Skip to main content
NCBI home page
European Journal of Sport Science logoLink to European Journal of Sport Science
. 2026 Apr 8;26(5):e70167. doi: 10.1002/ejsc.70167

Making Sense of Shoulder Exercise: Measuring the Accuracy of an Artificial Intelligence Model to Classify Shoulder Exercise via Wearable Sensors Among People With and Without Rotator Cuff Tendinopathy

Josh Naunton 1,2,3,✉, Yanran Jiang 4, Rodrigo Bini 2, Dawson Kidgell 5, Kim Bennell 6, Terry Haines 7, Dana Kulić 4, Peter Malliaras 1
PMCID: PMC13060652  PMID: 41949979

ABSTRACT

This study aimed to compare the accuracy of machine learning classification for three commonly prescribed shoulder exercises in people with and without rotator cuff tendinopathy. Eighteen participants with rotator cuff tendinopathy (mean age 54.2, SD 13.2; 50% female), followed by eighteen matched controls completed a laboratory‐based shoulder strength testing protocol. Three exercises were performed (shoulder press, lateral raise and bent over row) while wearing three inertial measurement (IMU) sensors (Axivity, Ax6 ‐ 3 axis accelerometry and gyroscope at 100 Hz and 1000°/sec respectively) positioned on the wrist arm and trunk. Data were analysed and accuracy was compared between common machine learning algorithms for those with rotator cuff tendinopathy and healthy matched controls in a subject‐dependent and subject‐independent model. The best accuracy scores for the subject‐dependent results were achieved by a random forest algorithm; 96.12% (3‐sensor combined system) for those with rotator cuff tendinopathy. For the subject‐independent results best accuracy scores were achieved by a convolutional neural network algorithm; 94.55% (3‐sensors) for the healthy controls without shoulder pain. K‐fold cross validation confusion matrix results by exercise type for the entire cohort show 97% accuracy (shoulder press), 95.5% (lateral raise) and 90.7% (bent over row) (3‐sensors, CNN subject‐independent analysis). Machine learning classification of 3 different shoulder exercises in people with rotator cuff tendinopathy and matched healthy controls demonstrate most accurate results using a CNN algorithm for subject‐independent analysis and a RF algorithm for subject‐dependent analysis. Results were similar for both those with rotator cuff tendinopathy and their matched healthy controls.

Keywords: exercise, IMU, machine learning artificial intelligence, rotator cuff tendinopathy, wearables

Highlights

  • Shoulder exercise can be classified with good accuracy scores using an inertial measurement unit (IMU) accelerometer and gyroscope sensor placed on the wrist.

  • Subject‐dependent and subject‐independent models demonstrate comparable results depending on the chosen machine learning algorithm.

  • Exercise classification based on a wearable IMU shows similar results between those with rotator cuff tendinopathy and matched healthy controls without shoulder pain.

1. Background

Shoulder pain is a prevalent musculoskeletal condition affecting up to 27% of adults aged under 70 at any one point in time (Luime et al. 2004). The most common cause of shoulder pain is rotator cuff tendinopathy, accounting for up to 70% of all musculoskeletal shoulder pain (van der Windt et al. 1995). Clinical practice guidelines recommend progressive exercise over 6–12 weeks as the first line treatment for rotator cuff tendinopathy (British Orthopaedic Association 2014; Diercks et al. 2014; Hopman et al. 2013; Klintberg et al. 2015; Littlewood et al. 2019). Despite the recommendations for exercise treatment, recent trials have reported low exercise adherence rates, potentially impacting on the effectiveness of exercise treatment (Clausen et al. 2021; Hopewell et al. 2021; Liaghat et al. 2020; Littlewood et al. 2015). Recording and monitoring exercise adherence currently relies on direct supervision or video recording as a gold standard measure with most trials relying on self‐reported exercise diaries. Few trials use technology to measure what exercise has been performed (Clausen et al. 2021; Hopewell et al. 2021; Liaghat et al. 2020; Littlewood et al. 2015; D. Burns et al. 2021). Measuring exercise adherence via objective means will improve our understanding of the most effective or optimal exercise prescription variables for rotator cuff tendinopathy (Littlewood et al. 2015; Malliaras, Johnston, et al. 2020; Naunton et al. 2020; Page et al. 2016). Exercise prescription is regularly based on one repetition maximum (1RM) strength results, yet there is a lack of normative data for 1RM tests for common shoulder exercises among people with rotator cuff tendinopathy, as isometric hand held dynamometry is commonly utlised (Ingwersen et al. 2017; Larsson et al. 2019; Chezar et al. 2013; Berkovitch et al. 2013; Almeida et al. 2020; Clausen et al. 2017; Akyol et al. 2013).

Recent advances in wearable technology, such as smartwatches, accelerometers, inertial measurement units (IMUs), activity trackers or smartphones may potentially provide a convenient, efficient and acceptable method to objectively monitor exercise adherence. Use of such technology is increasing within healthcare with devices now able to track step count, energy expenditure, maximal oxygen consumption (VO2max), sleep, oxygen saturation and even reporting a basic ECG signal (Shei et al. 2022; Li et al. 2022). Various smartwatches (Samsung Galaxy Watch, Garmin, Amazfit, Coros) have recently implemented repetition counting for a limited set of exercises, although this is reliant on manually starting and stopping the exercise recording. This technology often relies on artificial intelligence machine learning algorithms. However, the best machine learning algorithm and its accuracy in classifying shoulder exercise is not known.

Several studies have investigated the use of wearable technology to classify IMU sensor signal by the exercise performed, tracking step count or physical activity (Um et al. 2017; D. M. Burns et al. 2018; Chang et al. 2007; Cheng et al. 2017; Ebert et al. 2017; Morris et al. 2014; Muehlbauer et al. 2011; O'Reilly et al. 2017; Seeger et al. 2011; Shen et al. 2018; Gorzelitz et al. 2020). Despite this, few studies have investigated IMU sensor signal for exercise classification among clinical populations with only two other lab groups doing so among people with rotator cuff tendinopathy (D. Burns et al. 2021; Bavan et al. 2019; Bavan et al. 2020).

This study aims to: (1) compare the concurrent validity and accuracy of various machine learning algorithms' ability to automatically detect and classify recorded IMU sensor signal by exercise (i.e., shoulder press, lateral raise or bent over row) among people with and without rotator cuff tendinopathy; (2) compare the accuracy of IMU sensor placement between the wrist, arm, trunk or a combined 3‐sensor system; and (3) to measure one repetition maximum (1RM) strength of a shoulder press, lateral raise and bent over row for people with rotator cuff tendinopathy compared to matched healthy controls.

2. Methods

2.1. Experimental Design

This prospective non‐randomised observational matched case‐control study was reported according to the STROBE guidelines (Strengthening the Reporting of Observational Studies in Epidemiology) (Cuschieri 2019). Ethical approval was granted from the University Ethics Committee (No. 27192). Data were gathered from the Biomechanics Laboratory at a partner University campus subsequent to obtaining site approval from their own University Human Research Ethics Committee. All participants gave written informed consent prior to completing the study.

