Automatic Assessment of Functional Movement Screening Exercises with Deep Learning Architectures

2025-05-02 0 0 856.04KB 19 页 10玖币

侵权投诉





Article

Automatic Assessment of Functional Movement Screening

Exercises with Deep Learning Architectures

Andreas Spilz 1,† and Michael Munz 2,†,∗

1Research Group Biomechatronics, University of Applied Sciences Ulm, 89081, Germany; Andreas.Spilz@thu.de

2Research Group Biomechatronics, University of Applied Sciences Ulm, 89081, Germany; Michael.Munz@thu.de

*Correspondence: Michael.Munz@thu.de

Abstract:

(1) Background: The success of physiotherapy depends on the regular and correct performance

of movement exercises. A system that automatically evaluates these could support the therapy. Previous

approaches in this area rarely rely on Deep Learning methods and do not yet fully use their potential. (2)

Methods: Using a measurement system consisting of 17 IMUs, a dataset of four Functional Movement

Screening (FMS) exercises is recorded. Exercise execution is evaluated by physiotherapists using the

FMS criteria. This dataset is used to train a neural network that assigns the correct FMS score to an

exercise repetition. We use an architecture consisting of CNN, LSTM and dense layers. Based on this

framework, we apply various methods to optimize the performance of the network. For the optimization,

we perform an extensive hyperparameter optimization. In addition, we are comparing different CNN

structures that have been speciﬁcally adapted for use with IMU data. To test the developed approach, it

is trained on the data from different FMS exercises and the performance is compared on unknown data

from known and unknown subjects. (3) Results: The evaluation shows that the presented approach is

able to classify unknown repetitions correctly. However, the trained network is yet unable to achieve

consistent performance on the data of previously unknown subjects. Additionally, it can be seen that

the performance of the network differs depending on the exercise it is trained for. (4) Conclusions: The

present work shows that the presented deep learning approach is capable of performing complex motion

analytic tasks based on IMU data. The observed performance degradation on the data of unknown

subjects is comparably to publications of other research groups that relied on classical machine learning

methods. However, the presented approach can rely on transfer learning methods, which allow to retrain

the classiﬁer by means of a few repetitions of an unknown subject. Transfer learning methods could also

be used to compensate for performance differences between exercises.

Keywords: Automatic Exercise Evaluation; Deep Learning; Functional Movement Screening

1. Introduction

For successful physiotherapeutic treatment, the regular and correct execution of movement

exercises at home is very important. A typical physiotherapy treatment begins with a detailed

diagnosis, followed by initial treatment steps such as manual therapy. The patient is then taught

various stretching and strengthening exercises. These exercises must be performed by the

patient at home between appointments to ensure that the therapy progresses. However, training

at home comes with certain problems. The movements performed here cannot be controlled

and corrected by trained personnel. Thus, the patient can permanently perform the prescribed

movements incorrectly, which on the one hand slows down the success of the treatment, and

on the other hand can even aggravate the existing injury or add additional ones. Of course, it

is not possible that every home training session is supervised by professionals. However, the

work of physiotherapists could be signiﬁcantly supported by a new technological development

that monitors the home training of patients. Funded by the project MyPhysio@Home we

want to address this research topic. Based on the data of a portable measurement system,

arXiv:2210.01209v2 [cs.LG] 18 Nov 2022

2 of 19

machine learning algorithms will evaluate a performed movement exercise. In the ﬁrst step,

these algorithms are to classify the movement quality on a scale. In a future step, the patient

should additionally receive feedback that informs him about the errors in his execution. In this

publication we would like to present the developments made for the ﬁrst step.

