Wearable sensors and machine learning in post-stroke rehabilitation assessment: A systematic review

Using Wearable Inertial Sensors to Estimate Clinical Scores of Upper Limb Movement Quality in Stroke

“…However, this statement remains speculative and further research would be required to systematically investigate how the number and placement of the sensor units, as well as the direct and indirect measurement of movements, contribute to the accuracy of clinical scores estimations. This question has never been addressed so far, neither in studies that estimated different clinical scores with larger numbers of sensors ( Patel et al, 2010 ; Adans-Dester et al, 2020 ), nor in reviews of clinical assessments with wearable sensors ( Oña Simbaña et al, 2019 ; Kim et al, 2021 ; Boukhennoufa et al, 2022 ).…”

“…The hardware costs of commercially available inertial sensors, approximately $50 per sensor unit, are relatively low as compared to those of optoelectrical camera systems, approximately $10’000 per system, which are the current gold-standard for clinical motion analysis. Additionally, the same set up is frequently used to measure other aspects of motor functioning ( Oña Simbaña et al, 2019 ; Kim et al, 2021 ; Boukhennoufa et al, 2022 ). For example, many studies collected data during activities of daily living or free-living settings and aimed to develop new measures of ICF activity performance, such as quantifications of impaired arm use (e.g., ( Bailey et al, 2015 ; Lee et al, 2019 )).…”

“…Due to the difficulties with establishing kinematic measures of movement quality studies started to explore an intermediate goal. Supervised machine learning algorithms and low-cost sensor data were used to estimate clinical test scores (for reviews see Oña Simbaña et al, 2019 ; Kim et al, 2021 ; Boukhennoufa et al, 2022 ). This approach has the advantages that the clinical tests have established psychometric properties ( Kim et al, 2021 ), that clinical scores are easy to interpret ( Kim et al, 2021 ) and that wearable movement sensors can be used which are low cost and enable data collection across the entire continuum of care ( Oña Simbaña et al, 2019 ; Kim et al, 2021 ; Boukhennoufa et al, 2022 ).…”

“…Supervised machine learning algorithms and low-cost sensor data were used to estimate clinical test scores (for reviews see Oña Simbaña et al, 2019 ; Kim et al, 2021 ; Boukhennoufa et al, 2022 ). This approach has the advantages that the clinical tests have established psychometric properties ( Kim et al, 2021 ), that clinical scores are easy to interpret ( Kim et al, 2021 ) and that wearable movement sensors can be used which are low cost and enable data collection across the entire continuum of care ( Oña Simbaña et al, 2019 ; Kim et al, 2021 ; Boukhennoufa et al, 2022 ). Tests of ICF activity capacity assess limitations in the accomplishment of tasks that are relevant for activities of daily living ( Prange-Lasonder et al, 2021 ).…”

“…The ARAT is the most frequently used assessment of upper-limb functioning in clinical studies ( Murphy et al, 2015 ), as it is used in a broad range of neurological health conditions such as stroke, traumatic brain injury, multiple sclerosis ( Prange-Lasonder et al, 2021 ) and Parkinson’s Disease ( Song, 2012 ), has excellent psychometric properties ( Pike et al, 2018 ), is widely accepted and recommended by experts ( Kwakkel et al, 2017 ; Pohl et al, 2020 ; Prange-Lasonder et al, 2021 ), and is a significant predictor of motor recovery in stroke ( Wolf et al, 2021 ). Despite the importance of the ARAT, however, wearable sensor data were never utilized to estimate the test outcome, to the best of our knowledge ( Oña Simbaña et al, 2019 ; Kim et al, 2021 ; Boukhennoufa et al, 2022 ).…”