2.2. Participants

Thirty‐six participants (18 with shoulder pain and 18 matched controls) were recruited from a regional centre (Victoria, Australia) to take part in the trial. We aimed to recruit a minimum of 30 participants (minimum requirement not to be considered inadequate as per the COSMIN guidelines) (Mokkink et al. 2019) to detect a correlation of ≥ 0.6 (α = 0.05, power = 0.95), noting that power analysis is not usually applied to machine learning approaches (Rajput et al. 2023; Balki et al. 2019). 18 participants with rotator cuff tendinopathy were recruited via social media and clinical partners (Figure 1). Information regarding the trial was posted in free community and personal social media pages and posters were placed in the waiting rooms of clinical partners. Eligibility and screening were completed via phone call, with video when required in accordance with previously utilised methods (Malliaras, Cridland, et al. 2020). If the diagnosis remained uncertain, participants underwent face to face clinical examination by the primary author (more than 10 years clinical experience) prior to inclusion. Eligibility and exclusion criteria are listed in Table 1. Healthy participants without shoulder pain were recruited via the same social media channels and matched for age (within 2 years) and sex to the shoulder pain group. Participants utilised their dominant hand to match the painful group, for example, if the participant with pain used their dominant hand then the matched participant also used their dominant hand.

FIGURE 1.

FIGURE 1

Open in a new tab

Recruitment flow diagram.

TABLE 1.

Inclusion and exclusion criteria for the study.

Inclusion criteria for participants with rotator cuff tendinopathy Exclusion criteria (both groups)

  • Primary presenting complaint of anterolateral shoulder pain (with or without referral into the upper arm)

  • Shoulder pain that is primarily associated with movement, particularly overhead activity or increased shoulder load resistance

  • Preservation of passive shoulder range of movement

  • Age 18 years or older

  • Cervical spine referred pain

  • Hemiplegic shoulder pain

  • Systemic pathology including inflammatory joint conditions and neoplastic disorders e.g. rheumatoid arthritis

  • Neurological disease affecting the shoulder (e.g. Parkinson's disease)

  • Clinical diagnosis consistent with adhesive capsulitis or glenohumeral osteoarthritis
Open in a new tab

2.3. IMU Sensor Measures

Participants attended the university campus where a clinical assessment of their shoulder pain and anthropometry measures were completed. The testing protocol was explained, after which participants were fitted with three IMU activity tracking sensors (Axivity Ax6). The IMU sensors were attached using an elastic strap to secure the sensor in place on the trunk (anterior lower sternum), upper arm and at the wrist on the side of their painful shoulder or their matched arm (for dominance) in the pain free group (Figure 2). The IMU sensors consist of a 3‐axis accelerometer and 3‐axis gyroscope with inbuilt memory. The accelerometer was set to record at 100 Hz with a 16g range and the gyroscope was set to 2000° per second. Sensor data were downloaded via OMGui software (Open Movement, Newcastle University, UK) at the completion of the session. The session was recorded via video camera in order to correctly label the sensor signal for accurate data analysis. Data were time synced using a unique movement and repetition pattern, before matching data points.

FIGURE 2.

FIGURE 2

Open in a new tab

IMU Sensor setup. Circles show IMU devices fixed in place at the wrist, arm and trunk.

2.4. 1RM Strength Measures

Participants completed a short standardised warm up (one slow set of each exercise unweighted and one set at maximum speed unweighted), followed by a one‐repetition maximum (1RM) strength test for a single arm overhead dumbbell press (shoulder press), single arm dumbbell lateral shoulder raise (lateral raise) and single arm bent over row (bent over row). Participants were provided with 1–3 min rest between each effort. Three to five progressively loaded submaximal efforts of 1–3 repetitions were completed to determine the 1RM maximum according to previously published protocols (Schroeder et al. 2007; Levinger et al. 2009; Faigenbaum et al. 2012; Baechle et al. 2008). Participants were then instructed in varying order to complete each exercise (shoulder press, lateral raise and bent over row) at 40%, 60% and 80% of their 1RM maximum result. Participants were instructed to complete each repetition as fast as they could until they were no longer able to lift with adequate form or they reported pain that prevented them from completing the exercise. Each participant then completed a standardised cool down of each exercise without additional weight. Minimal cues were provided regarding exercise form in order to mimic the most realistic conditions outside of the university laboratory.

2.5. Data Analysis

Ground truth exercise labels (3 classes, shoulder press, lateral raise or bent over row) were applied to the data manually by reviewing the video recording of each session. Raw inertial sensor data were downloaded via OMGui (Open Movement, Newcastle University, UK) and subsequently imported into a python environment for further analysis. Figure 3, shows an example of raw IMU sensor signal plotted to demonstrate each exercise.

FIGURE 3.

FIGURE 3

Open in a new tab

Example of raw IMU recorded signal for each exercise.

Exercise classification was implemented in Python via the Scikit‐learn machine learning package (Pedregosa et al. 2011). Data were first filtered via a lowpass butterworth filter (20 Hz cutoff, order 5) before segmenting the data series via a 2 s sliding window. We utilised a supervised machine learning model comparing four common machine learning algorithms. These algorithms were chosen due to their use in a similar trial (Um et al. 2017; D. M. Burns et al. 2018; Bavan et al. 2019) and to test simpler algorithms before those that are more intensive or advanced. Three algorithms were the same for both a subject‐independent and subject‐dependent analysis; support vector machine with rbf kernel (SVM), random forest (RF) with n_estimators 500 and max depth 11 and multi‐layer perception classifier (MLP) (layer size 32.64). In addition, we used a convolutional neural network (CNN) (layer size 128, activation function dense: softmax) algorithm for subject‐independent analysis and a k‐nearest neighbours classifier (KNN) (2 nearest neighbours) for subject‐dependent analysis. SVM, RF, MLP and KNN all used a labelled encoding approach, whereas CNN used a one‐hot encoding approach (Potdar et al. 2017). 70 percent of data were used to train the model which was then tested against the remaining 30% (e.g., subject‐dependent model used 70% of data per participant for training and testing on the remaining 30% of data). For the subject‐dependent model, results were calculated based on each individual subject's data, then the mean scores of all subjects (and SD) were reported. Whereas for the subject‐independent model, accuracy was trained on the whole data set (70%) and tested against the next segment.

Of the 30% of data used to test the model, 15% of data were tested via k‐cross fold validation (valid accuracy) and 15% of the data were tested as a percentage calculation for accuracy (test accuracy). Lastly, we calculated F1 scores, a measure of the harmonic mean of the model's precision and recall. Precision is a measure of the false positives and recall is a measure of the false negatives (Grandini et al. 2020). The model was tested comparing the accuracy of a combined 3‐sensor system (trunk, arm and wrist), with the individual accuracy of the wrist, arm or trunk sensor. Lastly, we compared accuracy between participants with rotator cuff tendinopathy and matched participants without shoulder pain (separate models were trained for each group).

2.6. Statistical Analysis Descriptive Data

Descriptive statistical analysis was planned for all continuous outcomes. Normality was assessed for each measure using a Shapiro‐Wilk test to determine the appropriate significance test (parametric or non‐parametric) (Mishra et al. 2019). An independent samples two‐sided Welch t‐test was used to compare baseline characteristics and 1RM strength test results between the rotator cuff tendinopathy group and the matched controls for parametric data (West 2021). A Mann‐Whitney U‐test was used for non‐parametric data (West 2021). Significance was considered p < 0.05. All tests were conducted using R (R: A languag and e and environment for statistical computing 2020).

