WO2019154911A1 - System for personalized robotic therapy and related methods - Google Patents

Abstract

The present invention provides an apparatus and a method able to estimate motor improvement in real-time during three-dimensional rehabilitation tasks and to consequently dynamically personalize the therapy.The method can be carried out by a computer program. The use of said apparatus for restoring motor functions in a subject suffering from neuromotor impairment is also within the scope of the invention.

Description

System for personalized robotic therapy and related methods

FIELD OF THE INVENTION

The present invention refers to the field of robotic rehabilitation, in particular to systems and methods for improving rehabilitation of subjects after neurological disorders, in particular after stroke.

BACKGROUND OF THE INVENTION

Following a spinal cord injury or a stroke accident, the recovery of reaching and grasping abilities in upper limb and of skilful walking, running or jumping in lower limb is mandatory to allow patients to return to activities of daily living. In the past twenty years, robotic rehabilitation approaches showed great potential for the recovery of lost motor abilities.

Current robotic rehabilitation systems provide intense and highly repeatable treatment while also offering ways to control and quantify the therapy.

Methods and systems for improving computer- or robot-aided physical therapies and rehabilitation are mentioned, for example, in the following patent literature US2014287389, WO2017039553, WO2016096525, WO2017081647.

Nevertheless, an ideal robotic system should also be able to quantify motor improvement in patients and adapt the therapy accordingly, hence optimizing motor learning by matching the level of the training task to the patient’s level of ability (Guadagnoli and Lee 2004).

Current robotic rehabilitation devices do not embody this feature and training adjustments are based solely on the therapist’s abilities to choose proper exercises according to the patient’s performance level.

However, especially for three-dimensional movements, the visual evaluation of motor performance is a challenging task for the therapist.

Automatic detection of motor improvement and automatic adaptation of the motor task become even more compelling with the use of exoskeleton devices. Exoskeletons are wearable robotic devices where the limb is enclosed in an actuated robotic suit conform to the configuration of the limb (Maciejasz et al. 2014). They can be designed to cover as many degrees of freedom as the ones of the human limbs and to precisely determine the position and the delivered assistance torque at each articular joint (Lo and Xie 2012). Exoskeletons offer several advantages over end-effector robots, in particular for upper limb rehabilitation: they enlarge the task space to three dimensions, they follow the arm in its natural workspace with no restrictions, and they allow the independent or synergistic motion of shoulder, elbow and wrist joints during the execution of functional movements (Milot et al. 2013).

Whereas the extension to a three-dimensional workspace is important to enlarge the patient’s training workspace, the visual evaluation of the motor performance by the therapist becomes much more challenging in these more complex environments. The increased difficulty in the evaluation of the patient’s performance can lead to a maladaptation of the trained tasks and, thus, a longer and less efficient training. In this regard, the robotic system can provide an efficient way to autonomously and quantitatively assess the motor improvement of patients during their rehabilitative treatment, hence providing a measure to personalize the training.

Indeed, the personalization of the rehabilitative intervention has been suggested as a critical step to improve the outcome of robot-based rehabilitation (Fuhrer and Keith 1998, Krakauer 2006). Notable advances in the field of personalized robotic rehabilitation (Krebs, Palazzolo et al. 2003, Jezernik et al. 2004, Guerrero et al. 2010, Koenig, Omlin et al. 201 1 , Novak, Mihelj et al. 201 1 , Kan et al. 201 1 , Panarese, Colombo et al. 2012, Papaleo, Zollo et al. 2013, Metzger, Lambercy et al. 2014, Octavia and Coninx 2014, Wittmann et al. 2015, Badesa et al. 2016, Wu et al. 2016) included not only the personalization of the assistance provided by the robot, but also the adjustment of the proposed task. In particular, the movements proposed for the rehabilitation, as well as the difficulty of the training games, are based on kinematic performances recorded by the robot, such as accuracy, smoothness, speed, inter-joint coordination, range of motion and stiffness (Krebs, Palazzolo et al. 2003, Jezernik et al. 2004, Kan et al. 201 1 , Panarese, Colombo et al. 2012, Papaleo, Zollo et al. 2013, Metzger, Lambercy et al. 2014, Octavia and Coninx 2014, Wittmann et al. 2015, Wu et al. 2016), on game-related statistics (Octavia and Coninx 2014), on the muscle activity (Krebs, Palazzolo et al. 2003), or on the combination of kinematics and psychophysiological measurements (Guerrero et al. 2010, Koenig, Omlin et al. 201 1 , Novak, Mihelj et al. 201 1 , Badesa et al. 2016).

However, most of the aforementioned approaches are limited to specific conditions and setups (e.g., they can only be applied either in lower- or in upper-limb rehabilitation). Moreover, in the setups targeting upper- limb rehabilitation, the personalization was mostly implemented on planar robotic devices, limiting the range of motion to a two-dimensional workspace.

In addition, real-time implementation of the personalization is an essential feature to optimally exploit the advantages of a personalized rehabilitation training. However, only few approaches comprised a real-time implementation of the personalization (Jezernik et al. 2004, Guerrero et al. 2010, Koenig, Omlin et al. 201 1 , Papaleo, Zollo et al. 2013, Wittmann et al. 2015, Badesa et al. 2016, Wu et al. 2016).

More in particular, all the above mentioned approaches lacked the ability to integrate the information coming from multiple variables into one performance measure which meaningfully describes motor improvement of patients during their rehabilitation training.

This latter was addressed in the work of Panarese et al. (Panarese et al. 2012). The authors used a state- space model to merge the information from different performance measures and estimated the motor improvement of chronic stroke patients exercising with a planar robotic device. The estimates were computed on a daily basis and offline, i.e., the computations were executed after the completion of each training session. Similar approaches have been previously used to characterize learning processes in animals (Smith 2004, Prerau 2009). However, the method presented in the work of Panarese et al. was limited to the application on planar upper limb rehabilitation robots. Another downside of the implementation realised by Panarese et al. is that the personalization was not implemented in real-time, meaning that task adaptation could not be applied immediately but only after the completion of each training session, leading to a reduced efficiency of the personalization. In view of the above, there is still the need of a method and a system able to personalize robotic therapy in real-time, which could be used with any kind of robotic rehabilitation device, in particular with those supporting three-dimensional movements, and which is not limited to specific conditions and setups, in particular which can be used for both lower- and upper-limb rehabilitation.