We want to explore this topic, using screening exercises from Functional Movement Screen-

ing (FMS) [

]. FMS is an assessment system consisting of seven exercises that can be used to

systematically determine movement restrictions or weaknesses in the human musculoskeletal

system. Each exercise is assigned a score of 3 (perfect execution), 2 (complete execution with

compensation movements), 1 (incomplete execution, even with compensation movements)

and 0 (pain occured). For each exercise there is a well-deﬁned list of movement characteristics

that must be fulﬁlled in order to receive a certain score. FMS is chosen because it is widely

used and one of the most popular screening systems in the ﬁeld of sports physiotherapy. In

addition, Cook’s system also convinced with high interrater and intrarater reliability values

[

]. This circumstance is especially interesting for machine learning applications that depend

on unambiguous label information. Hart et al. and Ordonez et al. have made suggestions for

a suitable dataset based on existing approaches to automated exercise evaluation [

]. The

dataset should contain recordings from an appreciably large number of subjects. It should con-

tain many different variations of exercise execution both incorrect and correct. These variants

should not be staged. To test the applicability of a methodology to multiple exercise types,

the dataset should contain different exercises. Based on these recommendations, we recorded,

processed, and labeled an FMS dataset with 20 subjects of 4 different exercises.

Studies in the ﬁeld of exercise evaluation rely on different measurement systems. Depth

cameras [

–

] or RGB cameras in combination with human pose estimation [

] are often

used. In order to track the entire body posture, several cameras are required at different

positions in the training room. The training room requires a comparatively large empty area

for this. In addition, these cameras must be calibrated to a common coordinate system, before

a measurement is taken, in order to provide usable results. This process is time-consuming and

requires considerable computer resources and experience. These demands on training room,

prior knowledge, and resources cannot be expected of a user. A suitable alternative to these

camera-based systems are inertial measurement units (IMUs). These units record acceleration,

angular velocity and magnetic ﬁeld strength on three axis. Attached to the body segments,

these measuring units can be used to record the kinematic movement of a human being. In

addition, they are comparatively inexpensive, do not require additional space and do not cause

additional work for the user during operation in an appropriately designed measurement

system.

Based on this IMU data we want to automate the evaluation of FMS exercises. Current

studies mainly rely on classical machine learning methods such as decision tree [

–

], random

forest [

] or support vector machines [

]. In contrast, very few studies e.g. [

] to date

use deep learning methods in combination with IMU data and exercise evaluation, as also noted

in systematic reviews on this topic [

]. This is particularly interesting, as these techniques

are already widely used in related topics. Human Activity Recognition (HAR), for example,

has been performed with IMU data and deep learning methods for quite some time [

]. Deep

Learning methods offer some advantages over classical Machine Learning methods, such as

automatic feature engineering and the option to apply Transfer Learning methods. Ordonez

et al. have shown that using a combined structure of CNNs and LSTMs it is possible to

distinguish between different activities using IMU data [

]. Lee et al. have already compared

the performance of a random forest approach with an approach based on a combination of a

CNN and a LSTM on an exercise evaluation task [

]. When classifying a squat into different

performance variants based on the data of ﬁve IMUs, the deep learning approach (accuracy:

91.7%) achieved signiﬁcantly better results than the classical approach (accuracy: 75.4%).

3 of 19

We will explore this approach further, adapt it to our measurement setup and the FMS

exercises and look in particular at the following issues. Hart et al. already noted in their review

of various exercise evaluation publications that hyperparameter optimization is often not

performed, or not documented, even though the beneﬁts are widely recognized [

]. Therefore,

to obtain the best possible results, we performed a detailed optimization to determine the exact

parameters for the best possible performance. We also present a network structure that allows

repetitions of different lengths to be used for training without distorting the temporal context

of the individual repetition.

We also want to test alternative CNN structures explicitly designed to process IMU data.

CNNs are originally designed for the analysis of information in image form. By arranging the

individual IMU channels as rows of a two-dimensional matrix, CNNs can also analyze this time-

varying information. Nevertheless, the two types of information differ signiﬁcantly. In the case

of an image, CNNs perform spatial convolution, whereas in the case of IMU data, convolution

occurs over time, the individual measured quantities (acceleration, angular velocity, etc.), and

the spatial relationship between IMU units. We want to investigate whether the performance of

a classiﬁer improves when this fact is taken into account by an adapted CNN structure. Recent

studies [

–

] already investigate the inﬂuence of different CNN structures on classiﬁcation

performance, we want to adapt this approaches to a new network structure and apply it in

exercise evaluation. In summary, the main contributions of this paper are as follows:

•

Introduction of a novel IMU-based dataset for the automatic evaluation of FMS exercises

•

Development of a neural network based approach for the automatic evaluation of FMS

exercises

•

Evaluation of the inﬂuence of different hyperparameters and network structures on

classiﬁcation performance.

