randomly extracted trials were used as training data, and a deep learning model based on 5-fold cross validation to match KAM using those accelerations and angular velocities was created. Finally, acceleration parameter (iKAM) was calculated based on the deep learning algorithm. A peak iKAM (iKAM_peak) and impulse of iKAM (the timed integral of all the iKAM: iKAM_impulse) that correspond to the KAM peak and KAM impulse were calculated and evaluated.

Figure 2. The overview of our algorithm.

Acc: acceleration, ω: angular velocity, ROI: region of interest, Conv: convolution layer, ReLU: rectified linear unit, GeM: generalized mean.

The accuracy of the AI

The accuracy of the AI system was evaluated by comparing true KAM peak and KAM impulse, and iKAM_peak and iKAM_impulse, both on the train dataset (121 gait data from 35 subjects) and the test dataset (31 gait data from 9 subjects) that were randomly chosen from the whole gait data.

Reliability evaluation (Gait study 2)

The intra- and inter-observer reliabilities of the AI system was evaluated using gait data obtained from a 5m walk in the gait laboratory. IMU data was collected using a commercial IMU (AMWS020, ATR-Promotions, Kyoto, Japan), which installs identical accelerometer and gyroscope with TSND151, at sampling rate of 200Hz. A single IMU was attached on each leg at the tibial tubercle of the subjects using a custom-made strapped band with pocket (Child corporation, Tokyo, Japan). The IMU data were recorded on a tablet using a commercial software (SensorController for Android, ATR-Promotions, Kyoto, Japan), and iKAM_impulse were calculated using the AI system generated in Gait study 1. To calculate the intraclass correlation coefficients (ICCs), all subjects repeated a 5m walk three times and the reliability of iKAM_impulse was evaluated.

Statistical analysis

In this study, our system was implemented using Scikit-learn, a Python module that integrates machine learning algorithms. [22,23] To evaluate the accuracy of a linear regression model, the mean absolute percent error (MAPE) and the mean absolute error (MAE) were used.

Intra-assessor probabilities were tested for one-way assignment agreement　(ICC(1.1)(1.k)), and inter-assessor probabilities were tested for two-way assignment absolute agreement (ICC(2.1)(2.k)). Analyses were per-formed using IBM SPSS Statistics ver29.0.1.0.

Accuracy of AI system (Gait study 1)

Average gait speed of all gait trails were 1.03 ± 0.18 m/s.

MAPE, MAE of iKAM_peak of training data set and test data set were 13.0% and 0.0428 Nm.s/(BW*Ht), and 17.3% and 0.0505Nm.s/(BW*Ht) respectively, and those of iKAM_impulse were 20.4% and 0.0214 Nm.s/(BW*Ht), and 21.4% and 0.0212 Nm.s/(BW*Ht) respectively. Peason’s correlation coefficients between iKAM (peak, impulse) and KAM (peak, impulse) were 0.870 and 0.975 in train data set, and those in the test data set were 0.591 and 0.690. (Figure 3）

Figure 3

A: True KAM peak and predicted KAM peak for training data.

B: True KAM peak and predicted KAM peak for test data.　

C: True KAM impulse and predicted KAM impulse for training data.

D: True KAM impulse and predicted KAM impulse for test data.

Reliability of the AI system (Gait Study 2)

ICCs(1.1), (1.k), (2.1), and(2.k) of iKAM_impulse were summarized in Table 2.

We have developed deep learning algorithm using gait detection algorithm and 1D-CNN, the accuracy increased after 200 epoch number. The hypothesis was confirmed that our algorithm enables us to predict peak KAM and KAM impulse during 5m walk using a single IMU. Past studies used multiple wearable sensors and pressure sensor to estimate KAM during gait. [9,10] To compare with those studies, our system has several advantages in estimation of KAM; we require only single sensor on each knee which needs less calculation and hardware cost. Simple equipment of our system enables us to measure KAM during daily practice at clinic or hospital if there is free space to walk 5m.

We have shown that our system has reasonable reliability of ICCs to measure iKAM. The reliability of knee gait mechanics of healthy participants was reported as all spatiotemporal, joint angle, and joint moment measures had ICCs of 0.83–0.94, except for the knee adduction moment peak in late stance (ICC = 0.69 and 0.72). In particular, ICCs of peak KAM and KAM impulse were 0.877 to 0.886. [24] This fact suggests that there are variations among healthy gait, resulting in more than 10% difference in measured KAM in same subject. Gait mechanics of KOA likely to have more variations because of existing pain and abnormal knee motion such as the lateral thrust. Thus, our results seemed to have reasonable reliability to evaluate KAM of KOA during gait.