The present invention relates to an integrated system and a method able to estimate motor improvement in real-time during three-dimensional rehabilitation tasks and to consequently personalize the therapy. This renders robotic rehabilitation training more efficient and motivating, eventually leading to better training outcomes for the patients alongside a reduced workload for physical therapists.

The method and system of the invention overcome the above mentioned drawbacks of the prior art by (i) developing a generalizable method to estimate motor improvement at a subtask level, which can be used with any kind of robotic rehabilitation device, but in particular with those supporting three-dimensional movements; (ii) implementing it in real-time, a feature which is essential to optimally exploit the advantages of the personalization; (iii) developing a procedure to automatically personalize the rehabilitation training in real-time based on the motor improvement estimates provided by the model.

The method of the invention provides not only a tool to estimate motor improvement in real-time, but also a way to automatically update the difficulty of the motor tasks performed during the robotic rehabilitation training.

The method of the invention therefore provides a novel tool to identify patient-specific strengths and weaknesses in real-time during the rehabilitation training, which is accordingly customized dynamically to avoid over- and/or undertraining and to focus on identified motor deficits in order to make the treatment more effective and motivating.

SUMMARY OF THE INVENTION

An object of the present invention is a method for operating an apparatus for restoring voluntary control of upper and/or lower limbs in a subject suffering from a neuromotor impairment, wherein said apparatus comprises at least a robotic-assistive device, one or more sensors, a processor controlling said robotic- assistive device and means for data storage, said method comprising the following steps:

a. storing on a data medium a list of motor tasks and/or subtasks to be performed by the subject; b. performing an initial assessment of the subject-specific level of ability for any motor task and/or subtask of said list and ordering said tasks and/or subtasks by increasing difficulty, so that subtasks with the lowest performance values are classified as difficult, then saving the obtained order on a data medium;

c. choosing from step b) the subset of tasks and/or subtasks having the easiest tasks and/or subtasks as the initial training subset to be performed by the subject and storing it as the current training subset on a data medium; storing the set of remaining tasks and/or subtasks as the training queue on a data medium;

d. setting the apparatus to apply the current training subset; e. recording signals providing features of motion for each movement repetition of said subject executing said task or subtask from sensors integrated in said robotic-assistive device and/or sensors external from said robotic-assistive device;

f. computing on said processor the estimated motor improvement MI_k for any said task or subtask executed by the subject according to the following steps:

fi. modelling motor improvement as a random walk expressed by the following formula:

MI_k = MI_k- + e_k

wherein MI_k is the motor improvement, k = 1,2 ... K are the different repetitions of a movement required to accomplish the task and/or subtask and e_k are independent Gaussian random variables with zero mean and variance s ;

f₂. computing the continuous performance variables h , wherein j = 1, 2, ...] represents the different measures, from the data obtained in step e) using a multi-paradigm numerical computing environment installed on said processor and storing the results on a data medium; f₃. defining the log-linear probability model for the continuous performance variables h \ log 0}_,k) = oi_j + b_}MI„ + ¾ _k

where oc_j

are unknown parameters, S_{j k} are independent Gaussian random variables with zero mean and variance

f₄. tracking the completion of any task or subtask by using a binary discrete variable n_k e (0,1), with 1 meaning that the task is performed successfully and 0 meaning failure,

f₅. defining an observation model for n_k as the following Bernoulli probability model:

P ¾IPfc) = R (! ^_ Pfc)^1_n¾

where the probability p_k of performing the task successfully at repetition k is related to the motor improvement x_k by the logistic function:

exp( MI_k)

k 1 + exp( MI_k )

ensuring that p_k is constrained in [0, 1] ;

f₆. estimating the parameters (a_jt b_], s_{d .}, s_e, p_k) using the values of r_{j k} and n_k previously recorded by applying statistical estimation methods executed in a computer program loaded on said processor, so that the estimation of the parameters results in an estimate of the motor improvement MI_k,

wherein said step f) is only executed if K > -g- , whereas if K < - - step e) is repeated; g. evaluating for each subtask whether the motor improvement values are higher than 0 and the difference between two consecutive motor improvement values is smaller than 5% for at least four repetitions; in the affirmative case removing said task or subtask from the current training subset and replacing it by the next task or subtask from said training queue, unless the training queue is empty;

h. updating the current training subset and training queue according to task or subtask replacements performed in step g) and storing the updated current training subset and training queue on a data medium;

i. repeating steps d)-h) until the current training subset is empty.

In a preferred embodiment, said binary discrete variable n_k is defined by the utilization of the robotic assistance during the execution of a sub-task, wherein it is“1” if the subject reaches a target without robotic support and it is“0” if the subject is completely dependent from robotic assistance.

In a preferred embodiment, the initial difficulty evaluation of each motor task and/or subtask described in step b) is performed by evaluating their values for n_k (i.e., tasks and/or sub-tasks with a high proportion of n_k= 1 are considered less complicated than those with a high proportion of n_k = 0), with k = 1, 2 ... K being the different movement repetitions performed for a task and/or sub-tasks. If several tasks and/or subtasks have equal proportions for n_k, the order amongst them is determined by their average values of h (i.e., specific performance measures with j = 1, 2, ...J being the number of performance measures), while giving all performance measures h equal weight.

In a preferred embodiment, continuous performance variables h are selected from (i) the average movement velocity and (ii) the spectral arc length, which is a consistent measure for movement smoothness.

In a preferred embodiment, the method for the estimation of the unknown parameters in real-time is Bayesian Monte Carlo Markov Chain method.

In an embodiment, the subtask removed in step g) is re-inserted into the training queue, so that it can be subsequently re-executed.

It is also an object of the invention an apparatus for estimating motor improvement of a subject during robotic rehabilitation therapy and consequently adjusting said rehabilitation therapy, said apparatus comprising: a) a robotic rehabilitation device, in particular a robotic exoskeleton, having at least 6 degrees of freedom, comprising one or more joints which can be actuated by integrated motors;

b) sensors for recording signals providing features of motion of said subject;

c) at least one processor controlling said robotic device a) and receiving signals from said sensors b) and executing a rehabilitative motor task in a virtual environment comprising a three-dimensional workspace wherein one or more targets to be reached by the subject are set;

d) means for data storage;

wherein said apparatus is operated by the above disclosed method.

