Multimodal Sensing in Stroke Motor Rehabilitation

Applying sensors in biomedical institutions and home‐based stroke rehabilitation is now a global research focus. In this review paper, the relationship between stroke diseases’ physiology mechanism and diverse sensors’ functionalities is detailed explained. The review starts by interpreting how stroke influences motion abilities and then introduces broadly adopted physical training methods. After, the working principles of sensors and their use to objectively provide patients’ body information for stroke rehabilitation status are discussed. The content of the paper aims to not only review the state‐of‐the‐art works of developing sensors for assisting in evaluating stroke patients’ status but also bridging the gap between medical staff and engineers.


Introduction
Stroke, a type of acute cerebrovascular disease caused by the rupture or blockage of blood vessels in the brain, is now the second leading cause of death in the world. There are more than 12.2 million new stroke cases worldwide each year, at the same time, six and a half million people die from stroke annually. [1] 80% of stroke survivors have various degrees of motor dysfunction. [2] Studies have shown that 55-75% of stroke survivors still have movement disorders after 3-6 months, and most of them need regular daily care, seriously affecting their life qualities. [3] Effective stroke rehabilitation training can significantly restore patients' motion ability. Almost half of the patients (America, 49%;
Stroke rehabilitation training can be carried out during the whole after-stroke period: early rehabilitation in the emergency room or neurology department, routine rehabilitation in the rehabilitation ward or center, and active training in the community or at home. Stroke rehabilitation training plans are designed based on the doctors' subjective evaluations, which may lead to different training plans for the same patient, indicating that less experienced doctors may not be able to yield highly efficient training plans. The situation became worse when patients receive rehabilitation training in many developing countries where medical resources are relatively scarce. To address this issue, various sensor technologies, which give objective data in a convenient and flexible manner, start to play a significant role in evaluating the patients' rehabilitation status.
Diverse sensors have been utilized in assessing stroke patients' status. To report the state-of-the-art works to researchers in this field, numerical reviews have been generated. For example, Pablo summarizes the research on wearable sensors for improving upper limb movement ability and critically discusses their limitations and challenges. [5] Boukhennoufa evaluates recent advances in advanced wearable sensors and machine learning algorithms in the field of post-stroke rehabilitation and proposed a classification method that divides the evaluation system into activity recognition, motion classification, and clinical evaluation simulation. [6] The focus of Quiros's review is to evaluate the relationship between measurements based on wearable sensors and the level of function assessed by existing clinical stroke assessment methods. [7] Shi reviews the development and related research results of human-machine coordinated control in stroke rehabilitation, and discusses robotic rehabilitation, human-machine coupling system modeling and evaluation methods based on multimodal sensors. [8] Also in the area of robot-assisted therapy, focusing on the upper limb, Balasubramanian discusses how sensor-based robots quantitatively assess motor function in stroke patients. [9] Unlike previous reviews of site-specific exoskeletons, Lee provides a general overview of exoskeletons for the entire human body parts to illustrate the diverse strategies. Its content covers sensors, actuators, and motion prediction algorithms. [10] There are also some summaries shown in Table 1, which are not detailed here.
Based on our experience and discussions with colleagues in medical and engineering domains, we conclude that the gap Table 1. Reviews on the application of sensors in the field of stroke rehabilitation. Review Content Characteristic [5] Outlining wearable sensors for upper limb rehabilitation Presenting a roadmap for translating these technologies from "bench to bedside". [6] Assessing the progress made in the domain of stroke rehabilitation Making a status report of the different technological developments in smart upper and lower limb recovery [7] Providing an overview of setups to measure the quantitative and qualitative aspects of movements of stroke patients under free-living conditions using wearable sensors Evaluating the relation between the sensor-based outcomes that are obtained from moving in a free-living environment and the level of functioning as assessed by existing clinical evaluation methods. [8] Summarizing the development of human-robot coordination control and the associated research achievements Reviewing human-robot coordination control of lower-limb rehabilitation robots from four aspects, including demand analysis, system modelling, sensing and control strategies [9] Reviewing the current contributions of robot-assisted motor assessment of the upper limb Identifying the key problems to be solved in the measurement of robot upper limb motion [10] Offering a comprehensive summary of the recent advances of wearable exoskeletons and their constituting functional components 1) Reviewing the core components such as sensors, actuators and motion prediction algorithms that make up the soft robot exoskeleton 2) Giving a general overview of the wearable exoskeleton of the whole human body [11] Reviewing information technology that has helped stroke patients recover at home 1) The types of technologies available for home rehabilitation after stroke 2) The design requirements for such technologies. [12] Reviewing information on 10 external, non-invasive stroke sensor devices currently under development Providing a broad assessment of noninvasive exencephalic monitoring devices in an unbiased manner [13] Reviewing how conditions associated with the test of physical conditions can be determined, which can be detected, and which sensors are used for this purpose.
Verifying that the mean speed, the mean acceleration, the distance, and the different angles are the most extracted features for the identification of the results of the Functional Reach Test [14] Evaluating the current state of biofeedback literature pertaining to post-stroke gait training Determining future research directions related to gait biofeedback in context of evolving technologies [15] The application of VR in motor rehabilitation after stroke 1) The technologies used in VR rehabilitation, including sensors 2) The clinical application of and evidence for VR in stroke rehabilitation 3) Considerations for VR application in stroke rehabilitation. [16] Discussing the shortcomings of various sensors in compensation evaluation and detection Explore how technology-based methods were used to assess and detect compensation without the constant care of therapists [17] Aggregating both quantitative and qualitative knowledge from clinical research with wearable sensor technology in individuals with epilepsy, PD, and stroke Summarizing clinical application areas for all three diseases, key functions and clinical attributes measured by wearables, the proportion of reported missing data, compliance, and perceived experiences and preferences of wearables. [18] Reviewing patient demographics, the type of wearable technology used, gait parameter assessments, and measures of reliability and validity Identifying how wearable technologies have been used over the past decade to assess gait and mobility in persons with stroke [19] Presenting an overview on the state of the art regarding the use of the different Microsoft Kinect camera models to assess gait in post-stroke individuals Describing and summarizing the main features of the available works on gait in the post-stroke population, highlighting similarities and differences in the methodological approach and primary findings [20] Reviewing the technical and clinical specifications of biofeedback systems developed for post-stroke gait rehabilitation, including the complementary use with assistive devices and/or physiotherapist-oriented cues Comparing the sensors, actuators, and control strategies used to drive the sensory cues according to the sensor's output (technical specifications); and clinical protocols and evidence concerning the system's efficacy on post-stroke recovery (clinical specifications) [21] Reviewing the devices available for assessment of physical activity in stroke Comparing the effectiveness, feasibility and design of devices currently available for measuring physical activity [22] Reviewing position-sensing technologies and their application for human movement tracking and stroke rehabilitation.
Suggesting that it is feasible to build a home-based telerehabilitation system for sensing and tracking [23] Summarizing the state of the art in wearable movement sensors and their current applications to neurologic and orthopedic rehabilitation, followed by emerging clinical applications Concluding with an overview of next-generation sensor technologies that expand motion sensing through hybrid sensors, neural interfaces, and soft sensors [24] Inventorying and classifying interactive wearable systems for movement and posture monitoring during upper body rehabilitation, regarding the sensing technology, system measurements and feedback conditions between these two groups limits the development and utilization of sensors in evaluating disease progress. In many cases, doctors have difficulties explaining their needs in engineering language, and vice versa. Although increasing applications of sensors to assist in determining stroke patients' status have been found in literature, the issue explained above has not been successfully solved. To this end, this review article aims to build a bridge between doctors and engineers by explaining and linking the pathogenesis, clinical manifestations and rehabilitation exercises of stroke in medicine, to the development of desired sensing techniques according to the performances of patients and periods of disease, and thoroughly reviewing the methods of matching sensor data with current clinical scales. In this review, we organize the contents in a task-oriented structure. We first analyze the physiological basis of stroke rehabilitation, including the process of injury, principles of neuroplasticity, treatment, and assessment methods. For these methods, we introduce various sensors and their applications in the field of stroke rehabilitation, and the process that builds the mapping relationships between sensor data and clinical scales. By explaining the mechanism of the disease and the sensors used in different stroke periods, this article bridges the gap between doctors and engineers.