3. Results

Participant characteristics and demographics are summarised below in Table 2. Waist circumference measure was the only baseline outcome with a statistically significant difference between the groups (p value = 0.04, two sample t‐test). For the 1RM strength test results, the lateral shoulder raise showed a statistically significant difference between groups (see Table 3).

TABLE 2.

Mean sociodemographic and anthropometric characteristics of participants.

Baseline characteristic Rotator cuff tendinopathy mean (SD) (n = 18) Pain free shoulder control mean (SD) (n = 18) All mean (SD) (n = 36) p value
Age a (years) 54.2 (13.2) 54.3 (12.9) 54.3 (13.1) 0.96
Sex
Female 9 9 18 —
Male 9 9 18 —
Anthropology
Height (cm) 170.2 (9.6) 173.8 (8.5) 172.0 (9.2) 0.25
Weight (kg) 88.8 (17.5) 81.4 (9.0) 85.1 (14.4) 0.13
Waist (cm)* 98.6* (15.6) 88.9* (10.2) 93.8 (14.1) 0.04*
BMI a 30.8 (6.3) 27.1 (4.0) 28.9 (5.6) 0.05
Hand dominance
Right 15 18 33 —
Left 3 0 3 —
Painful shoulder
Right 14 — — —
Left 4 — — —
Pain duration (years) 5.4 (6.0) — — —
Pain (NRS past week) 4.0 (2.4) — — —
SPADI
Pain total 45.9 (21.6) — — —
Disability total 28.2 (21.3) — — —
Total SPADI 35.0 (20.2) — — —
Open in a new tab

Abbreviations: 1RM, One repetition maximum; SD, standard deviation.

*

p < 0.05.

a

non‐parametric test.

TABLE 3.

Mean 1RM shoulder strength tests.

1RM strength tests (kg) Rotator cuff tendinopathy mean (SD) (n = 18) Pain free shoulder control mean (SD) (n = 18) All mean (SD) (n = 36) p value
Shoulder press 12.8 (7.2) 14.5 (5.0) 13.7 (6.2) 0.19
Lateral raise* 6.4 (3.4) 8.4 (2.3) 7.4 (3.1) 0.01*
Bent row 27.9 (12.8) 30.7 (8.4) 29.3 (10.8) 0.18
Open in a new tab

Abbreviations: 1RM, One repetition maximum; SD, standard deviation.

*

p < 0.05.

3.1. Accuracy of Different Machine Learning Models

Results of the machine learning models are summarised in Table 4 (subject‐independent analysis) and Table 5 (subject‐dependent analysis). Of the machine learning algorithms tested, a convolutional neural network (CNN) produced the best results for the subject‐independent data, while the random‐forest (RF) algorithm was best for the subject‐dependent data.

TABLE 4.

Machine learning subject‐independent classification results.

Rotator cuff tendinopathy Pain free All
Valid accuracy % (SD) Test accuracy % (SD) F1 (SD) Valid accuracy % (SD) Test accuracy % (SD) F1 (SD) Valid accuracy % (SD) Test accuracy % (SD) F1 (SD)
Wrist
SVM 75.52 (18.4) 79.51 (19.2) 0.80 (0.18) 79.28 (16.0) 83.23 (12.8) 0.81 (0.142) 78.16 (17.6) 84.08 (16.3) 0.85 (0.125)
RF 75.4 (23.9) 80.36 (19.6) 0.81 (0.185) 79.79 (24.3) 80.87 (14.1) 0.79 (0.168) 78.46 (44.0) 85.61 (17.6) 0.87 (0.14)
MLP 76.85 (33.3) 67.86 (11.4) 0.77 (0.188) 76.91 (21.1) 70.57 (18.0) 0.70 (0.184) 77.11 (16.6) 79.85 (14.4) 0.81 (0.111)
CNN 88.2 (6.5) 88.55 (4.5) 0.85 (0.08) 90 (7.5) 90.1 (5.5) 0.89 (0.07) 92.2 (6.5) 95.55 (4.5) 0.89 (0.089)
Arm
SVM 74.64 (20.2) 76.87 (23.5) 0.76 (0.236) 76.49 (17.2) 77.53 (14.8) 0.75 (0.165) 77.29 (17.3) 82.96 (16.9) 0.83 (0.147)
RF 72.52 (27.9) 80.36 (19.6) 0.81 (0.185) 71.78 (21.4) 69.97 (24.9) 0.72 (0.223) 79.78 (23.9) 79.56 (16.9) 0.81 (0.149)
MLP 76.85 (33.3) 75.92 (19.6) 0.77 (0.188) 66.92 (31.2) 65.60 (21.6) 0.66 (0.206) 75.11 (26.7) 77.93 (15.2) 0.79 (0.129)
CNN 89 (3.5) 90.35 (4.5) 0.9 (0.04) 90.3 (4.5) 91.5 (4.3) 0.90 (0.05) 93.2 (5.5) 93.51 (4.5) 0.90 (0.08)
Trunk
SVM 68.85 (15.3) 67.65 (11.4) 0.66 (0.117) 74.08 (13.9) 58.66 (20.5) 0.58 (0.192) 68.23 (19.1) 63.00 (17.0) 0.63 (0.158)
RF 76.61 (31.5) 76.05 (9.79) 0.75 (0.103) 69.62 (21.3) 60.84 (20.2) 0.59 (0.194) 61.75 (21.9) 66.42 (19.1) 0.68 (0.167)
MLP 71.30 (31.7) 60.47 (18.3) 0.66 (0.121) 64.13 (31.4) 60.47 (18.3) 0.59 (0.177) 61.06 (18.4) 64.69 (17.0) 0.64 (0.167)
CNN 78.17 (7.97) 80.57 (6.7) 0.68 (0.105) 85.2 (5.5) 87.55 (6.5) 0.86 (0.05) 86 (14.6) 87 (15.7) 0.84 (0.12)
3‐Sensor
SVM 74.25 (18.1) 81.58 (12.9) 0.81 (0.152) 74.03 (22.8) 78.18 (24.4) 0.79 (0.223) 80.29 (12.3) 85.96 (10.9) 0.84 (0.16)
RF 78.10 (38.4) 86.75 (20.3) 0.89 (0.16) 61.20 (37.6) 80.58 (27.1) 0.82 (0.238) 78.78 (20) 81.56 (17.9) 0.81 (0.15)
MLP 75.20 (22) 80.5 (15) 0.81 (0.14) 64.13 (34.4) 77.05 (23.2) 0.78 (0.211) 77.11 (22.2) 80.93 (15.2) 0.80 (0.14)
CNN 90.20 (5.4) 93.05 (4.5) 0.90 (0.056) 91.20 (4.6) 94.55 (5.5) 0.90 (0.06) 93.00 (5.1) 94.00 (4.5) 0.93 (0.045)
Open in a new tab

Abbreviations: 3‐sensor = combined wrist, arm and trunk sensor; CNN, convolutional neural network layer size [128], activation_function_dense, softmax; MLP, Multi‐layer perceptron classifier, layer size [32.64]; RF, Random forrest (n_estimators_500_max_depth_11); SVM, support vector machine with rbf kernel.