All the components comprised in the apparatus, in particular items a)-d), are operatively connected. In particular, said processor operates the above-disclosed method to estimate and track in real-time the motor improvement of said subject and consequently adjust in real-time the therapeutic protocol modifying said rehabilitative motor task.

In a preferred embodiment, said sensors b) are integrated in said robotic rehabilitation device a).

In an embodiment, said sensors are force and/or position sensors.

In an embodiment, said sensors are external from the robotic device.

Also, both integrated and external sensors can be present in the apparatus of the invention.

In an embodiment, two processors are present, a first processor controlling said robotic device a) and executing a rehabilitative motor task in a virtual environment comprising a three-dimensional workspace wherein one or more targets to be reached by the subject are set and a second processor operatively connected to said robotic device a) and to said first processor wherein said second processor operates the above disclosed method and provides in real-time the adjusted therapeutic protocol to said first processor in order to modify said rehabilitative motor task. Optionally, said processors are integrated in the robotic device.

In a preferred embodiment, said means for data storage are embedded in the processor and/or in the robotic device. Data storage can be for example a hard drive.

In another embodiment, the apparatus comprises a computer system comprising said processor and said means for data storage.

The components of said system are preferably connected one with each other through a cable and/or wireless connection.

In a preferred embodiment, said processor includes a software control for gravity and friction compensation to allow the subject to move in the most natural and unconstrained way.

In a preferred embodiment, said targets are point-to-point reaching tasks.

For example, said rehabilitative motor task can be represented by a center-out reaching task, in which a cursor in the virtual environment represents the position of an end-effector of the robotic device.

In an embodiment, up to 18 targets are placed in said virtual environment on a three-dimensional sphere and the cursor is controlled by the subject through movements of his/her limb enclosed by the robotic exoskeleton.

The apparatus can further comprise a display wherein said targets are presented to the subject one after another.

The data from said sensors of the robotic device a) are preferably acquired at a frequency of at least 1000Hz.

Said robotic device a) preferably provides the modalities of active, passive and assist-as-needed.

Said robotic device is preferably a robotic exoskeleton mechanically compliant for the human upper or lower limb. For example, said robotic device can be the robotic exoskeleton named Armeo Spring and Armeo Power (Hocoma) or ALEx Rehabilitation Station (Wearable Robotics, Pirondini, Coscia et al. 2016). This system has been previously presented in WO2013186705.

It is also an object of the invention a computer program for carrying out the method above disclosed.

A data medium having said computer program is also within the scope of the invention.

A processor or a computer system on which said computer program is loaded is a further object of the invention.

A further object of the invention is the above-disclosed apparatus for use for restoring motor functions in a subject suffering from neuromotor impairment.

In an embodiment, the apparatus is for use for the recovery of reaching and grasping abilities in one or both upper limbs in a subject suffering from neuromotor impairment.

In another embodiment, the apparatus is for use for the recovery of locomotor functions in one or both lower limbs in a subject suffering from neuromotor impairment.

Said neuromotor impairment can be consequent to a stroke, a spinal cord injury, a neurodegenerative disease or a neurological disease.

Preferably, said subject suffered a stroke.

DETAILED DESCRIPTION OF THE INVENTION

Figures

Figure 1. Overall system description for a possible embodiment of the invention for the use in upper-limb rehabilitation

Figure 2. Design of a three-dimensional motor task for a possible embodiment of the invention for the use in upper-limb rehabilitation

Figure 3. Schematic of the system

Figure 4. Flow chart of the different steps of the method.

Figure 5. Experimental setup and protocols (a) Overall description of experimental setup (b) Design of three-dimensional point-to-point reaching task. Eighteen targets (representing the different subtasks) are positioned over a sphere of 19 cm of radius (equally distributed on the three planes). The empty circle represents the centre of the workspace (starting position) (c) Experimental protocol for patients. During the initial (A_{| ,}I ,A| ₂) and final (A_{F ,}I ,A_{F 2}) assessment sessions, all 18 targets were presented to the patients. Eight training targets for each treatment session were selected by a physical therapist (standard robotic therapy) or by the algorithm (personalized robotic therapy). The total number of repetitions performed in each session was determined by the therapist (d) Experimental protocol for healthy participants. Experiments were completed in a single session and divided into blocks (one block initial assessment A_| , five inversion blocks B _Ί _₅ , one block final assessment A_F). Each assessment block consisted of three runs, each composed of 18 reaching movements (one towards each target). Each training block consisted of five runs, each composed of eight reaching movements. The training targets for the inversion blocks were automatically selected by the algorithm. Breaks were allowed between the blocks to prevent muscle fatigue.

Figure 6. Analysis of performance measures for patients P01 and P02. (a) The first three rows depict the mean values for mean velocity (MV), spectral arc length (SAL) and rate of success (SUCC) for each assessment and treatment session for P01 (grey) and P02 (black). Measures were averaged for all targets presented during a session, shaded areas depict standard error of the mean (SEM). The last row depicts the scores on the Fugl-Meyer scale for upper extremities (FMA-UE) and grip strength assessed using the Jamar dynamometer before (A_| -_I ,A_{| 2}) and after (A_{F 1},A_{F 2}) the treatment. Dotted line indicates the maximum achievable score for FMA-UE. Shaded areas depict grip strength ranges for age- and gender-matched healthy subjects (values taken from existing literature (Mathiowetz et al. 1985)). (b) The first three rows depict the performance measures of P01 on a subtask level for two example targets (i.e., targets 6 and 16). Data for MV and SAL were low-pass filtered for visualization purposes (real data shown in light grey). Repetitions for each target are concatenated for all training sessions and presented in chronological order. Last row shows an offline estimation of motor improvement using the proposed model and the recordings of MV, SAL and SUCC. Data of the offline motor improvement estimates (M l₀ffiine) were low-pass filtered before being evaluated by the algorithm (raw motor improvement estimates shown in light grey). Dotted lines depict necessary condition (Ml₀ffiine > 0) for suggestion of target replacement. Grey area indicates the time point where the algorithm detects a performance plateau and suggests a target replacement.