Stroke Rehabilitation Physiology Basis
The current understanding of stroke recovery mechanisms is not fully clear yet. Various factors contribute to this phenomenon, such as the location of the disease, the size of the lesion, the degree of dysfunction, and so on, leading to difficulty in the selection of proper treatment and evaluation. The traditional coping strategy is that therapists make targeted and empirical treatment according to the location of the patient's movement disorder and the score of the scale, accompanied by subjective problems. This can be effectively addressed by sensor-based rehabilitation de-vices, and their designers have an obligation to understand the pathophysiology of stroke in order to avoid clinical disconnection.
This section mainly explains the pathophysiology of stroke, from the process of nerve's and muscle's damage, to the principle that motor rehabilitation repairs neuromuscular, then to motor rehabilitation strategies and evaluation methods. It is helpful for researchers to understand the process of stroke causing motor dysfunction and motor rehabilitation repairing neuromuscular circuits, and develop more targeted treatment strategies. The content is outlined in Figure 1.

Process of Nerve Damage
There are two main types of stroke, hemorrhagic stroke, and ischemic stroke. The former includes intracerebral hemorrhage and subarachnoid hemorrhage, while the latter is the most common type, accounting for about 85% of all strokes, consisting of thrombotic ischemia and embolic ischemia. [25] Whether the cerebral blood vessels are blocked or ruptured, the result is a local cerebral blood flow disorder, causing ischemia and hypoxia in the brain tissue. Deprived of oxygen and energy, neurons immediately fail to maintain their normal transmembrane ion gradient and homeostasis. This triggers several processes that lead to cell death: excitotoxicity, oxidative stress, inflammation, and apoptosis. [26][27][28] The schematic diagram is shown in Figure 2.
When the cerebral blood flow is interrupted, the ATP concentration in the neuron starts to decrease rapidly, resulting in the inability of the ion pump on the cell membrane to maintain ion homeostasis, and the neuronal membrane depolarizes, causing excess and neurotoxic glutamate to be released from nerve endings. [29][30][31] After the occurrence of cerebral ischemia, multiple mechanisms generate free radicals and cause oxidative stress, in-www.advancedsciencenews.com www.advsensorres.com cluding acidosis, excitotoxicity, calcium overload, mitochondrial dysfunction, nitric oxide synthase activation, leukocyte activation, neurotransmitter autoxidation, and inflammation. [28,[32][33][34] Excessive free radicals can directly oxidize and damage DNA, proteins, lipids, and other biological macromolecules. [35,36] When neurons are damaged, they release cytokines that trigger an inflammatory response, which exacerbates secondary brain injury by aggravating blood-brain barrier damage, microvascular failure, brain edema, oxidative stress, and direct induction of neuronal cell death. [37,38] Both endogenous apoptotic factors and exogenous inducible ligands activate the apoptotic program, resulting in DNA fragmentation, degradation of cytoskeletal and nuclear proteins, cross-linking of proteins, formation of apoptotic bodies, expression of ligands for phagocytic cell receptors and eventually the apoptotic cells will be engulfed by phagocytes. [28,39,40]

Injury of Skeletal Muscle
Motor dysfunction caused by stroke is not the result of neurological damage alone. Changes in the structure, metabolism, and function of skeletal muscles that cooperate with the nervous system also play an important role, which are often neglected in the traditional direction of rehabilitation. [41,42] The most visible change in muscle after a stroke is muscle atrophy, as known stroke-induced sarcopenia. [41] This sarcopenia is caused by a combination of multiple mechanisms, including immobilized and dysfunctional atrophy, impaired feeding, inflammation, sympathetic hyperactivation, and denervation. [43] In addition to sarcopenia, there is a partial shift in the type of fibers in skeletal muscle after stroke: Stroke survivors show a shift from slow-twitch to fast-twitch muscle fibers. [44] This shift leads to impaired metabolisms, such as increased tissue lactate and glycerol production, delayed and impaired glucose utilization, and slightly increased energy expenditure, [45] due to the easier fatigue of fast-twitch muscle fibers. By measuring the respiratory exchange ratio and rates of fat and carbohydrate oxidation, different from healthy people who mostly oxidize fatty acids, stroke patients' muscles rely heavily on carbohydrates, whose utilization rate is up to 70%, limiting long-term exercise capacity. [45,46] At the same time, immobilization, muscle weakness, or paralysis with or without spasticity are key factors that lead to muscle contracture development, the resulting structural alterations, leading to tissue that is stiffer than normal muscle, and eventually restricting joint mobility, causing deformity. [47]

Motor Dysfunction
After a stroke, under the influence of nerve and muscle damage, patients develop various types of hyperkinetic and hypokinetic movement disorders. [48] Hemiballism, dystonia, tremor, myoclonus, and parkinsonism have been reported. [49] Hemiballism is the most common motor dysfunction after stroke, with an incidence of 40%, and is characterized by vigorous, irregular, poorly patterned, high-amplitude movements of the limbs on one side of the body and is generally regarded as a severe form of chorea. [50,51] As the second most common movement dysfunction, accounting for 20%, dystonia involves involuntary persistent muscle contractions that cause twisting and repetitive movements or abnormal postures. [50,52] Tremor manifests as a resting tremor of the limbs, which is significantly aggravated by movement, intention, and goal-directed movement, with typical irregularity, low frequency (4.5 Hz), and usually involving the upper extremities. [53] Myoclonus refers to a brief, involuntary twitch of a muscle or muscle group. [48] The main clinical manifestations of parkinsonism are bradykinesia and rigidity, usually accompanied by a disturbed gait. [54]

Pathophysiological Basis of Stroke Motor Rehabilitation
There are four distinct periods after a stroke: the hyperacute, acute, subacute, and chronic phases. [55] Within hours of ischemia, depolarization, glutamate release, oxidative stress, and eventually neuronal death occur in the brain tissue, as known the hyperacute phase of stroke. [55,56] This is followed by an acute phase dominated by apoptosis. The third period of stroke is the subacute phase, during which inflammation is reduced and plasticity is strongest, lasting three months in humans approximately. [57,58] As the stroke progresses to the chronic phase, the plasticity window diminishes and the potential for recovery becomes limited, generally after three months.
Neuroplasticity is the physiological basis of motor rehabilitation reestablishing neuromuscular motor circuits. Stroke results in differentiated gene expression in focal neurons, inducing axon growth, synapse formation, and dendritic spines in adjacent neural tissues. [59][60][61] And exercise promotes this plasticity of synapses. [62,63] Studies have shown that exercise induces upregulation of neurotrophic factors, such as BDNF, GAP-43, IGF-1, and so on, and they promote neurogenesis, angiogenesis, and synaptic plasticity. [64][65][66] Microstimulation and functional mapping studies have shown that recovery from stroke injury can lead to surviving brain regions assuming the functional role of damaged brain tissue. [67] Animal studies have shown that constraint-induced motor therapy recruits more neurons to the innervation network, enhances the number of synapses in the contralateral cerebral cortex, and increases contralateral glucose metabolism. [68,69] These above results suggest that exercise may facilitate neural compensation beyond infarcted tissue and functional reorganization of the brain. [63] Exercise rehabilitation also has many positive effects on skeletal muscle, like treating sarcopenia, rather than just targeting neuroplasticity. By activating mTORC1, reducing oxidative stress, inhibiting inflammation, inactivating UPS, promoting mitochondrial biosynthesis, increasing IGF-1/muscle somatostatin ratio, and enhancing insulin sensitivity, exercise training can produce beneficial effects on skeletal muscle. [70,71]