TABLE 5.

Machine learning subject‐dependent classification mean results.

Rotator cuff tendinopathy Pain free All
Valid accuracy % (SD) Test accuracy % (SD) F1 (SD) Valid accuracy % (SD) Test accuracy % (SD) F1 (SD) Valid accuracy % (SD) Test accuracy % (SD) F1 (SD)
Wrist
SVM 88.57 (8.19) 91.63 (3.39) 0.90 (0.06) 87.98 (13.50) 90.01 (14.26) 0.88 (0.15) 88.28 (11.01) 90.92 (10.25) 0.89 (0.11)
RF 91.91 (4.25) 93.47 (3.04) 0.91 (0.08) 86.51 (14.06) 90.33 (12.92) 0.88 (0.15) 89.21 (10.60) 91.90 (9.39) 0.90 (0.12)
MLP 87.70 (5.31) 92.13 (3.61) 0.89 (0.08) 85.67 (14.83) 88.85 (12.53) 0.86 (0.14) 86.69 (11.03) 90.49 (9.23) 0.88 (0.11)
KNN 87.47 (7.10) 89.81 (5.22) 0.87 (0.08) 85.56 (13.26) 88.21 (13.08) 0.86 (0.14) 86.51 (10.52) 89.01 (9.85) 0.87 (0.11)
Arm
SVM 88.95 (6.21) 92.05 (3.78) 0.90 (0.06) 85.62 (13.75) 88.18 (13.51) 0.86 (0.14) 87.29 (10.66) 90.11 (9.97) 0.88 (0.11)
RF 90.25 (4.26) 93.45 (3.24) 0.91 (0.08) 85.89 (16.02) 91.32 (13.51) 0.89 (0.14) 88.07 (11.76) 92.38 (9.40) 0.90 (0.12)
MLP 89.24 (6.69) 91.23 (4.45) 0.89 (0.08) 85.12 (11.96) 89.46 (13.30) 0.87 (0.14) 87.18 (9.78) 90.35 (9.81) 0.88 (0.11)
KNN 88.09 (6.16) 90.31 (4.93) 0.87 (0.08) 85.71 (13.22) 88.07 (13.07) 0.86 (0.14) 86.90 (10.24) 89.19 (9.80) 0.86 (0.12)
Trunk
SVM 89.17 (7.09) 92.21 (3.40) 0.90 (0.06) 86.06 (13.21) 90.06 (13.08) 0.89 (0.14) 87.61 (10.57) 91.13 (9.48) 0.90 (0.10)
RF 91.75 (3.50) 93.72 (3.22) 0.91 (0.09) 86.11 (12.58) 90.86 (12.97) 0.89 (0.15) 88.92 (9.54) 92.28 (9.43) 0.90 (0.12)
MLP 89.02 (5.13) 92.49 (3.67) 0.90 (0.07) 85.49 (14.31) 89.48 (13.72) 0.87 (0.15) 87.25 (10.74) 90.98 (10.02) 0.89 (0.12)
KNN 87.63 (7.20) 89.10 (5.11) 0.86 (0.07) 86.51 (14.91) 88.34 (13.58) 0.86 (0.15) 87.06 (11.55) 88.71 (10.12) 0.86 (0.12)
3‐Sensor
SVM 93.73 (2.99) 94.76 (2.51) 0.93 (0.07) 89.29 (14.79) 91.67 (13.26) 0.89 (0.16) 91.51 (10.76) 93.22 (9.53) 0.91 (0.11)
RF 95.30 (2.85) 96.12 (2.89) 0.94 (0.08) 91.74 (12.12) 94.66 (11.11) 0.92 (0.16) 93.52 (8.86) 95.39 (8.03) 0.93 (0.13)
MLP 92.58 (5.62) 95.32 (2.37) 0.94 (0.07) 90.82 (12.22) 92.99 (11.64) 0.91 (0.14) 91.70 (9.42) 94.15 (8.36) 0.92 (0.11)
KNN 91.16 (4.73) 93.62 (4.49) 0.92 (0.06) 89.98 (13.07) 92.79 (10.56) 0.90 (0.14) 90.57 (9.70) 93.21 (8.03) 0.912 (0.11)
Open in a new tab

Abbreviations: 3‐sensor, combined wrist, arm and trunk sensor; KNN, K nearest neighbour classifier_nn_2; MLP, Multi‐layer perceptron classifier, layer size [32.64]; RF, Random forrest (n_estimators_500_max_depth_11); SVM, support vector machine with rbf kernel.

3.2. Accuracy of Different Sensor Positions

The combined 3‐sensor system (trunk, wrist and arm) outperformed the individual sensors for the subject‐dependent results with accuracy of 95.39% (SD 8.3). Accuracy scores for the wrist, arm and trunk were 91.90% (SD 9.3), 92.38% (SD 9.40) and 92.28% (SD 9.43) respectively. For the subject‐independent data, the wrist sensor performed best with 95.55% (SD 4.5) accuracy. Accuracy scores for the trunk, arm and 3‐sensor system were 87% (SD 15.7), 93.51% (SD 4.5) and 94% (SD 4.5) respectively. Statistically significant differences were found for the subject‐independent data comparing the wrist position and the trunk position (p = 0.003). All other comparisons were not statistically significant.

3.3. Comparisons Between Rotator Cuff Tendinopathy Group and Pain Free Group

Comparing accuracy results between the rotator cuff tendinopathy group and the matched control pain free group shows results for the subject‐independent analysis were less accurate in the rotator cuff tendinopathy group for all sensor positions (93.05%, SD 4.5 compared with 94.55%, SD 5.5 for the 3‐sensor system). This may be due to greater variability within the rotator cuff tendinopathy group. Whereas, for the subject‐dependent data the reverse was true and results were more accurate in the rotator cuff tendinopathy group (96.12%, SD 2.89 compared with 94.66%, SD 11.11 for the 3‐sensor system). Neither comparison was statistically significant.

3.4. Accuracy for Each Exercise

For the subject‐independent data, k‐fold cross validation confusion matrix results can be found in Figure 4 for the CNN algorithm. For both groups combined, best accuracy was achieved by the wrist sensor (94.33%), identifying the shoulder press with 96% accuracy, the lateral raise with 94% accuracy and the bent over row with 93% accuracy.

FIGURE 4.

FIGURE 4

Open in a new tab

K‐Fold cross validation confusion matrix—subject‐independent model, convolutional neural network (CNN).

4. Discussion

This study demonstrates the accuracy of artificial intelligence, utilising machine learning to automatically classify shoulder exercise for both patients with rotator cuff tendinopathy and a matched pain free cohort. Classification accuracy is similar between a subject‐dependent model and subject‐independent model. Overall, accuracy was greater for the entire cohort with a subject‐independent model which may reflect the larger data pool available. Interestingly, utilising a single wrist based sensor achieves similar results to the 3‐sensor system (within 3%–5% accuracy) for the entire cohort using a subject‐independent analysis.