Figure 7. Analysis of performance measures for the experiment with healthy participants. Average values of mean velocity (MV, panel a), spectral arc length (SAL, panel b), rate of SUCC (panel c), and directional error (DE, panel d) for each run (eight reaching movements) of fast (dark) and slow (light) learners. Measures were averaged for all targets presented during a run and for all subjects of a group. Blind trials were excluded from the calculations. Shaded areas depict standard error of the mean (SEM). Vertical bars indicate average number and standard deviation of new targets introduced by the algorithm in each block for both groups.

Figure 8. Analysis of the replacement of initial training targets for fast and slow learners. Bars show the percentage of subjects in each group for which a target was replaced in B₃, B₄, B₅ or was not replaced at all. No targets were replaced in B-i and B₂ due to lack of data needed for proper estimation of motor improvement. First and second column report the set of easy and difficult targets resulting from the analysis. Off-axis targets are grouped as OA.

Figure 9. Examples of motor improvement estimates and performance measures on a subtask level. Data is presented for the same two targets (i.e., targets 10 and 13) for a fast learner and a slow learner. Repetitions for each target are concatenated for all inversion blocks and presented in chronological order. Data for mean velocity (MV) and spectral arc length (SAL) were low-pass filtered for visualization purposes (raw data shown in light grey). Data for motor improvement (Ml) were low-pass filtered before being evaluated by the personalization routine (raw motor improvement estimates shown in grey). Dotted lines depict one of the necessary conditions (Ml > 0) for suggestion of target replacement. Dark grey areas indicate the time span where the algorithm detects a performance plateau and suggests a target replacement.

Figure 10. Analysis of performance measures during blind trials. Mean values of mean velocity (MV, panel a), spectral arc length (SAL, panel b), rate of SUCC (panel c), and directional error (DE, panel d) for each inversion block (40 reaching movements) for fast (dark) and slow (light) learners. Measures were averaged for all targets presented during a block and for all subjects of a group. Solid lines depict mean values for movements executed with visual feedback (regular trials), dashed lines depict mean values for movements executed without visual feedback (blind trials). Error bars depict SEM.

Figure 11. Analysis of target replacement for P03. The data show the number of repetitions needed for a target to be replaced by the personalization algorithm. Targets are ordered by subject-specific difficulty following the evaluation of the initial assessment sessions. The first eight targets were selected as initial training targets, the remaining ten were placed in the training queue, waiting to be inserted. Once all targets were replaced by the algorithm, the targets were reintroduced and presented alternatingly in the order they were replaced by the algorithm.

Figure 12. Analysis of performance measures for patient P03. (a) The first three rows show the mean values for mean velocity (MV), spectral arc length (SAL) and rate of success (SUCC) for each assessment and treatment session. Measures were averaged for all targets presented during a session, shaded areas depict standard error of the mean (SEM). The last row shows the scores on the Fugl-Meyer scale for upper extremities (FMA-UE) and grip strength assessed using the Jamar dynamometer before (Al ,1-2) and after (AF,1-2) the treatment. Dotted line indicates the maximum achievable score for FMA-UE. Shaded area depicts grip strength range of age- and gender-matched healthy subjects. Values are taken from existing literature (Mathiowetz et al. 1985) . (b) The first three rows show the performance measures at a subtask level for two example targets of P03. Data is shown until the moment each target was replaced by the personalization algorithm. Repetitions for each target are concatenated for all training sessions and presented in chronological order. Data for MV and SAL were low-pass filtered for visualization purposes (real data shown in light grey). Last row shows the real-time estimation of motor improvement using the presented model and the recordings of MV, SAL and SUCC. Data of the motor improvement estimates (Ml) were low- pass filtered before being evaluated by the personalization routine (raw motor improvement estimates shown in light grey). Grey area indicates the time point where the personalization algorithm detects a performance plateau and suggests a target replacement.

Figure 13. Results of simulated data for 25 repetitions of the same movement. Data shows the simulated performance measures and the corresponding motor improvement estimates under four different conditions. In the first column, the time constant for the mean velocity (MV) was reduced to t = 5 repetitions (t = 15 repetitions for remaining two measures), in the second column the time constant of the spectral arch length (SAL) was reduced t = 5 repetitions and in the third column the time constant reduced t = 5 repetitions was the one of the probability of success (p_k). In the last column, the time constant for all performance measures was set to t = 5 repetitions. The first two rows show simulated data (black dots) for MV and SAL. The third row depicts the simulated data for p_k (black dots) and the corresponding discrete performance measures SUCC (grey squares) deduced from p_k using a Bernoulli distribution model. Grey lines show approximations of the performance measures using the estimated parameters resulting from the algorithm. Shaded area depicts 95% confidence interval of the approximations. The last row shows the resulting offline motor improvement estimates (Ml₀ffiine) using an offline implementation of the algorithm. Dotted lines depict necessary condition (Ml₀ffiine > 0) for suggestion of target replacement. Grey area indicates the time span where the algorithm detects a performance plateau and suggests a target replacement.

Definitions

Within the frame of the present invention, the following definitions are provided.

“Apparatus”: means an ensemble of components or devices cooperating to provide a more complex function. “System” is herein used as a synonym of“apparatus”.

“Operatively connected” means a connection capable of carrying data flow between two or more input and/or output data ports. The connection can be of any suitable type, a wired or a wireless connection.

“Processor”: means any device capable of elaborating input signals and produce output signals. Said device can be incorporated in a more complex apparatus or system, such as for example a computer.

For“facilitating standing and walking” is intended an increase of the movements magnitudes of the hind-limb joints as well as an improvement of locomotor stability.

For“facilitating reaching and grasping abilities” is intended an increase of the movements magnitudes of the upper limb joints as well as an improvement of movement speed and accuracy.

For“task” is intended an ensemble of training movements to be performed by the subject (e.g., set of movements toward targets).

For“subtask” is intended an individual movement to be performed by the subject (e.g., movement toward a specific target).

For“robotic-assistive device” is intended a machine capable of assisting rehabilitation training. Preferably, said robotic device also measures training-related variables. “Robotic device” or “robotic rehabilitation device” are herein used as synonyms.