Treatment
In clinical practice, the treatment for stroke patients includes drug therapy, motor therapy, occupational therapy, physical therapy, speech therapy, psychological therapy, rehabilitation nursing, and traditional Chinese medicine treatment (acupuncture, traditional Chinese medicine). [72] Among them, autonomic movement rehabilitation training, such as motor therapy and occupational therapy, are the most effective methods to reduce motor impairments, restore the affected neuromuscular function, rebuild sensorimotor neural circuits and achieve independent body control in stroke patients. [73][74][75] At present, there are a variety of motor rehabilitation treatment methods, and each has a specific theoretical basis and scope of application, but there is no unified classification method for them. This paper divides them into basic theory strategies, empirical approaches, and data-based approaches.
In the 1960s, as traditional rehabilitation strategies, various neurodevelopmental techniques were proposed and improved successively. These are the methods to apply the basic rules and principles of neurophysiology and neurodevelopment to the rehabilitation treatment of movement disorders after brain injury, such as Bobath, Brunnstrom, proprioception neuromuscular fa-cilitation, and Rood, [76][77][78][79] which have been used as basic theories to the date.
Since the 1980s, many new treatments have been proposed that are effective, convenient, targeted, and personalized, such as motor relearning, mirror therapy, constraint-induced movement therapy, motor imagery, bilateral training, music therapy, and body weight support treadmill training. [80][81][82][83][84][85][86] These empirical strategies are the basis of stroke rehabilitation, which have been proven to be effective in clinical practice. Specific training items include the affected limb motor function enhancement training, balance control training, gait training, and upper limb operation training. [72] However, these approaches are hampered by limitations that the rehabilitation effect may be related to the therapist's experience.
In modern times, with the continuous development of sensor technology and artificial intelligence, and the continuous improvement of datamation requirements in the field of rehabilitation, on the basis of conventional rehabilitation treatment, intelligent rehabilitation technology continues to develop, with the advantages of informatization, standardization, and intelligence. [73] As non-invasive brain stimulation techniques, repetitive transcranial magnetic stimulation (TMS) and transcranial direct current stimulation (tDCS) use magnetic field pulses and constant low-intensity direct current to regulate the activity of cerebral cortical neurons, respectively, while adding traditional exercise rehabilitation training to the injured limbs. [87,88] The combination of magnetic or electrical stimulation techniques with rehabilitation tasks has produced encouraging results. Inspired by somatosensory games, virtual reality (VR) technology helps stroke patients to interact traditional exercise rehabilitation tasks with virtual environments through visual, auditory, and tactile senses. Compared with traditional rehabilitation technology, VR has great advantages in personalization, safety, fun, and timeliness, and experiments show that the combination of VR and traditional rehabilitation technology can induce the excitability of the brain motor cortex. [89] To reduce the cost of rehabilitation learning and reliance on therapists, robot-assisted therapy has been developed, with the most important advantage of providing repetitive, intensive, and accurate training for patients. [90,91] Brain-computer interfaces (BCI) is a system that measures brain activity, decodes motor thought information, and then converts it into an artificial output to control external devices. [92] The combination of BCI and motor rehabilitation can not only restore the limb control ability of patients with movement disorders, especially those with severe injuries who do not have the minimum motor ability required for traditional rehabilitation, but also monitor the neural activity during the rehabilitation process in real time. [93] Electromyographic (EMG) biofeedback technology can record the weak electrical signals of the patient's muscle contraction, and feedback to the patient or robot-assisted devices to adjust the muscle activity autonomously or passively, thus improving the motor function. Studies have shown that treatment with EMG feedback is more effective than passive training. [94] Combined with various wearable sensors, relying on Internet communication technology, some patients can achieve remote rehabilitation at home, which is not significantly different from the traditional method in the hospital, with a lower cost. [95,96] Although various intelligent rehabilitation technologies are developing rapidly, it should be emphasized that they are an aux-www.advancedsciencenews.com www.advsensorres.com iliary means of traditional rehabilitation therapy rather than a substitute. [73] Each rehabilitation strategy has its own advantages and disadvantages, therefore, different conditions may require a flexible combination of different approaches.

Evaluation
In the process of stroke rehabilitation, patients' motor function recovery needs to be accurately evaluated, because of its influence on the improvement of the patient treatment plan and the redistribution of medical resources. In clinical use, there are a variety of well-established and subjective scales for assessing motor function and the ability of daily living in stroke patients, including Brunnstrom Recovery Stages, [77] Fugl-Meyer Assessment, [97] Berg Balance Scale, [98] Postural Assessment Scale for Stroke Patients, [99] Falls Efficacy Scale, [100] Barthel ADL (Activities of Daily Living) Index, [101] the Functional Assessment of Stroke [102] and so on. The accuracy and reliability of these scales for rehabilitation assessment strongly depend on the skill and experience of the therapist. With the development of sensor technology, objective assessment methods have gradually developed, such as biomechanical testing, kinematic testing, electrical testing, machine vision, etc., and these methods make full use of the characteristics of mechanical sensors, inertial sensors, EMG sensors, vision sensors, etc., as a favorable supplement and improvement of subjective assessment scales.
Sensors have great potential in the field of stroke rehabilitation. The following content will focus on the application of various sensors in stroke rehabilitation in detail.

Single Mode Sensing Methods
The traditional treatment and evaluation methods of stroke are subjective to a certain extent, and the accuracy and reliability largely depend on the skills and experience of therapists. Therefore, it is necessary to introduce sensors that can support precise quantification of rehabilitation training dose, detailed assessment of injury and mobility impairment, [103] into the field of stroke rehabilitation. Medical workers use these sensing devices to accurately quantify various rehabilitation information of patients, such as muscle strength, posture information, electromyography information, joint angle, movement speed, and so on, and then give more appropriate rehabilitation strategies.
This section lists the clinical stroke rehabilitation methods and analyzes the information dimensions that sensors can perceive in these methods. According to the information on stroke rehabilitation, the basic principles, applications, and limitations of several kinds of sensors are introduced. The schematic diagram of this section is shown in Figure 3.

Information Dimension
Although there are a variety of rehabilitation training and assessment methods for motor disorders in different parts due to the uncertainty of the focal point of stroke, the information dimensions abstracted from them are similar.
Upper limbs: rotation control of shoulder joint and wrist joint, flexion and extension control of elbow joint, and hand coordination and dexterity training.
Lower limbs: rotation control, extension and flexion activities of ankle joint and hip joint, and gait training including one leg support, stepping, walking, and going up and down stairs.
Trunk: flexion, extension, and rotation training. Balance control: the training of static and dynamic balance of sitting and standing.
The vast majority of exercise rehabilitation processes require limb or trunk movement, whose kinematic information needs to be mastered, including position, velocity, acceleration, angular velocity, and range of motion. The basic physiology of exercise is that the central nervous system controls the muscles through electrical signals. Gait, described as the sixth vital sign, is a predictor of independence, mortality, family and community functional status, and quality of life, [104] intuitively reflected by plantar pressure distribution. In addition to gait monitoring, mechanical interactions between the human body and the outside world exist in grasping, bracing, sitting, and other training or daily activities. When focusing on the motion of a joint, such as the flexion and extension of the elbow or knee joint, we can measure the deformation of the skin around the joint. The above four kinds of information, kinematics, bioelectricity, force, and deformation, can basically cover the vast majority of stroke training and assessment content; the information can be measured by inertial measurement unit (IMU), bioelectricity sensors, force sensors, and optical sensors. The sensitive elements, physical quantities of perception, basic principles, functions, usages, and limitations of the above and other types of sensors applied in stroke rehabilitation are presented in Table 2.