The ability to automatically detect and recognise exercise via wrist worn wearable technology presents a promising opportunity to objectively provide clinically meaningful data to clinicians and researchers, to monitor patient adherence and to provide direct patient feedback. External feedback has been demonstrated to increase patient adherence to exercise, physical activity and improve exercise technique (Ferguson et al. 2022; Scheid and West 2019; Benjaminse et al. 2015; Benjaminse and Otten 2011; Wulf 2013; Brennan et al. 2019). However, there remains a number of barriers to implement this technology from the lab into meaningful and useful clinical practice or larger research trials investigating the effects of exercise. These barriers include implementing automatic exercise detection among everyday activity, time requirements to ensure the system operates as intended and usability of the hardware or software. Integration of objective exercise monitoring in larger research studies may help better understand the various mechanisms and pathways of exercise and how they may impact patient outcomes.

Our findings are consistent with other literature findings that have investigated exercise classification and activity recognition. Several prior studies have investigated predominantly lower limb or daily activity classification via IMU sensor signal (Cheng et al. 2017; Morris et al. 2014; O'Reilly et al. 2017; Seeger et al. 2011; Clarke et al. 2017; Cleland et al. 2013; Doherty et al. 2017; Khan et al. 2013; Narayanan et al. 2020; O’Reilly et al. 2018; Parkka et al. 2007; Song et al. 2010; Stewart et al. 2018; Tavakkoli et al. 2022; Yang and Hsu 2010; Brickwood et al. 2019). There are fewer studies that have investigated upper limb or shoulder exercise classification (Um et al. 2017; D. M. Burns et al. 2018; Chang et al. 2007; Muehlbauer et al. 2011; Carnevale et al. 2019; Pan et al. 2013; Shoaib et al. 2016; Steels et al. 2020). Similarly, to our knowledge, there are only three studies that investigated exercise classification among a cohort of patients with rotator cuff tendinopathy (D. Burns et al. 2021; Bavan et al. 2019; Bavan et al. 2020).

(D. M. Burns et al. 2018) reported accuracies up to 99.4% utilising a subject‐dependent convolutional recurrent network approach and an accuracy of up to 88.9% utilising a subject‐independent approach (data outside their training set). Likewise, our own model demonstrated significantly improved results using a deep learning convolutional neural network analysis as opposed to simpler similar machine learning methods (e.g., RF, SVM or MLP algorithms). In a population of patients with rotator cuff tendinopathy, (D. Burns et al. 2021) reported exercise classification (see Supporting Information S1: Appendix for list of exercises) to be 90% (F1 score 0.82). (Bavan et al. 2019) in their lab‐based study of patients with rotator cuff tendinopathy (subacromial pain) reported 97.2% accuracy using a 10‐fold cross validation support vector machine approach. These results are comparable to our own.

A key strength of this study is the inclusion and comparison of people with rotator cuff tendinopathy, with pain free matched controls, as well as the comparison of multiple machine learning classification approaches. The main limitation of this study is the small sample of subjects and the lower average SPADI and pain scores for the participants indicating more mild limitation in pain and function. It is not clear how accurate this model will perform in a real world setting or how easily it would be to detect shoulder exercise among everyday activity. However, the minimal feedback or exercise cues provided in this study aim to replicate movement patterns closer to those likely performed in unsupervised situations.

5. Practical Applications

This study demonstrates the feasibility of automatic exercise classification and recognition for people completing shoulder exercises, either with rotator cuff tendinopathy or matched pain free controls. IMUs are readily accessible, affordable and available in most smartphone devices. There is potential for software development to compliment research studies to implement widespread use of this technology in monitoring exercise adherence while providing feedback regarding performance. Automated systems reduce the barriers or number of steps required to utilise such monitoring devices currently available.

This system has the potential to provide clinically meaningful data to treating practitioners either directly in person or via remote monitoring. The use of technology to provide automated exercise monitoring presents new opportunities for better understanding exercise parameters and their effect on clinical outcomes. This may lead to new metrics to monitor exercise, such as in clinical measures for rate of force development or newly developed metrics. Further, the use of technology provides an efficient means to increase scale in either research based projects or with large numbers of patients such as training athletes or monitoring gym attendees, including determining exercise variables such as volume, intensity, frequency and duration.

6. Conclusion

This study indicates that classifying shoulder exercise using a practical wrist worn sensor is highly accurate. Further, a similar level of accuracy was identified for subject‐independent and subject‐dependent analyses and among people with rotator cuff tendinopathy and a pain free cohort. This technology can now be used to investigate exercise recognition and classification during everyday activity and explore the effects of feedback from that information on adherence and patient outcomes.

Funding

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting Information S1

EJSC-26-e70167-s001.docx (315.9KB, docx)

Acknowledgements

Open access publishing facilitated by La Trobe University, as part of the Wiley ‐ La Trobe University agreement via the Council of Australasian University Librarians.

References

  1. Akyol, Y. , Ulus Y., Durmus D., et al. 2013. “Shoulder Muscle Strength in Patients With Subacromial Impingement Syndrome: Its Relationship With Duration of Quality of Life and Emotional Status.” Turkiye Fiziksel Tip ve Rehabilitasyon Dergisi 59: 176–181. 10.4274/tftr.59837. [DOI] [Google Scholar]
  2. Almeida, E. Pd , Salgasso M. E., Oliveira A. Sd, et al. 2020. “Comparison of Isometric Muscle Strength Between Painful and Non‐Painful Shoulders in Patients With Symptoms of Subacromial Impingement.” BrJP 3, no. 4. 10.5935/2595-0118.20200201. [DOI] [Google Scholar]
  3. Baechle, T. R. , Earle R. W., National S., et al. 2008. Essentials of Strength Training and Conditioning. 3rd ed. Human Kinetics. [Google Scholar]
  4. Balki, I. , Amirabadi A., Levman J., et al. 2019. “Sample‐Size Determination Methodologies for Machine Learning in Medical Imaging Research: A Systematic Review.” Canadian Association of Radiologists Journal 70, no. 4: 344–353. 10.1016/j.carj.2019.06.002. [DOI] [PubMed] [Google Scholar]
  5. Bavan, L. , Surmacz K., Beard D., Mellon S., and Rees J.. 2019. “Adherence Monitoring of Rehabilitation Exercise With Inertial Sensors: A Clinical Validation Study.” Gait & Posture 70: 211–217. 10.1016/j.gaitpost.2019.03.008. [DOI] [PubMed] [Google Scholar]
  6. Bavan, L. , Wood J., Surmacz K., Beard D., and Rees J.. 2020. “Instrumented Assessment of Shoulder Function: A Study of Inertial Sensor Based Methods.” Clinical Biomechanics 72: 164–171. 10.1016/j.clinbiomech.2019.12.010. [DOI] [PubMed] [Google Scholar]
  7. Benjaminse, A. , Gokeler A., Dowling A. V., et al. 2015. “Optimization of the Anterior Cruciate Ligament Injury Prevention Paradigm: Novel Feedback Techniques to Enhance Motor Learning and Reduce Injury Risk.” Journal of Orthopaedic & Sports Physical Therapy 45, no. 3: 170–182. 10.2519/jospt.2015.4986. [DOI] [PubMed] [Google Scholar]
  8. Benjaminse, A. , and Otten E.. 2011. “ACL Injury Prevention, More Effective With a Different Way of Motor Learning?” Knee Surgery, Sports Traumatology, Arthroscopy 19, no. 4: 622–627. 10.1007/s00167-010-1313-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Berkovitch, Y. , Haddad M., Keren Y., et al. 2013. “The Diagnostic Value of the Shoulder Rotator Cuff Muscles°Ø Isometric Force Testing.” Open Journal of Clinical Diagnostics 3, no. 6. 10.4236/ojcd.2013.33025. [DOI] [Google Scholar]
  10. Brennan, L. , Dorronzoro Zubiete E., and Caulfield B.. 2019. “Feedback Design in Targeted Exercise Digital Biofeedback Systems for Home Rehabilitation: A Scoping Review.” Sensors 20, no. 1: 181: [published Online First: 2020/01/08]. 10.3390/s20010181. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Brickwood, K.‐J. , Watson G., O'Brien J., and Williams A. D.. 2019. “Consumer‐Based Wearable Activity Trackers Increase Physical Activity Participation: Systematic Review and Meta‐Analysis.” JMIR mHealth and uHealth 7, no. 4: e11819. 10.2196/11819. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. British Orthopaedic Association . 2014. “Commissioning Guide: Subacromial Shoulder Pain.” https://www.boa.ac.uk/uploads/assets/uploaded/f4bfe04a‐0450‐4eab‐b9acad9dbb5d8c86.pdf.
  13. Burns, D. , Boyer P., Razmjou H., Richards R., and Whyne C.. 2021. “Adherence Patterns and Dose Response of Physiotherapy for Rotator Cuff Pathology: Longitudinal Cohort Study.” JMIR Rehabilitation and Assistive Technologies 8, no. 1: e21374: [published Online First: 2021/03/12]. 10.2196/21374. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Burns, D. M. , Leung N., Hardisty M., Whyne C. M., Henry P., and McLachlin S.. 2018. “Shoulder Physiotherapy Exercise Recognition: Machine Learning the Inertial Signals From a Smartwatch.” Physiological Measurement 39, no. 7: 075007. 10.1088/1361-6579/aacfd9. [DOI] [PubMed] [Google Scholar]
  15. Carnevale, A. , Longo U. G., Schena E., et al. 2019. “Wearable Systems for Shoulder Kinematics Assessment: A Systematic Review.” BMC Musculoskeletal Disorders 20, no. 1: 546. 10.1186/s12891-019-2930-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Chezar, A. , Berkovitch Y., Haddad M., Keren Y., Soudry M., and Rosenberg N.. 2013. “Normal Isometric Strength of Rotatorcuff Muscles in Adults.” Bone & Joint Research 2, no. 10: 214–219: [published Online First: 2013/10/09]. 10.1302/2046-3758.210.2000202. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Clarke, C. L. , Taylor J., Crighton L. J., Goodbrand J. A., McMurdo M. E. T., and Witham M. D.. 2017. “Validation of the AX3 Triaxial Accelerometer in Older Functionally Impaired People.” Aging‐Clinical & Experimental Research 29, no. 3: 451–457: [published Online First: 2016/07/19]. 10.1007/s40520-016-0604-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Clausen, M. B. , Hölmich P., Rathleff M., et al. 2021. “Effectiveness of Adding a Large Dose of Shoulder Strengthening to Current Nonoperative Care for Subacromial Impingement: A Pragmatic, Double‐Blind Randomized Controlled Trial (SExSI Trial).” American Journal of Sports Medicine 49, no. 11: 3040–3049. 10.1177/03635465211016008. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Clausen, M. B. , Witten A., Holm K., et al. 2017. “Glenohumeral and Scapulothoracic Strength Impairments Exists in Patients With Subacromial Impingement, but these Are Not Reflected in the Shoulder Pain and Disability Index.” BMC Musculoskeletal Disorders 18, no. 1: 302. 10.1186/s12891-017-1667-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Cleland, I. , Kikhia B., Nugent C., et al. 2013. “Optimal Placement of Accelerometers for the Detection of Everyday Activities.” Sensors 13, no. 7: 9183–9200: [published Online First: 2013/07/23]. 10.3390/s130709183. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Cuschieri, S. 2019. “The STROBE Guidelines.” Supplement, Saudi Journal of Anaesthesia 13, no. Suppl 1: S31–s34: [published Online First: 2019/04/02]. 10.4103/sja.SJA_543_18. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Diercks, R. , Bron C., Dorrestijn O., et al. 2014. “Guideline for Diagnosis and Treatment of Subacromial Pain Syndrome: A Multidisciplinary Review by the Dutch Orthopaedic Association.” Acta Orthopaedica 85, no. 3: 314–322: [published Online First: 2014/05/23]. 10.3109/17453674.2014.920991. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Doherty, A. , Jackson D., Hammerla N., et al. 2017. “Large Scale Population Assessment of Physical Activity Using Wrist Worn Accelerometers: The UK Biobank Study.” PLoS One 12, no. 2: e0169649. 10.1371/journal.pone.0169649. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Parkka J., Ermes M., Antila K., van Gils M., Manttari A. and Nieminen H.. 2007. “Estimating Intensity of Physical Activity: A Comparison of Wearable Accelerometer and Gyro Sensors and 3 Sensor Locations.” 2007 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society: 1511–1514. 10.1109/IEMBS.2007.4352588. [DOI] [PubMed] [Google Scholar]
  25. Um T. T., Babakeshizadeh V. and Kulić D.. 2017. “Exercise Motion Classification From Large‐Scale Wearable Sensor Data Using Convolutional Neural Networks.” 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS): 2385–2390. 10.1109/IROS.2017.8206051. [DOI] [Google Scholar]
  26. Faigenbaum, A. D. , McFarland J. E., Herman R. E., et al. 2012. “Reliability of the One‐Repetition‐Maximum Power Clean Test in Adolescent Athletes.” Journal of Strength & Conditioning Research 26, no. 2: 432–437. 10.1519/JSC.0b013e318220db2c. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Ferguson, T. , Olds T., Curtis R., et al. 2022. “Effectiveness of Wearable Activity Trackers to Increase Physical Activity and Improve Health: A Systematic Review of Systematic Reviews and Meta‐Analyses.” Lancet Digital Health 4, no. 8: e615–e626. 10.1016/S2589-7500(22)00111-X. [DOI] [PubMed] [Google Scholar]
  28. Gorzelitz, J. , Farber C., Gangnon R., and Cadmus‐Bertram L.. 2020. “Accuracy of Wearable Trackers for Measuring Moderate‐ to Vigorous‐Intensity Physical Activity: A Systematic Review and Meta‐Analysis.” Journal for the Measurement of Physical Behaviour 3, no. 4: 346–357. 10.1123/jmpb.2019-0072. [DOI] [Google Scholar]
  29. Grandini, M. , Bagli E., and Visani G.. 2020. “Metrics for Multi‐Class Classification: An Overview.” ArXiv: abs/2008.05756.
  30. Hopewell, S. , Keene D. J., Marian I. R., et al. 2021. “Progressive Exercise Compared With Best Practice Advice, With or Without Corticosteroid Injection, for the Treatment of Patients With Rotator Cuff Disorders (GRASP): A Multicentre, Pragmatic, 2x2 Factorial, Randomised Controlled Trial.” Lancet 398, no. 10298: 416–428. 10.1016/S0140-6736(21)00846-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Hopman, K. , Lee K. S., Lukersmith S., et al. 2013. Clinical Practice Guidelines for the Management of Rotator Cuff Syndrome in the Workplace. Port Macquarie. University of New South Wales. [Google Scholar]
  32. Ingwersen, K. G. , Jensen S. L., Sørensen L., et al. 2017. “Three Months of Progressive High‐Load Versus Traditional Low‐Load Strength Training Among Patients With Rotator Cuff Tendinopathy: Primary Results From the Double‐Blind Randomized Controlled RoCTEx Trial.” Orthopaedic Journal of Sports Medicine 5, no. 8: 2325967117723292: [published Online First: 2017/09/07]. 10.1177/2325967117723292. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Khan, A. M. , Kalkbrenner G., and Lawo M.. 2013. Recognizing Physical Training Exercises Using the Axivity Device.
  34. Klintberg, I. H. , Cools A. M. J., Holmgren T. M., et al. 2015. “Consensus for Physiotherapy for Shoulder Pain.” International Orthopaedics 39, no. 4: 715–720. 10.1007/s00264-014-2639-9. [DOI] [PubMed] [Google Scholar]
  35. Larsson, R. , Bernhardsson S., and Nordeman L.. 2019. “Effects of Eccentric Exercise in Patients With Subacromial Impingement Syndrome: A Systematic Review and Meta‐Analysis.” BMC Musculoskeletal Disorders 20, no. 1: 446: [published Online First: 2019/10/16]. 10.1186/s12891-019-2796-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Levinger, I. , Goodman C., Hare D. L., Jerums G., Toia D., and Selig S.. 2009. “The Reliability of the 1RM Strength Test for Untrained Middle‐Aged Individuals.” Journal of Science and Medicine in Sport 12, no. 2: 310–316. 10.1016/j.jsams.2007.10.007. [DOI] [PubMed] [Google Scholar]
  37. Li, K. , Elgalad A., Cardoso C., and Perin E. C.. 2022. “Using the Apple Watch to Record Multiple‐Lead Electrocardiograms in Detecting Myocardial Infarction: Where Are We Now?” Texas Heart Institute Journal/from the Texas Heart Institute of St Luke's Episcopal Hospital, Texas Children's Hospital 49, no. 4: [published Online First: 2022/08/05]. 10.14503/thij-22-7845. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Liaghat, B. , Skou S. T., Søndergaard J., Boyle E., Søgaard K., and Juul‐Kristensen B.. 2020. “A Randomised Controlled Trial of Heavy Shoulder Strengthening Exercise in Patients With Hypermobility Spectrum Disorder or Hypermobile Ehlers‐Danlos Syndrome and Long‐Lasting Shoulder Complaints: Study Protocol for the Shoulder‐Mobilex Study.” Trials 21, no. 1: 992. 10.1186/s13063-020-04892-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Littlewood, C. , Bateman M., Connor C., et al. 2019. “Physiotherapists’ Recommendations for Examination and Treatment of Rotator Cuff Related Shoulder Pain: A Consensus Exercise.” Physiotherapy Practice and Research 40, no. 2: 87–94. 10.3233/PPR-190129. [DOI] [Google Scholar]
  40. Littlewood, C. , Malliaras P., and Chance‐Larsen K.. 2015. “Therapeutic Exercise for Rotator Cuff Tendinopathy: A Systematic Review of Contextual Factors and Prescription Parameters.” International Journal of Rehabilitation Research 38, no. 2: 95–106: [published Online First: 2015/02/26]. 10.1097/mrr.0000000000000113. [DOI] [PubMed] [Google Scholar]
  41. Luime, J. J. , Koes B. W., Hendriksen I. J., et al. 2004. “Prevalence and Incidence of Shoulder Pain in the General Population; A Systematic Review.” Scandinavian Journal of Rheumatology 33, no. 2: 73–81: [published Online First: 2004/05/28]. 10.1080/03009740310004667. [DOI] [PubMed] [Google Scholar]
  42. Malliaras, P. , Cridland K., Hopmans R., et al. 2020. “Internet and Telerehabilitation‐Delivered Management of Rotator Cuff‐Related Shoulder Pain (INTEL Trial): Randomized Controlled Pilot and Feasibility Trial.” JMIR mHealth and uHealth 8, no. 11: e24311: [published Online First: 2020/11/19]. 10.2196/24311. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Malliaras, P. , Johnston R., Street G., et al. 2020. “The Efficacy of Higher Versus Lower Dose Exercise in Rotator Cuff Tendinopathy: A Systematic Review of Randomized Controlled Trials.” Archives of Physical Medicine and Rehabilitation 101, no. 10: 1822–1834. 10.1016/j.apmr.2020.06.013. [DOI] [PubMed] [Google Scholar]
  44. Mishra, P. , Pandey C. M., Singh U., et al. 2019. “Descriptive Statistics and Normality Tests for Statistical Data.” Annals of Cardiac Anaesthesia 22, no. 1: 67–72: [published Online First: 2019/01/17]. 10.4103/aca.ACA_157_18. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Mokkink, L. B. , Prinsen C., Patrick D. L., et al. 2019. COSMIN Study Design Checklist for Patient‐reported Outcome Measurement Instruments, 1–32.
  46. Morris, D. , Saponas T. S., Guillory A., et al. 2014. “Recofit: Using a Wearable Sensor to Find, Recognize, and Count Repetitive Exercises.” In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 3225–3234. Association for Computing Machinery. [Google Scholar]
  47. Narayanan, A. , Stewart T., and Mackay L.. 2020. “A Dual‐Accelerometer System for Detecting Human Movement in a Free‐Living Environment.” Medicine & Science in Sports & Exercise 52, no. 1: 252–258: [published Online First: 2019/07/31]. 10.1249/mss.0000000000002107. [DOI] [PubMed] [Google Scholar]
  48. Naunton, J. , Street G., Littlewood C., Haines T., and Malliaras P.. 2020. “Effectiveness of Progressive and Resisted and Non‐Progressive or Non‐Resisted Exercise in Rotator Cuff Related Shoulder Pain: A Systematic Review and meta‐analysis of Randomized Controlled Trials.” Clinical Rehabilitation 34, no. 9: 1198–1216. 10.1177/0269215520934147. [DOI] [PubMed] [Google Scholar]
  49. O’Reilly, M. , Caulfield B., Ward T., Johnston W., and Doherty C.. 2018. “Wearable Inertial Sensor Systems for Lower Limb Exercise Detection and Evaluation: A Systematic Review.” Sports Medicine 48, no. 5: 1221–1246. 10.1007/s40279-018-0878-4. [DOI] [PubMed] [Google Scholar]
  50. O'Reilly, M. A. , Whelan D. F., Ward T. E., et al. 2017. “Technology in Strength and Conditioning Tracking Lower‐Limb Exercises With Wearable Sensors.” Journal of Strength & Conditioning Research 31, no. 6: 1726–1736. 10.1519/jsc.0000000000001852. [DOI] [PubMed] [Google Scholar]
  51. Page, M. J. , Green S., McBain B., et al. 2016. “Manual Therapy and Exercise for Rotator Cuff Disease.” Cochrane Database of Systematic Reviews 2016, no. 6. 10.1002/14651858.CD012224. [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Pan, J.‐I. , Chung H.‐W., and Huang J.‐J.. 2013. “Intelligent Shoulder Joint Home‐Based Self‐Rehabilitation Monitoring System.” International Journal of Smart Home 7, no. 5: 395–404. 10.14257/ijsh.2013.7.5.38. [DOI] [Google Scholar]
  53. Pedregosa, F. , Varoquaux G., Gramfort A., et al. 2011. “Scikit‐Learn: Machine Learning in Python.” Journal of Machine Learning Research 12: 2825–2830. 10.5555/1953048.2078195. [DOI] [Google Scholar]
  54. Potdar, K. , Pardawala T., and Pai C.. 2017. “A Comparative Study of Categorical Variable Encoding Techniques for Neural Network Classifiers.” International Journal of Computer Applications 175, no. 4: 7–9. 10.5120/ijca2017915495. [DOI] [Google Scholar]
  55. R: A language and environment for statistical computing . 2020. [Program]. R Version 4.0.2 (2020‐06‐22) Version. R Foundation for Statistical Computing. [Google Scholar]
  56. Rajput, D. , Wang W.‐J., and Chen C.‐C.. 2023. “Evaluation of a Decided Sample Size in Machine Learning Applications.” BMC Bioinformatics 24, no. 1: 48. 10.1186/s12859-023-05156-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  57. Cheng L., Guan Y., Zhu K. and Li Y.. 2017. “Recognition of Human Activities Using Machine Learning Methods With Wearable Sensors.” 2017 IEEE 7th Annual Computing and Communication Workshop and Conference (CCWC): 1–7. 10.1109/CCWC.2017.7868369. [DOI] [Google Scholar]
  58. Scheid, J. L. , and West S. L.. 2019. “Opportunities of Wearable Technology to Increase Physical Activity in Individuals With Chronic Disease: An Editorial.” International Journal of Environmental Research and Public Health 16, no. 17: 3124: [published Online First: 2019/08/31]. 10.3390/ijerph16173124. [DOI] [PMC free article] [PubMed] [Google Scholar]
  59. Schroeder, E. T. , Wang Y., Castaneda‐Sceppa C., et al. 2007. “Reliability of Maximal Voluntary Muscle Strength and Power Testing in Older Men.” Journals of Gerontology: Series A 62, no. 5: 543–549. 10.1093/gerona/62.5.543. [DOI] [PubMed] [Google Scholar]
  60. Seeger, C. , Buchmann A., and Van Laerhoven K.. 2011. Myhealthassistant: A Phone‐based Body Sensor Network that Captures the Wearer's Exercises Throughout the Day. 10.4108/icst.bodynets.2011.247015. [DOI]
  61. Ebert A., Kiermeier M., Marouane C. and Linnhoff‐Popien C.. 2017. “SensX: About Sensing and Assessment of Complex Human Motion.” 2017 IEEE 14th International Conference on Networking, Sensing and Control (ICNSC): 327–332. 10.1109/ICNSC.2017.8000113. [DOI] [Google Scholar]
  62. Shei, R.‐J. , Holder I. G., Oumsang A. S., Paris B. A., and Paris H. L.. 2022. “Wearable Activity Trackers–advanced Technology or Advanced Marketing?” European Journal of Applied Physiology 122, no. 9: 1975–1990. 10.1007/s00421-022-04951-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  63. Shen, C. , Ho B. J., and Srivastava M.. 2018. “Milift: Efficient Smartwatch‐Based Workout Tracking Using Automatic Segmentation.” IEEE Transactions on Mobile Computing 17, no. 7: 1609–1622. 10.1109/TMC.2017.2775641. [DOI] [Google Scholar]
  64. Shoaib, M. , Bosch S., Incel O. D., Scholten H., and Havinga P.. 2016. “Complex Human Activity Recognition Using Smartphone and Wrist‐Worn Motion Sensors.” Sensors 16, no. 4: 426. 10.3390/s16040426. [DOI] [PMC free article] [PubMed] [Google Scholar]
  65. Song, J. , Semanik P., Sharma L., et al. 2010. “Assessing Physical Activity in Persons With Knee Osteoarthritis Using Accelerometers: Data From the Osteoarthritis Initiative.” Arthritis Care & Research 62, no. 12: 1724–1732. 10.1002/acr.20305. [DOI] [PMC free article] [PubMed] [Google Scholar]
  66. Steels, T. , Van Herbruggen B., Fontaine J., De Pessemier T., Plets D., and De Poorter E.. 2020. “Badminton Activity Recognition Using Accelerometer Data.” Sensors 20, no. 17: 4685. 10.3390/s20174685. [DOI] [PMC free article] [PubMed] [Google Scholar]
  67. Stewart, T. , Narayanan A., Hedayatrad L., Neville J., Mackay L., and Duncan S.. 2018. “A Dual‐Accelerometer System for Classifying Physical Activity in Children and Adults.” Medicine & Science in Sports & Exercise 50, no. 12: 2595–2602: [published Online First: 2018/07/27]. 10.1249/mss.0000000000001717. [DOI] [PubMed] [Google Scholar]
  68. Tavakkoli, O. S. , Malliaras P., Hill K. D., Clark R., and Perraton L.. 2022. “Evaluating Daily Physical Activity and Biomechanical Measures Using Wearable Technology in People With Achilles Tendinopathy: A Descriptive Exploratory Study.” Musculoskeletal Science and Practice 58: 102534: [published Online First: 2022/02/28]. 10.1016/j.msksp.2022.102534. [DOI] [PubMed] [Google Scholar]
  69. Chang, K.‐H. , Chen, M. Y. and Canny, J. F. 2007. Tracking Free‐Weight Exercises. Springer Berlin Heidelberg. [Google Scholar]
  70. van der Windt, D. A. , Koes B. W., de Jong B. A., and Bouter L. M.. 1995. “Shoulder Disorders in General Practice: Incidence, Patient Characteristics, and Management.” Annals of the Rheumatic Diseases 54, no. 12: 959–964. 10.1136/ard.54.12.959. [DOI] [PMC free article] [PubMed] [Google Scholar]
  71. West, R. M. 2021. “Best Practice in Statistics: Use the Welch t‐Test When Testing the Difference Between Two Groups.” Annals of Clinical Biochemistry 58, no. 4: 267–269. 10.1177/0004563221992088. [DOI] [PubMed] [Google Scholar]
  72. Muehlbauer, M. , Bahle, G. and Lukowicz, P. 2011. “What Can an Arm Holster Worn Smart Phone Do for Activity Recognition?” 2011 15th Annual International Symposium on Wearable Computers. [Google Scholar]
  73. Wulf, G. 2013. “Attentional Focus and Motor Learning: A Review of 15 Years.” International Review of Sport and Exercise Psychology 6, no. 1: 77–104. 10.1080/1750984X.2012.723728. [DOI] [Google Scholar]
  74. Yang, C.‐C. , and Hsu Y.‐L.. 2010. “A Review of Accelerometry‐Based Wearable Motion Detectors for Physical Activity Monitoring.” Sensors 10, no. 8: 7772–7788. 10.3390/s100807772. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supporting Information S1

EJSC-26-e70167-s001.docx (315.9KB, docx)

Articles from European Journal of Sport Science are provided here courtesy of Wiley

RESOURCES