For“training session” is intended the period during which the subject uses the robotic-assistive device to perform the task.

For“training subset” is intended the set of subtasks performed by a subject during a training session.

The present invention will be now disclosed in detail referring to the exemplary embodiment of facilitating reaching and grasping abilities, with particular reference to upper limbs, being intended that the teaching of the invention is applicable to every kind of neuromotor impairments, such as, for example, impairment of lower limbs.

Detailed description of the invention

Apparatus

An exemplary representation of the system of the invention is represented in figure 1 and herein explained. According to an embodiment of the invention, the integrated system of the invention 1 comprises a robotic upper-limb exoskeleton device 2 that allows the execution of three-dimensional tasks. The three-dimensional workspace enables to design rehabilitative tasks similar to daily life scenarios, which can be realized using virtual environments. The robotic exoskeleton is equipped with force and position sensors and has a microprocessor 3 for the high-level control of the robot. Moreover, the integrated system comprises a rehabilitative motor task embedded in a virtual environment stored in the microprocessor 3 as well as a second microprocessor 4 for the real-time estimation of the motor improvement and the implementation of the personalization.

Figure 1 shows a possible embodiment of the invention for the use in upper-limb rehabilitation. The subject 5 is seated in a comfortable chair integrated in the robotic exoskeleton. His impaired upper-limb is fastened to the limb of the robotic system 2. A motor task is executed on microprocessor A (3) and displayed on a monitor 6 in front of the subject. Microprocessor B (4) is used for the real-time estimation of motor improvement during the rehabilitation and for the implementation of personalization methods. The arrows show the data flows between the different components of the system.

Communication between the components is established through cable and/or wireless connection.

In an embodiment, the motor task for upper-limb rehabilitation is a center-out reaching task, which is currently widely used in robotic rehabilitation. This kind of task allows the subject to repeat general or task- oriented movements (such as planar or 3D reaching movements in multiple directions) with different sensory feedback (visual, audio, haptic, etc.), while possibly receiving assistance and/or resistance from the robotic device (Marchal-Crespo and Reinkensmeyer 2009).

Data exchange and communication between the components of the system can be established through cable and/or wireless connection.

A schematic overview of the components of the system and of their interactions is depicted in figure 3.

The following section describes the components of the system in more detail.

Robotic rehabilitation device

The robotic device is preferably a robotic exoskeleton. It should have at least 6 degrees of freedom, with one or more joints that can be actuated by integrated motors. The robotic device preferably has integrated force and position sensors at all relevant locations, which allow a full characterization of every movement performed. The data from the sensors are preferably acquired at a frequency of at least 1000Hz to provide a better functionality of the system. Data are preferably immediately accessible after the completion of a movement. Furthermore, the robotic exoskeleton preferably provides different modalities such as active (i.e., the subject moves the upper limb and the robot measures the movement), passive (i.e., the robot guides the upper limb of the subject during movement execution using the integrated actuators), and assist-as-needed (i.e., the robot guides the arm of the subject if he/she is unable to perform the movement).

First processor (Robot controller)

The first processor controls the robotic device and provides the motor tasks. It preferably includes a gravity and friction compensation to allow the subject to move in the most natural and unconstrained way. The rehabilitative motor task is executed on the first processor. Preferably, the rehabilitative motor task comprises a three-dimensional workspace, allowing the subject to train movements that are beneficial for the transfer to activities of daily living. Moreover, it preferably provides means to focus the rehabilitative treatment on the specific motor deficits and impairments of the subject.

In a preferred embodiment, a center-out reaching task is used, in which the subject, wearing the robotic exoskeleton, controls a cursor in a virtual environment using the position of the end-effector of the exoskeleton. The subject repeats a set of movements in different directions always starting from the same point, which defines the“center” of the workspace. Figure 2 depicts a possible design of such a center-out reaching task, wherein each numbered point represents a target; 18 targets are placed on a three- dimensional sphere and the cursor is controlled by the subject through movements of his/her upper limb enclosed by the robotic exoskeleton. During the rehabilitation task, the targets are presented to the subject one after another, and the subject is asked to move the cursor to the position of the target that is currently displayed. Once a target is reached, the subject is asked to move back to the center, before the next target is presented. Such design allows exploring a complete three-dimensional workspace and provides a way to evaluate the subject’s performance on a sub-task level, therefore for each single movement direction.

Second processor, for motor improvement estimation and personalization

A second processor is connected to the robotic device and the first processor typically through a cable and/or wireless connection. This processor is used for the real-time estimation of motor improvement and the implementation of the personalization, according to the method above described. It receives the data recorded from the integrated sensors of the exoskeleton, and/or from external sensors, after each reaching movement completed by the subject during the rehabilitative training. Through the real-time method herein disclosed and integrated in the processor, these data are used to estimate and track the subject’s motor improvement on a sub-task level. Based on these motor improvement estimations, the processor adjusts the task difficulty based on the current performance level of the subject, as described below in more details.

Preferably both processors are microprocessors.

As mentioned above, the two processors can be integrated in a single processor. Also, the processor or the processors can be integrated in the robotic device.

Method The present application provides a method for operating an apparatus, in particular for tracking and estimating motor improvement of a subject during a robotic rehabilitation therapy and consequently adjusting the rehabilitation therapy protocol.

An exemplary flowchart of the steps of the method of the invention is depicted in figure 4.

According to an exemplary embodiment of the method of the invention, motor improvement x_k is assumed to be an unobservable variable and modelled as a random walk:

MI_k = MI_k- + e_k

where k = 1,2 ... K are the different repetitions of a movement required to accomplish the task or subtask and e_k are independent Gaussian random variables with zero mean and variance s .

Since the motor improvement MI_k is assumed to be unobservable, a set of observation equations is defined. These equations relate the motor improvement to discrete and continuous performance measures, which are recorded by the integrated sensors of the exoskeleton or by external sensors.

Said sensors record signals providing features of motion of the subject. Said signals can be joints displacement and/or positions of the handle. In particular, said signals are kinematic data, such as three- dimensional Cartesian positions of the handle.

Continuous performance variables h , wherein j = 1, 2, ... J represents the different measures, are computed from kinematic data provided by the sensors using a multi-paradigm numerical computing environment installed on the robot’s processor. For example, variables can be computed from the Cartesian coordinates of the three-dimensional trajectory of the robotic handle or from the coordinates of the joints. An example of multi-paradigm numerical computing environment is MATLAB (matrix laboratory), a commercial programming language developed by MathWorks.

The continuous variables h are defined by the following log-linear probability model log 0}_,k) = oi_j + b_}MI„ + S_{j k}

where S_{j k} are independent Gaussian random variables with zero mean and variance s . The use of log- linear models captures the rapid increases (or decreases) of the performance measures at the beginning of the training, while towards the end a convergence towards subject-specific upper (or lower) bounds is expected.

The binary discrete variable n_k e (0,1) is used to track the completion of the given task, with 1 meaning that the task was performed successfully and 0 meaning failure. The observation model for n_k is assumed to be the following Bernoulli probability model

where the probability p_k of performing the task successfully in trial k is related to the motor improvement MI_k by the logistic function exp( MI_k )

k 1 + exp( MI_k )

ensuring that p_k is constrained in [0,1]. Furthermore, this formulation guarantees that p_k approaches 1 when the motor improvement is increasing.

The unknown parameters {ct_j, b^, s_d ., s_£, p_k} can be estimated using the recordings of r_{j k} and n_k and applying appropriate statistical estimation methods. Such methods are of common use in the field and well known to the skilled in the art. Consequently, the estimation of the unknown model parameters results in an estimate of the motor improvement MI_k.

In order to ensure accuracy of the model, the number of recordings of r_{j k} and n_k should exceed the number of unknown model parameters, hence estimation is only executed if K >—j—.

According to the present invention, the binary performance variable can be defined by the utilization of the robotic assistance during the execution of a sub-task, such that it is“1” if the subject reaches a target without robotic support and“0” otherwise.

Possible choices for continuous performance variables could be (i) the average movement velocity and (ii) the spectral arc length, a consistent measure for movement smoothness.

Bayesian Monte Carlo Markov Chain methods with appropriate initial guesses can be applied as a suitable statistical method for the estimation of the unknown parameters in real-time.

The following exemplary method is then used to personalize the training.

1. Initial Assessment of sub-tasks

Before the rehabilitation training starts, the performance of the subject for all sub-tasks is assessed. In particular, sub-tasks can be ordered by increasing difficulty (i.e., subtasks with low performance values are classified as difficult). Tasks and/or sub-tasks can be first ordered by their values for n_fc (i.e., tasks and/or sub-tasks with n_k = 1 are considered less complicated than those with ( n_k = 0). If several tasks and/or subtasks have equal values for n_k , the order amongst them can be determined by their values of h , while giving all performance measures h equal weight.

The subset with the easiest sub-tasks can then be chosen for the initial training session.

2. Sub-task replacement on improvement

The above-disclosed method is used to continuously track the motor improvement for each sub-task throughout the rehabilitation training. When the subject executes said sub-task in such a manner that the motor improvement values exceed the threshold ( MI_k > 0), and the difference between two consecutive motor improvement values is smaller than 5% for at least four repetitions, the processor removes this subtask from the current training subset and replaces it with a more difficult one according to the previously established training queue. The process goes on until both the training queue and the training subset are empty, i.e., the above condition on motor improvement has been satisfied for all the tasks and subtasks of the training session.

Following this scheme, a subject gradually explores new challenges, based on his/her current and past performance.

In an embodiment of the invention, the method is for the use in upper limb rehabilitation of stroke patients. However, the actuation of the method is not limited to such application. Indeed, it can be adapted for the use with any robotic rehabilitation device and motor learning task, using different performance measures and/or different observation equations.

In particular, it can also be applied to lower limb rehabilitation.

The real-time functionality and the identification of subject-specific strengths and weaknesses on a subtask level tremendously enhance robotic rehabilitation training in general, making it more purposive, efficient and motivating for the patients. Thanks to the fully automated character of the personalization procedure, the therapists can shift their attention from the selection of the movements to be performed to other important aspects of the training, hence further improving the rehabilitation process of the patients.

More in particular, the method of the invention implements a model-based approach for the personalization of robotic rehabilitation training based on the real-time estimation of motor improvement in three-dimensional training tasks.

With respect to the prior art, the extension from planar to three-dimensional workspaces, according to the present invention, permits the training of a wider range of natural and functional movements. Capitalizing on the enhanced potential for plasticity in the early stage after the injury (Biernaskie et al., 2004; Cramer et al., 2008), the method is able to continuously track motor improvement in subacute stroke patients.

The model-based approach to personalize robot-aided rehabilitation therapy in real-time herein disclosed was validated in experiments with three subacute patients and seventeen healthy subjects.

In the following an exemplary embodiment of the application of the method and apparatus of the invention is described.

In order to test the model, we used the robotic upper limb exoskeleton ALEx Rehabilitation Station (Wearable Robotics, Pirondini, Coscia et al. 2016) and designed a three-dimensional point-to-point reaching task, a training exercise that is commonly used in robotic rehabilitation therapy (Coscia et al., 2014; Frisoli et al., 2007; Krebs et al., 2004). Such design allowed the patients to explore an extensive workspace and at the same time it provided a way to easily identify the different subtasks of the exercise. The results showed that there were indeed remarkable differences in performance for different subtasks (i.e., movements towards different targets). Due to the design of the point-to-point reaching task, it was possible to detect these performance differences on a subtask level, hence justifying the choice of this training exercise for the use with the developed model. Our model was able to differentiate between (i) subject-specific improvement and (ii) target-specific improvement (i.e., improvement on a subtask level) for each individual. Three performance measures, computed from movement kinematics, were selected and further used by the model: movement velocity (MV), spectral arc length (SAL) and robot assistance dependency (SUCC). Indeed, the selected performance measures were able to capture different dynamics of recovery. The data of two stroke patients (P01 and P02) were used in order to support this, and the analysis emphasized the ability of the model to compute reliable motor improvement estimates based on those measures. Moreover, the evolution of the performance measures showed strong correlations with clinical scores, such as the Fugl- Meyer assessment for upper extremities and the Jamar dynamometer for grip strength.