IMU Sensor
IMU is an integrated accelerometer and gyroscope device that can measure the acceleration and angular velocity of an object in three axes. Due to its small size, ease to wear, and low cost, IMU can assess numerous information in stroke rehabilitation, such as hand function, [106] coordinated arm movement, [120] trunk injury, [105] gait, [121] fall risk, [107] posture of sitting or standing [122] and so on, through its wear on fingers, wrists, arms, chest, waist, hips, legs, and ankles. Detailed explanations are given in the following.
Jovana puts the IMU on the tip of the damaged hand (Figure 4a) for an objective assessment of recipient function, unlike the sports glove model limited by size and joint alignment issues. [106] Biswas recognizes three basic arm movements (reach and retrieve, lift the cup to mouth, rotation of the arm), achieving accuracies in the range of 91-99% for healthy subjects and 70-85% for stroke patients, through a three-axis accelerometer worn on the wrist as a means of assessing upper limb rehabilitation in stroke patients. [120] Williams explores the Functional Reach Test (FRT) for real-time fall risk assessment, and implements the FRT function in mStroke, using a chest-worn NODE sensor (the basic module is essentially an IMU, as shown in Figure 4b). [107] The fusion of multiple IMUs worn on different body parts can feedback  [105] Copyright 2020, The Authors, published by MDPI. Hand function assessment by IMU. Reproduced unter terms of the CC-BY license. [106] Copyright 2021, The Authors, published by IEEE. Assessment of fall risk by accelerometer. Reproduced unter terms of the CC-BY license. [107] Copyright 2017, The Authors, published by HINDAWI. Gait assessment by IMU. Reproduced unter terms of the CC-BY license. [108] Copyright 2020, The Authors, published by MDPI. An EMG-driven wrist/hand exoneuromusculoskeleton. Reproduced unter terms of the CC-BY license. [109] Copyright 2021, The Authors, published by BMC. An EMG-based gait monitoring system.Reproduced under terms of the CC-BY license. [110] Copyright 2021, The Authors, published by MDPI. Trunk Muscle activity detection based on EMG. Reproduced under terms of the CC-BY license. [111] Copyright 2022, The Authors, published by PLOS ONE. Elbow torque measurement. Reproduced unter terms of the CC-BY license. [112] Copyright 2019, The Authors, published by FRONTIERS MEDIA SA. EEG-based brain motion imaging. Reproduced under terms of the CC-BY license. [113] Copyright 2021, The Authors, published by IEEE. Kinematic analysis of the upper limbs based on Kinect. Reproduced under terms of the CC-BY license. [114] Copyright 2022, The Authors, published by MDPI. Trunk compensation detection based on the camera. Reproduced under terms of the CC-BY license. [115] Copyright 2018, The Authors, Published by IEEE. Trunk compensatory movements detection based on FBG (fiber Bragg grating sensor). Reproduced under terms of the CC-BY license. [116] Copyright 2022, The Authors, published by MDPI. Force myography (FMG) for monitoring grasping. Reproduced under terms of the CC-BY license. [117] Copyright 2017, The Authors, published by FRON-TIERS MEDIA SA. Compensatory detection robot based on the force sensor. Reproduced under terms of the CC-BY license. [118] Copyright 2020, The Authors, published by IEEE. A Finger Grip Force Sensor. Reproduced under terms of the CC-BY license. [119] Copyright 2019, The Authors, published by MDPI. on more complex movement information. Alhwoaimel proposed an instrumented Trunk Impairment Scale (iTIS) as an objective measure of trunk injury after a stroke. The iTIS, using the Valedo system with three IMUs placed on the front chest, back thoracic vertebrae, and lumbar vertebrae, respectively (Figure 4c), can detect small changes in abnormal trunk movement that cannot be observed clinically, confirming a moderate relationship between it and clinical scores. [105] With five IMUs wearing on two wrists, two arms, and hip, respectively, combined with machine learning algorithms, the study of Chen can monitor task-specific activities of daily living (ADLs) and model them, improving the recognition accuracy to 90% and providing treatment references for clinicians and researchers. [123] Veerubhotla's study compares the accuracy of three machine learning algorithms in identifying and distinguishing sitting, standing, and walking in stroke patients by wearing IMUs on their wrists, waist, and ankles, finding that RUSBoosted trees have the highest accuracy, 99.1%. [122] The vast majority of studies have focused on assessing motor function, especially motor quality, through standardized clinical tests, which may explain the greater prevalence of IMU use to date. [5]

EMG Sensor
Electromyography (EMG) is a noninvasive and wearable technique that measures the response of a muscle to the electrical stimulation of a nerve by collecting the electrical potential of the muscle as it contracts through electrodes. [5,124] To a certain ex-tent, the specific changes in EMG signal amplitude, frequency, and other parameters can reflect the ability of central nervous control and muscle excitation conduction closely related to the control dysfunction of stroke patients. [125] In the process of rehabilitation therapy and evaluation, EMG can assist therapists or devices to judge the status of patients'  . IMUs worn on fingers (Reproduced under terms of the CC-BY license. [106] Copyright 2021, The Authors, published by IEEE), chest (Reproduced under terms of the CC-BY license. [107] Copyright 2017, The Authors, published by HINDAWI), and back (Reproduced under terms of the CC-BY license. [105] Copyright 2020, The Authors, pubished by MDPI).

Figure 5.
EMGs worn on the a) arm (Reproduced under terms of the CC-BY license. [109] Copyright 2021, The Authors, published by EMC), b) stomach (Reproduced under terms of the CC-BY license. [111] Copyright 2022, The Authors, published by PUBLIC LIBRARY SCIENCE), and c) legs (Reproduced under terms of the CC-BY license. [110] Copyright 2021, The Authors, published by MDPI). muscles, such as the arm muscles [109] for upper limb movement, the trunk muscles [111] for stabilizing and moving the seat, and the leg muscles [110] for gait, as shown in Figure 5. In emerging therapies, EMG can monitor the neuromuscular potentials triggered by TMS during treatment, allowing for more accurate doses. [126] In occupational therapy, patients tend to use compensatory strategies to accomplish tasks, and these compensations are sometimes difficult for therapists to detect and correct, but can be captured and quantified by EMG. [127] Control scheme based on EMG has proven their superiority in man-machine cooperation, as EMG signals provide a good estimation of movement intention. [128] Combined with pattern recognition technology, Zhang has developed an EMG control system to improve the rehabilitation of stroke patients. The detection accuracy of upper limb movement type reaches 96.1%. [129] EMG can monitor muscle activity even when there are no signs of muscle activity on traditional examinations. The study of Papazian evaluates the arm muscle activity of stroke survivors with limited or no arm movement during acute care, providing a reference for EMG to predict future function. [130]

Force Sensor
The force sensors can detect tension, pressure, weight, torque, internal stress and strain, and other mechanical quantities. Here mainly introduces the most widely used normal and tangential force measurement devices, namely pressure sensors and strain sensors.
The pressure sensor can feel the pressure (normal force) signal and convert the pressure signal into usable output electrical signal according to a certain rule. The signal could be the interaction force of the human-operated facility or the ground reaction force [133] that feeds back gait information. Khoo embeds Figure 6. a) Measuring the force applied to the handrail using force sensors on the force plate to assess the sit-to-stand motion. [131] Copyright 2021, IEEE. b) Assessing ADL based on flexible strain sensor embedded in upper limb garment. Reproduced under terms of the CC-BY license. [118] Copyright 2020, The Authors, published by IEEE c) Evaluating the trunk compensation strategy through pressure sensors on the rehabilitation robot. [132] Copyright 2005, EMC. d) Monitoring grasping based on FMG. Reproduced under terms of the CC-BY license. [117] Copyright 2017, The Authors, published by FRONTIERS MEDIA SA. six pressure sensors in the insole to calculate the gait parameters of stroke patients, including gait time and asymmetry, while incorporating the idea of biofeedback to provide corresponding real-time biofeedback through auditory and electrotactile stimuli to actively correct the patient's gait. [134] The recovery of skeletal muscle strength is the most intuitive feedback of the degree of stroke recovery, and this recovery process can be captured by force sensors, especially pressure sensors. Park integrates pressure sensors into the grasping rehabilitation robot; the sensors detect the patient's grasp intention and then drive the robotic device to move the patient's hand to form a normal grasping behavior. [135] Using gloves integrated with pressure sensors, Hossain proposes a hand rehabilitation service framework, based on the technology of cloud computing and augmented reality, to remotely detect patients' hand rehabilitation training. [136] Combined with pressure distribution data on the handrail of the rehabilitation facility, the pressure sensor also provides feed-back on the patient's sit-to-stand motion (Figure 6a), [131] trunk compensation strategy (Figure 6b), [118] and balance ability. [137] As a prerequisite for rehabilitation and training, adequate sleep for stroke patients can be detected by pressure sensors placed under the mattress. [108] In recent years, a technique called force myography (FMG), essentially consisting of pressure sensors to record changes in the stiffness patterns of muscles around the limb during different movements, has gained traction among researchers. [138] Sadarangani uses six pressure sensors embedded linearly and isometric in a circular flexible band (Figure 6d)that can detect the precise movements of the upper limbs of stroke patients who wear the band on the forearm, and this FMG-based grasp detection shows a high accuracy of 92.2%. [117] As an alternative to EMG, FMG can use off-the-shelf pressure sensors, without complex signal processing circuits and placing the sensor in specific anatomical points of the body. [139,140] [115] Copyright 2018, The Authors, published by IEEE) and b) marked visual measures (Reproduced under terms of the CC-BY license. [114] Copyright 2022, The Authors, published by MDPI). c) Trunk compensation detection based on FBG (Reproduced under terms of the CC-BY license. [116] Copyright 2022, The Authors, published by MDPI).
The strain sensor is a kind of sensor based on measuring the strain caused by the tangential force deformation of the object. Strain sensors applied in stroke rehabilitation have the characteristics of high sensitivity, low hysteresis, good linearity, and low cost. Particularly, some of them have good flexibility and mechanical properties similar to skin, making strain sensors have great advantages in rehabilitation fields such as joint bending angle measurement and motion tracking. [141,142] Denen embeds a capacitive strain sensor into the insole to assess gait by measuring information about plantar deformation. The sensor is made of a composite elastic polymer combined with single-walled carbon nanotubes, which has a high dielectric constant and a low hardness. [143] Yao attaches small capacitive strain sensors based on silver nanowires to various joints of the hands to measure skin deformation recordings that are highly correlated (>93%) with joint angles, to assess the hand mobility of stroke patients. [141] The strain sensors, made of elastic metal in a linear or ribbon-like shape, are also used to measure hand function, upper limb daily mobility, and swallowing function when inserted into gloves, [144] embedded in upper limb garment (Figure 6c), [132] and attached to the neck, [145] respectively.

Optical Sensor
The most widely used optical sensor is the visual imaging device that is generally composed of a photosensitive element (such as CCD or CMOS) and an imaging device (usually an imaging lens).
The way it works is to evaluate the rehabilitation process through optical imaging that directly captures the limb information of the patient's movement or the position information of the patient-driven device. Rehabilitation training, based on visual sensing, the most convenient and concise sensing mode, can make use of equipment commonly found at home, such as Kinect. Based on Kinect, Hoang designs a rehabilitation training system that directly photographs stroke patients doing Tai chi exercises, tracking their skeletal information and comparing it with that of normal people, completing the construction of 18 recognition poses. [146] Also using Kinect, Zhi photographs the upper limb task of the patient to quantify the compensation (Figure 7a), combining with support vector machine classifier and recurrent neural network classifier, reaching a high AUC (area under curves) score of 0.98 in lean-forward compensation recognition. [115] The marker-free vision-based approach described above is convenient, economical, adaptable, and simpler than other approaches. [147,148] Markers can be added to the limb for a wider variety of applications. Fan places three markers on the hand, elbow, and shoulder, respectively, using infrared cameras to image the markers to assess the quality of rehabilitation training, with the support of audiovisual feedback and personal goal setting. [149] Achieving the quantification of the trunk, shoulder, and elbow compensation (Figure 7b), Faity also points out that Kinect does not quantify the number of velocity peaks and the peak hand velocity. [114] In addition to geometric optics, there is a class of optical sensors that apply physical optical properties, FBG. Also sensitive to deformation, different from the strain sensor based on electricity, the principle of FBG is that when the fiber is deformed, the wavelength of the outgoing light is different from the incident light, and the rate of the change is closely related to the degree of deformation. Due to its small size, light weight, multiplexing capability, biocompatibility, and immunity to electromagnetic interference, FBG has been applied in stroke rehabilitation in recent years, such as hand function assessment [150] and trunk compensation detection (Figure 7c). [116]

Others
In bioelectricity, in addition to skeletal muscle, the electrical activity of the brain and heart, measured by electrodes attached to the head and chest, respectively, can be utilized in the field of stroke rehabilitation. The electroencephalogram (EEG) waveforms measured by EEG sensors, such as Delta, Theta, Alpha, or Beta waves, can reflect the patient's motor, cognitive, and emotional states and their complex relationships, can be used as the input of the BCI technology described above to drive an external device such as a robot or exoskeleton. [93,151,152] Some studies have shown that ECG sensors can detect atrial fibrillation, one of the key causes of stroke, preventing the occurrence and recurrence of stroke disease. [153,154] Similar to the visual sensing model, Griffith uses ultrasonic echolocation sensors, changing the perceived content from light to sound, to detect compensatory movements associated with excessive trunk flexion, [155] which may solve the fixed defects of visual sensors, and privacy disclosure. Besides the above, some studies have also added angle sensors, [156] force torque sensors, [157] optical encoders, [158] radio frequency identification sensors, [159] etc., to the training and evaluation of stroke rehabilitation that are not listed in detail here.
The sensor can give some relatively intuitive feedback on the patient's physical state indicators. The IMU acceleration and angular velocity information can be used to calculate the ROM of the limb. Parameters such as peak value or mean value of EMG or mechanical sensor can reflect the intensity of muscle activity according to certain rules. Optical sensors generally capture markers to determine the position of the limb, vector and other information, and then evaluate the patient's mobility. When deep learning is introduced into the field, more dimensions of these indicators are generally required. This sometimes requires additional calculation of the time-domain or frequency-domain characteristics of the data, or classification of the original data by frequency, window and other parameters to increase the dimension. Study characteristics related to the single-mode sensor applica-tions in stroke rehabilitation and their placement, evaluation content, performance or result, algorithm, feature and indicator for the included papers are presented in Table 3.

Limitation of Single Mode Sensing
A training or evaluation often involves multiple information dimensions. Limited by the inherent defects of the sensing mode, the use of a single type of sensor may lead to incompleteness and inaccurateness of measurement results. The limitations of each type of sensor are detailed below.
IMU sensors have inherent disadvantages that the displacement and angle information is calculated by integrating the inertial signal, and it's not sensitive to slow movement. The former leads to the accumulation of errors over time; the latter may make the acquired inertial signal unable to accommodate enough information because of the tendency of stroke patients to move slowly. [5] Moreover, IMU's position changes during patients' exercise, reducing the validity of the data. [160] Medically, there is no clear clinical interpretation of the motor quality indicators obtained from devices based on IMU. [161] As physiological limitations, interference between muscles, [162] reliability of deep muscles, and the influence of adipose tissue, [163] sweat, [164] and hair [117] can affect EMG recording. Needle electrodes are invasive and surface electrodes are poorly targeted to specific muscles; [165] the electrodes need the hardware of complex signal acquisition and method, as well as a high sampling rate. [117] These limitations result in low SNR and high variability of EMG, so models based on EMG alone cannot achieve accurate identification. [166] The main disadvantage of force sensors (mainly resistive) is that temperature changes can affect the output, in turn leading to thermal zero offset and sensitivity drift. [167] Some pressure sensors are manufactured and packaged at high temperatures, so there are thermal residual stresses that can affect the measurement results. [168] Creep effects are common in strain sensors, causing deviations that reduce long-term stability. [169] Detection devices based on visual sensors may lead to privacy leaks, causing stroke patients to engage in unnatural behavior as a result of monitored discomfort. [155] It is also limited by the lighting environment [105] and the shooting field of view. [170]

Multimodal Sensing Method
The information obtained by a single type of sensor is very limited, and it is also affected by device performance and other objective factors, so there are often multiple types of sensors applied in stroke rehabilitation training and evaluation in recent years. Focusing on one specific action or feature in the process of stroke rehabilitation, the multimodal sensing method complements and optimizes the combination of information of various sensors in multi-level and multi-space, and finally generates a consistent interpretation of the patient's state. Currently, there is no uniform classification of the application of sensors in the field of stroke rehabilitation. It is classified according to the three periods of rehabilitation, flaccid period, spasm period, and recovery period. This classification method is not only because of the different application time points, but more importantly, the roles of the sensor are  Figure 8. Applications of the multimodal sensor in three rehabilitation periods of stroke. From the picture directly above, in clockwise order: Evaluating the spasm level of the elbow joint by EMG, angle sensor, and torque sensor. [177] Copyright 2021, IEEE. Detecting hand spasms by IMU and pressure sensors. [178] Copyright 2021, WILEY. Knee pendulum test to assess knee spasm. [179] Copyright 2018, BMC. Comparing muscle synergy by EMG and force plate. [180] Copyright 2019, IEEE. Detecting bilateral coordination by the binocular camera and pressure sensor. [181] Copyright 2016, MDPI. Gait assessment by FSR and EMG. [182] Copyright 2021, IEEE. VR technology applied in upper limb rehabilitation. [183] Copyright 2022, NATURE PORTFO-LIO. Detecting trunk compensatory by the pressure sensor and EMG. [184] Copyright 2019, BMC. Recognizing lower limb movement tasks by IMU and EMG. [185] Copyright 2021, MDPI. Measuring the exoskeleton angular velocity and interaction force by angle sensors and torque sensors. [175] Copyright 2021, HINDAWI LTD. Monitoring torque and angle information of the ankle exoskeleton by torque and angle information. [174] Copyright 2021, IEEE.
relatively clear, assisting passive training, detecting spasms, and evaluating active training, respectively. The schematic diagram of this section is shown in Figure 8.

Flaccid Period
The main manifestations of the relaxation period are muscle relaxation, low muscle tension, and no voluntary movement. Passive activity, as the most suitable rehabilitation mode for a relaxation period, can maintain the range of motion of the joint, prevent muscle contracture, stimulate muscle recovery, and induce the body that has lost function to regain function. Passive training is mainly performed by the healthy limb, the therapist, and the robot where the sensor is applied centrally, and there is evidence that robot-assisted therapy is significantly more effective than traditional methods. [172] In general, the position, angle, and force information of the rehabilitation assistance robot need to be perceived. The Rehabilitation Institute of Chicago designs an ankle exoskeleton system that monitors torque and angle information in real time through force and angle sensors. In passive stretching mode, the ankle joint can be forcefully and safely extended to extreme dorsiflexion. [173] They then display the torsion information of the joint on a monitor for patients and therapists to manipulate (Figure 9). [174] The upper limb assisted training robot of Huashan Hospital also uses angle sensors and force torque sensors to measure the exoskeleton angular velocity and human-computer interaction force, respectively, so as to identify the motion intention of the patient. [175] Wang places IMU and angle sensors at the joints of the lower limb exoskeleton to guide gait synchronization between the passive leg and the healthy leg. [176] [174] Copyright 2021, IEEE.
Unlike the above assistive robots following predefined trajectories, passive assistive robots driven by biological signals of human-computer interaction (EMG or EEG) give the patient more subjectivity, facilitating active closed-loop stimulus feedback for motor function reconstruction. Chen proposes a control strategy for a lower limb exoskeleton based on EMG signals of leg muscles, with plantar pressure sensors detecting movement quality and real-time feedback. [186] The upper limb exoskeleton robot controlled by EMG generally needs to ensure the real-time and accuracy of angle and torque information, and meanwhile due to the complexity of upper limb movement, angle sensors, and torque sensors are required. [94,187] In addition to its similarity to EMG, driving assistive robots, BCI technology based on EEG signals can also allow targeted control of functional electrical stimuli (FES) as sensory feedback for motor intention synchronization. [188] With bioelectricity alone, using EEG as signals to drive the exoskeleton, in combination with the angle information calculated from EMG, Kawase achieves precise control of the passive movements of fingers, wrists, and elbows (Correlation coefficients: 0.91 ± 0.01). [189] Passive training transmits proprioceptive information not only to the motor cortex, but also to the sensory cortex, thereby activating the complete sensorimotor system. [190][191][192] Therefore, it is effective to give the patient with barely moving a normal-like and real-time sensation in passive training, such as giving electrical stimulation to corresponding muscles, or pressure irritation to connected skin, while the exoskeleton controls joint movement. Passive ankle rehabilitation is typically repetitive dorsiflexion and extension exercises driven by robots in the clinic. To mimic normal walking, with the calf muscles electrically stimulated and the planta squeezed suitably, this kind of training based on sensory feedback should be more effective. The so-called "suitable" location, intensity, time, and other information should be collected from normal people in advance by IMU, EMG, angle sensors, and so on.

Spasm Period
Spasticity is a common complication in patients after stroke, manifesting as an abnormal muscle stiffness that leads to dyskinesia and dysfunction, causing great discomfort and pain, and affecting the quality of life of stroke patients. [193][194][195] So, assessment of the extent of spasticity is essential for early intervention during rehabilitation. The Ashworth Scale, Modified Ashworth Scale (MAS), and Tardieu Scale are widely used in the clinical measurement of spasticity, however, as mentioned above, are limited by the subjectivity of the therapist. [196,197] There is increasing interest in the application of sensors in spasticity measurement due to their quantitative and reliable properties.
There is a high correlation coefficient between spasticity levels and EMG parameters such as the root mean square (RMS) of the signal, [198] therefore, most spasm measurement systems include EMG. Due to the inherent limitations of EMG mentioned above, spasm measurement systems generally include other types of sensors. Yeh assesses the degree of spasticity in the pendulum experiment (Figure 10b) by attaching angle sensors to the knee joint in conjunction with muscle activity measured by EMG, showing effective discrimination ability between spastic and nonaffected limbs using the method (P<10 -4 ). [199] The parameter obtained in the pendulum experiment is the angle time curve of the joint falling freely under the action of gravity. Due to spasms, the damping effect of abnormal muscle contraction will affect the angle time curve that can be sensed by angle sensors. For the same purpose of detecting joint angular motion, Hu places an angle sensor at the elbow joint of the upper limb exoskeleton, combines it with EMG information, and uses HMSEN (Hilbert-Huang transform marginal spectrum entropy) method to improve the recognition rate of upper limb spasm to 95.45%. [198] The torque sensor has the ability to measure the mechanical impedance of the joint more intuitively, therefore, Wang integrates the advantages of EMG, angle sensor, and torque sensor simultaneously to Figure 10. a) Components and application of upper limb spasm assessment system. [177] Copyright 2021, IEEE. b) Schematic of the knee pendulum experiment. c) Quantifying the maximum voluntary elbow contraction task based on IMU, EMG, and force sensor. [200] Copyright 2022, IEEE.
evaluate the spasm level of the elbow joint (Figure 10a), with the combination of genetic algorithm, observing strong correlations between the proposed spasticity assessment and the severity level measured by clinical scales (R = 0.86, P = 1.67*10 -5 ). [177] Spasms can cause errors in the patient's movement and force control, which can be used as a criterion to identify the type of spasm. [201,202] Using three supervised machine learning algorithms, Zhang evaluates the neurogenic and kinematic components of muscle spasms during passive movement in patients, which are sensed by EMG and IMU sensors affixed to the upper arm and wrist, and the final evaluation score could be used as a prediction of the scale. [203] The final conclusion is that the method using SVR to fuse the biomarkers of the two models is superior to the other two methods, with the lowest MSE of 0.059. Although using the same type of sensor, unlike the study of Zhang, Chen explores the method of evaluating upper limb spasm under active movement, and concludes that a random forest with a window length of 256 ms gives the best results. [204] In addition to IMU and EMG, Wang adds an additional pressure sensor placed between the hand and the table, quantifying the maximum voluntary elbow contraction task, when the patient lifts the table with one hand. [200] The system's assessment of muscle activation is 96% accurate, well above the 71% achieved with EMG alone.
The linear flexible strain sensor can measure the change of radial spasm force, and it is easy to embed into gloves and finger exoskeletons, so some studies use it to combine with IMU to detect hand spasms. [178,205] Visual measurement has great advantages in the measurement and modeling of whole-body motion. Using labeled visual sensors and IMU, Li proposes a dynamic model based on a bone vector to reconstruct the body posture of hemiplegic children during exercise training to evaluate the degree of spasticity, and verify the accuracy in the measurement direction very close to that of OptiTrack. [206] The traditional indicators for the evaluation of spasticity include the RMS of electromyography, [198] the angular time curve of pendulum test [199] and the threshold of tetanic stretch reflex based on IMU and EMG. [204] On this basis, some new indicators are proposed, such as the novel indicator based on PAC between angular velocity and sEMG amplitude, [199] and identifying the disparity between the impaired and unaffected side in the MVC test. [200] Zhang proposes a brand new evaluation indicator, a five-dimensional vector that simultaneously integrates TSRT from the lambda model and four biomarkers from the kinematic model, realizing the complementation of the two. [203]

Recovery Period
During the recovery period, patients ought to positively participate in active training to improve their self-care ability and shorten the process of returning to family and society. Only active movement indicates that a particular motor function has formed a circuit in the central nervous system.
Due to the diversity of active training types, sensors are used more in active training, especially the advantage of multi-sensing technology is more obvious. For example, in gesture recognition, the study of Song has shown that the classification accuracy of the combination of FMG-EMG-IMU (81%) is significantly higher than any single sensing mode (EMG, 69.6%, FMG, 63.2%, IMU, 47.8%). [207] The detection of motion intention by the sensor is helpful for human-computer interaction. Park introduced elastic tension sensors and pressure sensors into the hand orthotics to assist EMG in identifying the motion intention of patients, thus achieving accurate assistance in grasping training (global accuracy = 86%). [208] Three types of rehabilitation performance in convalescent patients need continuous attention: inhibition of abnormal posture, enhancement of balance ability, and improvement of coordination ability.
One type of injury after a stroke is the loss of independent movement of a joint or body part. In the early stages of active rehabilitation, the movement of a patient's single joint often leads to excessive movement of adjacent joints, [209] such as "basket on upper limb, circle on lower limb," therefore, when the affected limb can be actively lifted, training should focus on the detection and correction of abnormal posture. By analyzing the principle of compensated motion, Xu designs a motion pattern detection system composed of force sensors, angle sensors, and EMG, and identifies and classifies three abnor- Figure 11. a) Trunk compensatory measuring device based on EMG and pressure distribution mattress and operation demonstration. [184] Copyright 2019, BMC. b) Motion prediction based on EMG and IMU and experimental process. [185] Copyright 2021, MDPI. mal motion modes, namely trunk rotation, trunk forward tilt, and scapular elevation, with recognition rates of 91.56%, 91.90%, 82.62%, respectively. [210] Based on the combination of pressure distribution data of occupational therapy with EMG data of eight trunk muscles, Cai achieves excellent performance in distinguishing abnormal movements, with an average classification accuracy of 98.1% for various trunk compensatory movements (Figure 11a). [184] All these studies show that support vector machine is effective in abnormal motion pattern recognition. [118,184,210] Patients have been trying to complete active training tasks with greater accuracy, but in fact, their movements do not match the target task in a way that even professional therapists have trouble distinguishing it. The quality and anomalies of these tasks can be evaluated and quantified by sensors. Giuffrida uses the most classical sensor combinations, IMU and EMG, to recognize upper limb rehabilitation tasks with an accuracy of more than 80%. [211] Meng applies this combination to the recognition of lower limb movement tasks (Figure 11b), with the comparison of four classifiers, and finally concludes that the SVM model with a sliding window size of 80 ms has the best recognition performance. [185] Balance is the ability to maintain body orientation in space under static and dynamic conditions, as postural stability at rest and as a response to activation or external disturbance, respectively. [212] At present, IMU, pressure sensors, and EMG are the main sensors for balance capability evaluation. The IMU provides accurate information about body spatial orientation and movement, the EMG assesses specific patterns of muscle activation in response to static and dynamic postural perturbations, and pressure sensors are placed with a fixed insole or integrated into the shoe to measure changes in plantar forces. [213] Gait dysfunction is closely related to balance dysfunction. [214] Zhao proposes an intelligent sensing system for gait pattern recognition; the system (Figure 12) utilizes 48 plantar stress distribution sensors (PSD) integrated into the insole and flexible EMG to achieve gait recognition with an accuracy rate of 96.53%. [182] At the same time, this study points out that in future work, IMU should be combined to visually reflect the 3D motion information of lower limbs to compensate for the shortcomings of EMG and PSD sensors in sensing small changes in position information. Different from the insole integration, Cui implants force sensors into the floor, combining with EMG data, achieving 98.21% accuracy in Figure 12. Gait detection system based on PSD and EMG. The upper part is locations of sensor deployment and an overview of front-end EMG and PSD signals detection, and the lower part is plantar stress distribution sensor and EMG sensors. [182] Copyright 2021, IEEE. gait recognition. [215] Insole pressure measurement systems are limited by size and are more susceptible to damage; in contrast, floor pressure measurement systems are more stable but prone to some unnatural gaits. In addition to traditional physical therapy, VR technology, generally including IMU, and pressure sensors, is increasingly used in balance training. The advantages of VR technology are to enable patients to perform balance training in a multi-sensory, fully immersive, and obsessive environment, providing task-specific objectives, objective real-time feedback of actions, and facilitating enhanced personalized repetition. [216] In addition to IMU and pressure sensors receiving real-time mo-tion feedback from patients, more visual sensors such as infrared cameras and Kinect are used for recording. [217][218][219] Limb coordination refers to the spatiotemporal relationships between the kinematic, kinetic, and physiological variables of two or more limbs performing a motor task with a common goal. [220] It is widely agreed that quantifying patient coordination is crucial. Lewis points out in his study comparing the muscle activity of one and both hands that the peak of EMG activity can be determined as an indicator of the degree of muscle activation. [221] By observing the peak intensity and the relative time between the start of muscle activity and the peak muscle activity, the coordination Figure 13. a) Gesture recognition based on Leap Motion and pressure sensor. [181] Copyright 2016, MDPI. b) Comparing STS muscle synergy between normal and stroke patients. [180] Copyright 2019, IEEE.
ability of the movements of both hands can be analyzed. Some studies have shown that IMU with kinematic measurement ability can also assist in the evaluation of coordination. [222,223] Yu discusses hand coordination by using Leap Motion and pressure sensors (Figure 13a) to complete the task of two-handed gesture recognition, making it possible to coordinate bilateral rehabilitation training based on sensor fusion. [181] Yang compares muscle synergy between normal and stroke patients during sit-to-stand (STS) motion using EMG and pressure sensors (Figure 13b), and concludes that patients after a stroke still utilize the same amount of muscle synergy as healthy people, but the temporal structure of muscle synergy changes due to motor injury. [180] Gavrilovi´applies principal component analysis to the measurement of gait, comparing the temporal coordination of stroke patients before and after treatment through the most traditional combination of gait measurements, IMU, and pressure sensors, statistically confirming the hypothesis of invariant temporal synergies at different gait velocities. [163] Compared with the traditional evaluation indicators of gait coordination (such as symmetry and temporal gait parameters), the two-dimensional PCA cyclograms is more direct.
In this section, the application of multi-sensor fusion in various stages of stroke is introduced. There are some other studies whose application periods are not clear, which can also pro-vide some reference possibilities for the future combination of medicine and medicine. [224][225][226][227][228][229][230][231] Study characteristics related to the multimodal sensor applications in stroke rehabilitation and its placement, evaluation content, performance or result, algorithm, and feature for the included papers are presented in Table 4.

Medical Scales and Sensor Data
Many applications of sensors in stroke rehabilitation have been described above. But the use of some sensors is relatively unfamiliar to therapists, especially how to translate raw sensor data into clinical interpretations familiar to doctors-accepted clinical scales that have been clinically validated for years.
There are many researches on sensor in the field of stroke rehabilitation, but the researches on matching sensor data to medical scale are relatively few and the regularity is not strong. Through the analysis of these studies, the authors divided the matching process into three steps: • Characteristic data extraction: The original data first needs to undergo preliminary processing, such as noise reduction, filtering, interpolation. Then calculate the required characteristic scale, select the appropriate threshold value, fitting curve, algorithm, and process the characteristic data into the grade or score of the same dimension as the scale. • Verification: After matching is completed, the grade or score after processing is compared with the score of experienced doctors using the clinical scale in advance for verification, to draw the correlation conclusion.
This section explains some of these studies and divides them into three categories according to the processing methods of characteristic data: threshold based, fitting based, and model based.

Threshold Based
Clinical scales that need to calculate thresholds are generally graded, like TIS and Brunnstrom Scale. Alhwoaimel takes the AUC in ROC curve analysis as the cutoff point (the maximum AUC is the best cut-off value), and grades the data recorded by the IMU of patients' trunk ROM. PCC analysis is used to analyze the correlation between clinical classification of TIS and instrument score. [105] Similar to Alhwoaimel's threshold method, Han grades the sagittal motion amplitude and rotation data of a single inertial sensor fixed on the instep, and uses sensitivity, specificity, positive predictive value, negative predictive value, accuracy and F1 score to test the algorithm effect. The overall accuracy reached 86.8% in the Brunnstrom stage of lower limb. [233]

Fitting Based
Scales that require data fitting or general relationships tend to be of the rating type. Using force sensors, an records the STS motion of patients and calculates six mechanical characteristics of this process, including the maximum force during the STS motion, force per motion time, peak time, start time, end time and duration time. PCC analysis showed that FMA data is negatively correlated with 6 characteristics. [131] Through IRTATS, Fan defined 4 kinematic variables: MT (the movement time in the reaching and returning phases), MV (the velocity of the maximum tangential linear movement of the hand), NVP (the number of velocity peaks in the reaching and returning phases), and HPR (the actual hand path traveled/an ideal straight line between the start and end positions). These variables quantify the duration, speed, smoothness, and efficiency of the movement of the upper limb to the target, and obtain an excellent result (ICC = 0.95) by inter-rater reliability detection with the FMA score. [149] In view of the limitations in the MAS evaluation process, Kim proposes to quantify MAS by using angle sensor and sEMG to calculate TSRT, and obtains a negative correlation result (0.74). When the EMG number increases by two times the standard deviation from the mean baseline EMG, the joint angle and angular velocity values should be recorded. TSRT is the intercept value obtained by linear fitting the values of multiple groups of records. [234] The calculation process is similar to the above, but the difference of Zhang is that the angle information required by TRST is calculated by IMU. [203]

Model Based
In model-based matching methods, the relationship between characteristic data and clinical scales is not intuitively clear because the characteristic data may not have clear clinical significance. At this time, some models are needed, constantly "learning" the relationship between data and scale, and constructing the numerical relationship between the two.
For MAS, the correlation of Wang model (0.86) is higher than that of TSRT. The algorithm first uses an improved genetic algorithm to identify a parametric impedance model consisting of inertia, damping and stiffness, and then extracts the electrophysiological features from the biceps electromyographic signals by an EEMD (ensemble empirical mode decomposition) process that breaks them down into a series of IMF (intrinsic mode functions) signals. The fusion model SAS (spasticity assessment score) is constructed based on SVR by connecting the quantitative output of biomechanical and neurophysiological levels. [177] To cope with the complexity of FMA, Yu extracts five features from the original sensor data: amplitude (AMP), mean value (MEAN), root mean square value (RMS), root mean square value of derivative (JERK), approximate entropy (ApEn), uses the extremum learning machine (ELM) algorithm to map the eigenvalues to the clinical FMA score, and applies RRelief algorithm to find the optimal feature subset, finally, achieving a determination coefficient of 0.917 for predicting FMA scores in the upper limb. [144] Cui proposes a walking ability score (WAS) algorithm with a maximum correlation of 0.84 with Wisconsin Gait Scale (WGS). He carries out cubic spline interpolation and resampling on the data of the three modes of marked trajectory (MT), ground reaction force (GRF) and electromyogram, normalizes the data of each mode to 101 points per channel time (representing the gait period from 0% to 100%), and defines the prediction probability of the classifier as WAS. Chen calculates 18 kinds of characteristic data, such as A_AreaComp_F (Forefoot pressure to body weight ratio in affected side), by using motion data of the lower limb inertial sensor and pressure data of the planar pressure sensor. With the help of k-Nearest Neighbor, the accuracy rate of Brunnstrom stage reaches 94.2%. [235] Through leave-onesubject-out cross-validation based hierarchical discriminant analysis (LCHDA), Datta grades the Time Domain Lagged Cross-Correlation (TD-LCC), Time Domain Activity Cross-Coherence (TD-ACC), and Frequency Domain Magnitude Squared Coherence (FD-MSC) extracted from the IMU data, and obtains an average match rate of 91% with the National Institutes of Health Stroke Scale (NIHSS). [236] For Berg balance scale (BBS) matching, via SVD, Greene classifies three new parameters: CAAS (center of all active sensors), CHAT (centroid of heel and toe points), MOP (magnitude of pressure), obtained by pressure sensors and IMUs. [237] Table 5 lists sensor types, scale types, characteristic data, matching methods, and verification methods of each study in this section.

Discussion
In the previous section, we have discussed a variety of applications where smart sensors are utilized for retrieving objective data from stroke patients. Nevertheless, some issues are still waiting to be addressed. In this section, we discuss the three most significant ones from our point of view.

Standardization
As explained in the first two sections, stroke rehabilitation is a continuous process and sensors could be employed during the entire process. One of the most powerful functions of sensors is the ability to capture the continuous tiny change in patients' bodies and display objective data to medical staffs. However, in the clinic, sensors' data are not fully utilized. Currently, we do not have a well agreed standard for different types of sensory data. No matter what a high sensitivity can a sensor provide, we merely classify patients into traditional evaluation scales. It is expected that new standardizations could be developed based on the new objective multidimensional sensing data. This requires long-lasting work on interpreting sensor data into medical and clinical explanations.

Low Cost and IoHT
Stroke is a worldwide disease, and it is more severe in lower and lower-middle income countries (86.0% of deaths and 89.0% of disabilities). In these countries, wearable systems should be affordable for patients. In addition, many patients living there lacks sufficient medical resources. Hence, the deployment of the Internet of Health Things (IoHT) based on low-cost wearables would be a potential solution for improving the rehabilitation effects in these areas.

Development of Rehabilitation Aids
Most of the stroke patients, if not all, suffer different levels of motion ability loss. As stated previously, nearly half of patients cannot regain self-care abilities after rehabilitation training, Hence, the development of rehabilitation aids becomes important. In aids design, sensors are majorly utilized in transferring patients' intentions into control signals for external aids such as exoskeleton. During this procedure, advanced algorithms are taking the same weight as accurate sensing. Because time delays will inevitably be generated from sensing to controlling, the long delays will give a poor user experience or even result in unexpected danger. Nevertheless, there is always a trade-off between accuracy and delay. More data sources and volume often indicate higher accuracy, while will give longer delay. Hence, algorithm develop-www.advancedsciencenews.com www.advsensorres.com ment and sensor design should be considered together for optimizing the aids' performance.

Conclusion
Helping stroke patients restore a certain self-care ability requires worldwide efforts, and designing a customized training plan for specific patients for achieving an optimized rehabilitation effect is highly desired. In this review, we discussed how sensors function in the whole stroke rehabilitation process, and link clinical needs to engineering specifications. One thing we omitted here is the discussion of the impact of emotion on the rehabilitation training effect. This is owing to that current research on emotion detection is strongly based on EEG and emotion induction, which are not convenient for most stroke patients at the current stage. Stroke rehabilitation is a highly interdisciplinary area, and a universal language for researchers of different backgrounds to understand each other is expected. We are toward this target, building the universal language, although it is challenging.