• Performance evaluation of the developed system on different exercises from the FMS

2. Materials and Methods

2.1. Dataset

In the context of a study, we create a labeled dataset of four different FMS exercises. The

exact speciﬁcation of this dataset and our methodology is explained in the following.

2.1.1. Exercises

The described dataset contains four exercises from the FMS. Speciﬁcally, these are the Deep

Squat (DS), Hurdle Step (HS), Inline Lunge (IL), and Trunk Stability Pushup (TSP) exercises.

These are modiﬁed versions of the commonly known sports exercises squat, pushup and lunge.

The HS movement is rather unknown. Here, a subject stands upright in front of a hurdle, the

height of which depends on the person’s anatomy. Now the subject must lift one leg over this

hurdle and touch the ground on the other side with his heel. A sample image of each exercise

can be found in ﬁgure 1. Additional information on the exact movement sequence of each

exercise can be found in Cook et al. [25].

2.1.2. Study population

The exercises presented are performed by a total of 17 healthy volunteers. During re-

cruitment, care is taken to ensure that approximately equal numbers of men (9) and women

(11) participate in the study. The age of the volunteers ranges from 24 to 62 years, trying to

recruit as balanced as possible from the different age groups (41

13.53 years). The height

of the subjects ranges from 160 to 185 cm (173.5

8.3 cm) and the weight ranges from 40 to

110 kg (71.35

14.9 kg). We try to include athletic, average and rather inactive subjects in the

study population. Subjects self-assess their level of athleticism. Before the trial begins, each

participant signs an informed consent form. The study protocol was evaluated and accepted

4 of 19

Figure 1.

FMS exercises included in the recorded dataset. From left to right, from top to bottom: Deep

Squat (DS), Hurdle Step (HS), Inline Lunge (IL) and Trunk Stability Pushup (TSP). [25]

by the ethics committee of the University of Applied Sciences, Ulm. The study presented is

registered in the German Clinical Trials Register and can be found under the ID DRKS00027259.

2.1.3. Measurement setup

Each subject is instrumented with a total of 17 IMUs. We have divided the human body

into 17 segments, to each of which an IMU is attached. The number and position of the

individual IMUs is based on the Xsens MVN fullbody tracking system [

]. The setup deﬁnition

of Xsens seems suitable to us, because Xsens has years of experience in professional motion

analysis with IMUs and the products are already used in a wide variety of sports. In addition,

the products are already used by other research groups in many areas of motion analysis. The

exact positioning of the IMUs is shown in ﬁgure 2. The individual IMUs are attached to the

subject with elastic straps. The position on a segment is chosen in a manner that minimizes

displacement of the IMUs while in motion, e.g. by placing them between muscle bellies. In

the described setup we use Shimmer3 IMUs by ShimmerSensing. These IMUs feature a three-

axial accelerometer, gyrometer and magnetometer and an air pressure sensor. The individual

metrics are recorded with a sampling frequency of 120 Hz on a measuring range of

16 g

(accelerometer),

200

/s (gyrometer),

49.5 Ga (magnetometer) and 300-1100 hPa (air pressure

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

.ArticleAutomaticAssessmentofFunctionalMovementScreeningExerciseswithDeepLearningArchitecturesAndreasSpilz1,andMichaelMunz2,,1ResearchGroupBiomechatronics,UniversityofAppliedSciencesUlm,89081,Germany;Andreas.Spilz@thu.de2ResearchGroupBiomechatronics,UniversityofAppliedSciencesUlm,89081,Germany;Mi...

展开>> 收起<<

Automatic Assessment of Functional Movement Screening Exercises with Deep Learning Architectures.pdf

共19页,预览4页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Automatic Assessment of Functional Movement Screening Exercises with Deep Learning Architectures

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: