Skip to main content

Table 1 Studies involving the use of wearable sensors in the study of stroke. Only studies including actual patients shown. Most of the studies listed focused on the assessment of motor function through standardized clinical tests, which focus mainly on movement quality. This might explain the much more common use of IMU’s so far

From: Wearable technology in stroke rehabilitation: towards improved diagnosis and treatment of upper-limb motor impairment

Device Category

Sensors

Product / Units

Author

Assessment Type

Patients / Healthy controls

Methods and Results

IMU

6 accelerometers (hand, forearm, upper arm, sternum)

“Vitaport 3” (Temec®, Heerlen, NL) / m/s2

Hester et al., 2006 [19]

Upper limb motor assessment score prediction

12 / 0

10% relative error in prediction of clinical scores (leave-one-out cross-validation).

6 accelerometers (hand, forearm, upper arm, sternum)

“Vitaport 3” / m/s2

Patel et al., 2010 [23]

FAS- WMFT

24 / 0

5.76% relative error when predicting FAS.

1 IMU worn at the wrist

Non-commercial, / m/s2, deg/s

Parnandi et al., 2010 [24]

FAS- WMFT

1 / 0

“Prediction” error close to zero. *Model likely overfits the data, reasoning behind analysis might be incorrect.

6 accelerometers (hand, forearm, upper arm, sternum)

“Vitaport 3” / m/s2

Del Din et al., 2011 [25]

Motor function, FM Test

24 / 0

4-points off when using a single item of the WMFT to predict total UEFM score (max 66).

1 IMU attached to the forearm

MotionNode® (Seattle, USA) / m/s2, deg/s

Zhang et al., 2012 [48]

Upper limb movement trajectory comparison

2 / 1

Similarity between feature vectors for 5 exercises of the UEFM is evaluated using cosine distance. The authors claim that higher similarity (close to 0.9) corresponds to higher FM scores assigned by therapists. They compare feature vectors from affected and unaffected limbs in patients, but they never show how similar the vectors are in the healthy person.

10 IMUs (upper and lower extremities and trunk)

Non-commercial / m/s2, deg/s

Strohrmann et al., 2013 [52]

Changes in motor function over time

2 / 0

Longitudinal look at changes in motor function over the course of 4 weeks. Used linear regression. Mean RMSE was of 0.15 and correlation between regression estimate and ground truth (expert assessment) was of 0.86.

2 accelerometers, one per wrist

“GTX+” (ActiGraph®, Pensacola, USA) / m/s2

Bailey et al. 2015 [53]

Upper-limb bilateral activity to detect limb neglect during activities of daily life

48 / 74

Had participants wear accelerometers on each wrist for 26 h. Calculated the magnitude of the acceleration vector every second for each wrist, and a ratio of said vectors between the affected and non-affected hands. Were able to detect limb neglect in impaired patients (impairment level measured using ARAT test).

IMU from smartphone worn on the right-front hip

“Blackberry Z10” (BlackBerry®, Waterloo, CAN)/ m/s2, deg/s

Capela et al., 2015 [33]

Human activity recognition (6 activities)

12 / 15

Found common features for healthy individuals (young and elderly) and stroke patients to discriminate between different conditions of movement and stillness using a smartphone. Classification accuracy was over 80% for most of the levels of comparison (e.g. mobile vs. immobile, large movements vs. stairs, etc.) when using decision trees, and similar (if slightly lower) when using SVM or Naive Bayes.

3 IMUs (Lower arm, upper arm and trunk)

“ArmeoSenso” (Hocoma®, Volketswil, CH), “MotionPod 3” (Movea Inc.®, Pleasanton, USA) / m/s2, deg/s, Gauss

Wittman et al. 2016 [39]

Home-based rehabilitative training

11 / 0

Significant improvement of motor function as assessed by the FMA (4.1 points) and by metrics native to the “ArmeoSenso” system

2 accelerometers (forearm and upper arm)

Not specified / m/s2

Yu et al.,2016a [20]

Brunnstrom stage classifier

23 / 4

Used ELM to classify people into 5 of the 6 stages of the Brunnstrom Stage Evaluation. 80% of samples were used as training set. No cross-validation was done. All patients belonged to stages from II to V. Stage VI is considered to be unimpaired, so data acquired from healthy participants were used. Data were acquired during a single exercise (repeated several times) and used to predict Brunnstrom stage. Accuracy was above 85% when using ELM.

1 IMU worn at the forearm

“MTi-300” (Xsens®, Ensched, NL) / m/s2, deg/s

Zhang et al., 2016a [35]

Upper limb motion classification

14 / 0

Recorded inertial data from 14 patients (6 were relatively unimpaired). Used PCA and used top 7 components to label recordings according to the motion that generated them. The Fuzzy Kernel algorithm achieved an error rate of 0% when dealing with the 6 well-recovered patients, and of 0.56% for more impaired patients.

1 IMU worn at the forearm

“MPU-6050” (InvenSense®, San Jose, USA) / m/s2

Zhang et al., 2016b [49]

Upper limb motion assessment

21 / 8

Proposed a mobility index based on DTW to characterize patients’ movements and assign them to Brunnstrom stages from III-VI. Their index used with a KNN classification algorithm (k = 3) achieved an accuracy of 82% in leave-one-out cross validation.

2 accelerometers, one per wrist

“LSM9DS0” (Adafruit®, New York, USA)/ m/s2

de Lucena et al., 2017 [30]

Bimanual symmetry, jerk and clinical function to explain variance in upper limb recovery

9 / 0

Used PCA and concluded that the first component relates to functional status, whereas they suggest the second component might be related to movement quality (as it described a strong correlation (≥0.75) between acceleration asymmetry and jerk asymmetry). Both principal components were found to explain 86% of the variance.

1 IMU worn at the wrist

“ReSense” [54] / m/s2, deg/s

Leuenberger et al., 2017 [29]

Affected limb neglect during activities of daily life

10 / 0

Proposed a new measure for arm use called Gross Arm Movements, which detects changes in arm orientation larger than 30 degrees. This measure has large correlation to clinical tests (r > = 0.9) even when not removing signals acquired while patients walk.

2 IMUs, one per wrist

“Shimmer3” (Shimmer Research®, Ireland) / m/s2, deg/s

Lee et al., 2018 [28]

Neglect and exercise quality at home

20 / 10

Detection accuracy of goal-directed movements was described with a ROC curve, with an AUC of 87%. F-score (harmonic mean of precision and recall) of 84.3% when classifying movements into feedback vs no-feedback groups, an F-score of 73.7% when detecting feedback due to accuracy issues and an F-score of 65.3% when detecting feedback due to compensatory movement.

EMG

10 EMG electrodes on forearm and hand

Noraxon® (Scottsdale, USA) / mV

Lee et al., 2011 [55]

Classification of hand postures (6 classes)

20 / 0

LDA to classify signals into 6 hand gestures, with accuracies ranging from 37.9% (severely impaired subjects, Chedoke stage 2 and 3) to 71.3 (moderately impaired, Chedoke stage 4 and 5). Single model built for each patient, gradually adding more data to it.

89 EMG electrodes

“Refa 128” (TMSI®, Twente, NL) / mV

Zhang & Zhou, 2012 [56]

Classification of hand postures (20 classes)

12 / 0

Used Fisher linear discriminant analysis (PCA + LDA) for dimensionality reduction. Best performance (96% classification accuracy) was obtained using time-domain features and an SVM classifier. Achieved comparable results with only 8 electrodes, but do not specify which ones

 

2 EMG electrodes

Not specified / mV

Donoso Brown et al., 2015 [57]

Home-based gamified, rehabilitative training

10 / 0

Proved feasibility of this approach at home, and the system was described as engaging and motivating, but there were no reports of improved functionality transferred to activities of daily life

Pots. & Encoders

2 potentiometers

“SP12S-1 K” (ETI Systems®, Carlsbad, USA) / V mapped to angular displacement

Durfee et al., 2009 [58]

Hand joint angles tracked during proof of concept rehabilitative game

24 / 0

System of beams tracking wrist and index finger joint angles. No classification or other form of accuracy reported, as position was mapped directly from potentiometer readout.

1 encoder

“E4” (US Digital®, Vancouver, USA) / V mapped to angular displacement

Chen et al., 2017 [59]

Hand joint angles in 4-bar hand orthosis

10 / 0

Patients trained at home (only 7 finished training) during 4 weeks, training 5 times a week. By the end of training, patients showed motor improvement of 4.9 +/−  4.1 points in FM score, with a strong correlation (0.90) between amount of movements performed during training and score improvement.

Flexible sensors

Flex sensors along the dorsal side of fingers

“Flexpoint bend sensor” (Flexpoint Sensor Systems®, Draper, USA) / V mapped to joint flexion

Prange-Lasonder et al., 2017 [60]

Hand gesture tracking during rehabilitative training and assistive grasping

5 / 0

Presented a glove with two possible modalities (assistive and rehabilitative). Modest improvement in pinch force and execution of other tasks was reported, hinting towards potential benefits of its use as a rehabilitative/assistive tool.

Combinations

2 accelerometers (forearm and upper arm) and 7 flex sensors (dorsal side of fingers, wrist)

“ADXL345” (Analog Devices Inc.®, Norwood, USA) / m/s2, V mapped to joint flexion

Yu et al., 2016b [61]

Motor function, FM Test

24 / 0

Evaluated shortened version of UEFM using accelerometers and flex sensors. Model built using SVM (after using RRelief algorithm for feature selection) had a 0.92 correlation with clinical scores given by a therapist.

2 IMUs (wrist, upper arm) and 10 sEMG (forearm, biceps and triceps)

“MPU-9250” (InvenSense®, San Jose, USA), EMG not stated / m/s2, deg/s, mV

Li et al., 2017 [62]

Motor function, FM Test

18 / 16

34-leave-one-out cross validation, achieved determination coefficients (correlation between their proposed measure and FM score) of 0.85 using different unsupervised (PCA, NMF) and supervised (LASSO) algorithms.

8 EMG + IMU, worn at the forearm

“Myo” armband (Thalmic Labs®, Kitchener, Canada) / m/s2, deg/s, arbitrary units (Myo EMG)

Ryser et al., 2017 [63]

Control signals for a hand orthosis

3 / 0

Classification accuracy between 78 and 98% in stroke patients when discriminating between 3 hand gestures using SVM.

9 IMUs (hands, wrists, upper arms, forearms, sternum) and 16 EMG electrodes (worn at the forearms)

“Myo” armband, IMU not stated / m/s2, deg/s, arbitrary units (Myo EMG)

Repnik et al., 2018 [27]

ARAT test

28 / 12

Correlation of 0.60 between movement time and movement smoothness with respect to ARAT score. EMG data revealed significant differences in muscle activity of healthy subjects when grasping objects of different sizes. Normalized muscle activation revealed that, in more affected patients (ARAT score 2), maximal muscle activation was present when grasping the largest object, while in healthy participants activation was close to one third of maximal output.