Wang et al. used artificial neural network and XGBoost to train the data from IMU attached on the ankle, and estimated standard KAM measured by motion capture [4]. They reported successful results to estimate KAM from 6 axis IMU data, as the correlation coefficient between predicted and true KAM were up to R² of 0.956. While the excellent estimation of KAM using an unique technology and a single IMU on the ankle, several points should be addressed about their study. First, there was a lack in data of advanced KOA (only 3%, 5 of 156 knees were advanced KOA and majority 59%, 92 of 156 knees were KL grade 1 and there is a concern to estimate KAM of advanced KOA by their system. It is known that there were considerable differences in KAM between knees with KL1 and KL3. The differences in reported KAM impulse were more than 30% (average 1.10 and 1.40 Nm·s/(BW *HT)%) in KL1 and KL3, respectively) [25]. Second, there was a limitation in the data range in their KAM value. The true and predicted KAM ranged from 0 to 4 %Nm/(BW*Ht), and the average seemed to be around 2 %Nm/(BW*Ht). (Although there was only figure and no raw data were provided in their paper) The reported KAM value of knee OA patients ranged from 3.66 (1.20) %Nm/(BW*Ht) with KL grade 1 to 4.23 (1.12) %Nm/(BW*Ht) with KL grade 3 [25]. The KAM data range of our train data obtained from motion capture were from 1.59 to 6.78 (average 3.51) %Nm/(BW*Ht). Thus, the system by Wang et al. likely to have limitation to estimate the knee with KAM beyond 4%Nm/(BW*Ht), that should be majority of the KAM of OA knees.

Several limitations should be noted in the current investigation. First of all, data were collected from 60 participants, raising concerns regarding the external validity of the data. Nevertheless, multiple measurements were conducted for the majority of the subjects, ensuring as much diversity in the data as possible. Additionally, we have incorporated techniques such as Dropout and GeM pooling within the small neural network, taking utmost consideration to maximize the robustness of the output. Second, diversity in race, age, severity of illness, and BMI must also be considered when considering the population for IMU use.

We have successfully developed a deep learning algorithm using gait phase detection algorithm and 1D-CNN, and obtained acceleration parameters(iKAM) from a single IMU data that correspond to peak KAM and KAM impulse. The acceleration parameters showed reasonable accuracy and reliability in estimating KAM. Simple equipment of our system enables us to measure KAM during daily practice at clinic or hospital if there is free space to walk 5m.

KAM:knee adduction moment, KOA:knee osteoarthritis, CNN: Convolutional Neural Network, MAPE: mean absolute percent error, MAE: mean absolute error ICCs: intraclass correlation coefficients, IMUs:inertial measurement units, KL: Kellgren-Lawrence, BMI: Body Mass Index, ROI: Regions of Interest, GeM: Generalized Mean, Acc: acceleration, ω: angular velocity, Conv: convolution layer, ReLU: rectified linear unit.

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki, and approved by the Ethics Committee of Keio University School of Medicine (number: 20190294, approval date: March 24,2020). Informed consent was obtained from all subjects involved in the study.

Consent for publication

Not applicable.

Availability of data and materials

All supporting data can be provided based on request to the authors.

Competing interests

T.N. is the founder and shareholder of iMU, Co.,Ltd. T.N., K.H and Y.I. are patent holders for the processing device, program, method, and processing system.

Funding

This research received no external funding.

Author Contributions

A.A. analyzed the background and gait data of the subjects, and edited the manuscript. Y.I, K.H and T.N conceived and designed the experiments. H.Y. developed AI algorithm. K.N. collected the gait data. M.N. organized and designed the study. All authors approved the final version of the manuscript.

Acknowledgments

The author would like to acknowledge to Masaki Nagashima MD, Ryo Sasaki MD, and Rie Tanabe PT for helping the gait measurements and data collection.

Turkiewicz A, Gerhardsson de Verdier M, Engström G, Nilsson PM, Mellström C, Lohmander LS, Englund M. Prevalence of knee pain and knee OA in southern Sweden and the proportion that seeks medical care. Rheumatology (Oxford). 2015 May;54(5):827-35.
Amin S, Luepongsak N, McGibbon CA, LaValley MP, Krebs DE, Felson DT. Knee adduction moment and development of chronic knee pain in elders. Arthritis Rheum. 2004 Jun 15;51(3):371-6.
Hurwitz DE, Ryals AB, Case JP, Block JA, Andriacchi TP. The knee adduction moment during gait in subjects with knee osteoarthritis is more closely correlated with static alignment than radiographic disease severity, toe out angle and pain. J Orthop Res. 2002 Jan;20(1):101-7.
Miyazaki, Miyazaki T, Wada M, Kawahara H, Sato M, Baba H, Shimada S. Dynamic load at baseline can predict radiographic disease progression in medial compartment knee osteoarthritis. Ann Rheum Dis. 2002 Jul;61(7):617-22.
Chen C, Ding S, Wang J. Digital health for aging populations. Nat Med. 2023 Jul;29(7):1623-1630.
Muro-de-la-Herran A, Garcia-Zapirain B, Mendez-Zorrilla A. Gait analysis methods: an overview of wearable and non-wearable systems, highlighting clinical applications. Sensors (Basel). 2014 Feb 19;14(2):3362-94.
Stone E, Skubic M. Evaluation of an inexpensive depth camera for in-home gait assessment. J. Ambient Intell. Smart Environ. 2011 Jan 3(4):349–361
Pfister A, West AM, Bronner S, Noah JA. Comparative abilities of Microsoft Kinect and Vicon 3D motion capture for gait analysis. J Med Eng Technol. 2014 Jul;38(5):274-80.
Khurelbaatar T, Kim K, Lee S, Kim YH. Consistent accuracy in whole-body joint kinetics during gait using wearable inertial motion sensors and in-shoe pressure sensors. Gait Posture. 2015 Jun;42(1):65-9.
Liu T, Inoue Y, Shibata K., Shiojima K, Han M M. Triaxial joint moment estimation using a wearable three-dimensional gait analysis system. Measurement. 2014 Jan;47: 125-129.
Wang C, Chan PPK, Lam BMF, Wang S, Zhang JH, Chan ZYS, Chan RHM, Ho KKW, Cheung RTH. Real-Time Estimation of Knee Adduction Moment for Gait Retraining in Patients With Knee Osteoarthritis. IEEE Trans Neural Syst Rehabil Eng. 2020 Apr;28(4):888-894.
Wang S, Chan PPK, Lam BMF, Chan ZYS, Zhang JHW, Wang C, Lam WK, Ho KKW, Chan RHM, Cheung RTH. Sensor-Based Gait Retraining Lowers Knee Adduction Moment and Improves Symptoms in Patients with Knee Osteoarthritis: A Randomized Controlled Trial. Sensors (Basel). 2021 Aug 19;21(16):5596.
Chang A, Hayes K, Dunlop D, Hurwitz D, Song J, Cahue S, Genge R, Sharma L. Thrust during ambulation and the progression of knee osteoarthritis. Arthritis Rheum. 2004 Dec;50(12):3897-903.
Wink AE, Gross KD, Brown CA, Guermazi A, Roemer F, Niu J, Torner J, Lewis CE, Nevitt MC, Tolstykh I, Sharma L, Felson DT. Varus thrust during walking and the risk of incident and worsening medial tibiofemoral MRI lesions: the Multicenter Osteoarthritis Study. Osteoarthritis Cartilage. 2017 Jun;25(6):839-845.
Kuroyanagi Y, Nagura T, Kiriyama Y, Matsumoto H, Otani T, Toyama Y, Suda Y. A quantitative assessment of varus thrust in patients with medial knee osteoarthritis. Knee. 2012 Mar;19(2):130-4.
Mahmoudian A, van Dieen JH, Bruijn SM, Baert IA, Faber GS, Luyten FP, Verschueren SM. Varus thrust in women with early medial knee osteoarthritis and its relation with the external knee adduction moment. Clin Biomech (Bristol, Avon). 2016 Nov;39:109-114.
Costello KE, Eigenbrot S, Geronimo A, Guermazi A, Felson DT, Richards J, Kumar D. Quantifying varus thrust in knee osteoarthritis using wearable inertial sensors: A proof of concept. Clin Biomech (Bristol, Avon). 2020 Dec;80:105232.
Iwama Y, Harato K, Kobayashi S, Niki Y, Ogihara N, Matsumoto M, Nakamura M, Nagura T. Estimation of the External Knee Adduction Moment during Gait Using an Inertial Measurement Unit in Patients with Knee Osteoarthritis. Sensors (Basel). 2021 Feb 18;21(4):1418.
KELLGREN JH, LAWRENCE JS. Radiological assessment of osteo-arthrosis. Ann Rheum Dis. 1957 Dec;16(4):494-502.
Schiphof D, Boers M, Bierma-Zeinstra SM. Differences in descriptions of Kellgren and Lawrence grades of knee osteoarthritis. Ann Rheum Dis. 2008 Jul;67(7):1034-6.
Filip R, Giorgos T, Ondrej C. Fine-tuning CNN Image Retrieval with No Human Annotation. arXiv:1711.02512v2 [cs.CV] 10 Jul 2018
Gouwanda D, Gopalai AA. A robust real-time gait event detection using wireless gyroscope and its application on normal and altered gaits. Med Eng Phys. 2015 Feb;37(2):219-25.
Fabian P, Gaël V, Alexandre G, Vincent M, Bertrand T, Olivier G, Mathieu B, Peter P, Ron W, Vincent D, Jake V, Alexandre P, David C, Matthieu B, Matthieu P, Édouard D, Scikit-learn: Machine Learning in Python. J Machine Learn Res 2011, 12, 2825-2830.
Charlton JM, Birmingham TB, Leitch KM, Hunt MA. Knee-specific gait biomechanics are reliable when collected in multiple laboratories by independent raters. J Biomech. 2021 Jan 22;115:110182.
Hall M, Bennell KL, Wrigley TV, Metcalf BR, Campbell PK, Kasza J, Paterson KL, Hunter DJ, Hinman RS. The knee adduction moment and knee osteoarthritis symptoms: relationships according to radiographic disease severity. Osteoarthritis Cartilage. 2017 Jan;25(1):34-41.

Tables 1 and 2 are available in the Supplementary Files section.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Development of a deep learning algorithm to estimate knee adduction moment during gait using a single inertial measurement unit

Status:

Version 1

Abstract

Figures

Background

Materials and methods

Results

Discussion

Conclusions

Abbreviations

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1