The offline estimation of motor improvement for P01 highlighted the potentiality of the chosen approach to enhance robotic rehabilitation therapy by introducing a real-time adaptation of the task difficulty. For instance, following the therapist’s decision, this patient performed 25 repetitions for the presented target 6, although the performance for this target was sufficiently good from the very beginning of the training. In this context, the application of the proposed real-time personalization algorithm would have prevented the redundant repetitions by replacing target 6 with a more difficult one, hence making the training more challenging and motivating. The potential of this solution was further supported by the pilot test performed with a third patient (P03), during which the algorithm was applied in real-time throughout the rehabilitation therapy.

The test showed that targets were indeed replaced by the algorithm as the motor improvement estimates for these targets approached a plateau. Notably, the performance for these targets was retained when they were reintroduced as training targets, indicating the usability and efficacy of the chosen approach.

We also put emphasis on the psychological potentials of the personalization. We noticed that P03 showed increased motivation and satisfaction every time the algorithm introduced a new target. Motivation is known to be a crucial factor in rehabilitation and finding ways to maintain and improve it has always been a matter of interest (Maclean et al. 2000, Maclean et al. 2002, Colombo et al. 2007). With regard to this issue, it seems like the fully automated and dynamic character of our approach can actually have positive impacts on the patient’s engagement.

The experiments with the healthy participants further validated the algorithm’s capability to dynamically track subject-specific motor improvement using the chosen performance measures. The participants were grouped into fast and slow learners, based on the number of new training targets introduced by the algorithm. This division was strongly supported by the differential evolution of the directional error (DE) for the two groups. Interestingly, this confirmation stems from an independent performance measure, which was not used by the algorithm for the motor improvement estimation. While the fast learners adapted almost instantaneously to the inverted vision, the group of slow learners needed considerably more time to reach similar performances.

The subject-specific dynamics were well captured by the algorithm, on both the overall performance level and the subtask performance level. Furthermore, the developed model provided well-timed manner task difficulty adjustments, i.e., targets were replaced when subjects reached a performance plateau for that target.

Medical uses A further object of the invention is the above-disclosed apparatus for use for restoring motor functions in a subject suffering from neuromotor impairment.

In an embodiment, the apparatus is for use for the recovery of reaching and grasping abilities in one or both upper limbs in a subject suffering from neuromotor impairment.

In another embodiment, the apparatus is for use for the recovery of locomotor functions in one or both lower limbs in a subject suffering from neuromotor impairment.

Said neuromotor impairment can be partial or total paralysis of upper or lower limbs.

Said neuromotor impairment may have been caused by a spinal cord injury, Parkinson’s disease (PD), an ischemic injury resulting from a stroke, or a neuromotor disease as, for example, Amyotrophic Lateral Sclerosis (ALS) or Multiple Sclerosis (MS) or a neurodegenerative disease or neurological disease.

Preferably, the device is used for facilitating motor functions in a subject after stroke.

The skilled person in the field knows how to use the apparatus of the invention in view of the desired therapeutic application and his general knowledge in the medical field.

The invention will be further described by means of examples.

EXAMPLES

Methods

The development and validation of the model was divided into several steps. We developed a model to continuously estimate and track patient’s motor improvement in real-time and in a three-dimensional workspace. We then developed a personalization routine, which automatically adapts the difficulty of the rehabilitative motor task to the current performance level of the patient using these estimates. The motor improvement was estimated using different performance measures identified through the analysis of the kinematic data recorded from two patients who completed a robot-assisted rehabilitation therapy based on three-dimensional point-to-point reaching movements while wearing a robotic upper limb exoskeleton (Bergamasco 2013, Pirondini, Coscia et al. 2016). We then tested the model on a group of 17 healthy participants, who completed a modified version of the point-to-point reaching task. As a proof of concept of the usability and efficacy of the developed method, the model was finally validated in a pilot test with a third subacute stroke patient who completed a personalized rehabilitation training, where the difficulty of the point- to-point reaching task was automatically adapted by the developed personalization algorithm.

1. Motor improvement model

In order to continuously track patient’s motor improvement in real-time and at a subtask level (i.e., for different movement directions), we used a state-space model. In a nutshell, motor improvement is assumed to be an unobservable variable ( MI_k ) and is modelled as a random walk

MI_k = MI_k-1 + e_k where k = 1, 2, ... K are the different repetitions of a movement and e_k are independent Gaussian random variables with zero mean and variance s . A set of observation equations is defined in order to estimate the motor improvement. These equations relate the motor improvement to continuous performance measures h , which can be computed from the recordings provided by the robotic rehabilitation device.

The continuous variables h (with j = 1, 2, .../) are defined by the log-linear probability model: log 0}_,k) = oi_j + b_}MI„ + ¾ _k

where S_{j k} are independent Gaussian random variables with zero mean and variance s . The use of log- linear models allows capturing the rapid increase (or decrease) of the performance measures at the beginning of the training, as well as the expected convergence towards subject-specific upper (or lower) bounds at the end of the training.

Similarly, an observation equation for a discrete performance measure n_k is defined. The binary discrete variable n_k e (0,1) is used to track the completion of the exercised subtask, with 1 meaning that the subtask was performed successfully and 0 meaning failure. The observation model for n_k is assumed to be a Bernoulli probability model:

where p_k, the probability of performing the subtask successfully at repetition k, is related to the motor improvement MI_k by the logistic function:

exp( MI_k)

k 1 + exp (MI_k)

ensuring that p_k is constrained in [0,1]. Furthermore, this formulation guarantees that p_k approaches 1 when the motor improvement is increasing.

The unknown model parameters {ct_j, b_], s_{d .}, s_e, p_k} are estimated using the recordings of r_{j k} and n_k and by applying Bayesian Monte Carlo Markov Chain methods with appropriate initial guesses. The estimation of the unknown model parameters also results in an estimate of the motor improvement MI_k. In order to ensure accuracy of the model, it is necessary that the number of recordings of r_{j k} and n_k exceeds the number of unknown model parameters. Consequently, no reliable motor improvement estimates can be obtained if K, the total number of performed repetitions for a movement, is too small.

The model was implemented in the task control unit of the robotic rehabilitation device in order to track the motor improvement of patients and healthy participants in real-time. As we aimed at estimating motor improvement on a subtask level, separate state-space models were used for each subtask of the training exercise. In order to validate the capability of the proposed model to appropriately capture variable dynamics of the performance measures, we simulated different rehabilitation scenarios under varying conditions (see figure 13).

2. Personalization routine Using the model described in point 1 , motor improvement estimates were computed for each subtask and used to build a personalized training routine, where specific subtasks were selected based on these estimates. We identified the subject-specific level of difficulty for each subtask following an initial assessment of the training exercise (see 6.1 Experimental protocol patients). The subtasks were then ordered by increasing difficulty and the easiest ones were selected for the initial training set. The remaining subtasks were placed in a training queue, waiting to be inserted. During the training, a subtask was removed from the set of current training subtasks when the motor improvement estimate for this subtask exceeded a given threshold and started to reach a plateau. Specifically, the probability of performing the subtask successfully ( p_k ) had to be greater than 0.5, and the difference between two consecutive motor improvement values (i.e., between two repetitions of the same subtask) had to be smaller than 5% for at least four repetitions. Given the observation equation for p_k, the former condition ( p_k > 0.5) can be equally expressed in terms of the motor improvement, i.e., MI_k > 0. Once the conditions were satisfied, the subtask was replaced by a more difficult one from the training queue. The removed subtask was placed back into the training queue, so that it could be reintroduced at a later stage.

3 Participants

3.1 Subacute stroke patients

Three subacute stroke patients (a male, two females, respectively 86, 68 and 34 years old) from the inpatient unit of the Hopitaux Universitaires de Geneve (HUG, Geneva, Switzerland) were included in the study. A summary of the patient information is reported in table 1. All patients suffered a right hemiplegia with at least 10° of residual motion in shoulder and elbow joints. The patients were enrolled in the study within 2-4 weeks after the stroke. The first two patients (P01 and P02) underwent a therapy following the standard robotic rehabilitation protocol (see 6. Experimental protocols). Their data were used to identify the performance measures, which were integrated in the model presented in point 1. The developed real-time algorithm was finally validated in a third patient (P03), who underwent a therapy following the personalized robotic rehabilitation protocol (see 6. Experimental protocols).

3.2 Healthy participants

Seventeen right-handed subjects (eight males, nine females, 25.4±3.3 years old) participated in the experimental validation of the algorithm. The participants did not present any evidence or known history of skeletal and neurological diseases and they exhibited normal ranges of motion and muscle strength. All participants gave their informed consent to participate in the study, which had been previously approved by Commission Cantonale d'Ethique de la Recherche Geneve (CCER, Geneva, Switzerland, 2017-00504).

Table 1. Demographics and information about stroke patients included in the study

4. Robotic exoskeleton and motor task

We implemented our model and the personalization routine in the robotic exoskeleton ALEx (Wearable Robotics srl, Ghezzano, Italy, (Bergamasco 2013, Pirondini, Coscia et al. 2016). ALEx is a six degrees of freedom (DOFs) mechanically compliant exoskeleton for the human upper limb. Four DOFs are sensorised and actuated (shoulder abduction, rotation, flexion and elbow flexion), and two DOFs are sensorised and passive (forearm prono-supination and wrist flexion). ALEx allows simultaneous recordings of the six joint angles along with the three-dimensional trajectories of the handle. When wearing ALEx, the subjects are seated in a chair and the posture of the back is ensured with seat belts. The user’s shoulder acromion is aligned to the centre of rotation of ALEx’s shoulder joint to guarantee the alignment of ALEx’s joints axes with the corresponding axes of the human articular joints.

During the experiments, patients and healthy participants were instructed to perform point-to-point reaching movements using the robotic handle of the exoskeleton (figure 5a). All reaching movements started from the centre of the workspace and the goal was to reach one of the eighteen targets distributed over a sphere of 19 cm of radius (figure 5b). Such design of the motor task made it possible to exploit an extensive three- dimensional workspace and provided means to easily identify all subtasks of the exercise (i.e., each target represents a subtask). The sphere was positioned so that its centre was at a comfortable distance for the subject and aligned with the acromion of the right arm. The targets were displayed on a monitor mounted in front of the subjects and visual feedback was provided by means of a cursor mapping the position of the handle of the exoskeleton into the virtual environment.

If a subject was unable to reach a target (i.e., the subject did not move for more than 3 seconds), ALEx activated its assistance mode to guide the subject towards the target according to a minimum jerk speed profile (Sadaka-Stephan, Pirondini et al. 2015).

5. Performance measures

Two measures were selected as continuous performance variables h for the use with the state-space model presented in Section 1 : (i) the mean velocity of a movement (MV) and (ii) the spectral arc length (SAL), a robust and consistent measure of movement smoothness (Balasubramanian et al. 2012). SAL is a dimensionless measure quantifying movement smoothness by negative values, where higher absolute values are related to jerkier movements. With regard to the rehabilitation training, values of SAL close to zero are desirable as well as high values of MV. Both variables were computed from the Cartesian coordinates of the three-dimensional trajectory of the robotic handle. The discrete variable n_k was denoted as success (SUCC) and was defined separately for the experiments with patients and healthy participants. For the patients, the value of SUCC was determined by the robotic assistance (i.e., SUCC = 1 if the patient performed the movement without robotic assistance, SUCC = 0 otherwise). On the other hand, healthy participants were expected not to rely on the robotic assistance and we therefore, defined the value of SUCC based on the execution time (i.e., SUCC = 1 if a healthy participant completed the movement within a time threshold, SUCC = 0 otherwise). The time threshold t_th was set to 4 seconds based on previous experiments with healthy subjects.

6. Experimental protocols

6.1 Experimental protocol patients

The experimental protocol for the patients consisted of 4 weeks of robot-assisted rehabilitation therapy (figure 5c). Patients trained with ALEx three times a week for 30 minutes, performing the previously described point-to-point reaching movements (see 4 Robotic exoskeleton and motor task). In order to evaluate the outcome of the rehabilitation training, patients went through two assessments before the treatment ( -i and A_{! 2}) as well as two assessments following the therapy (A_{F 1} and A_{F 2}). The initial assessment sessions Ai_,-i and A_F2 were completed two weeks and one week before the beginning of the therapy. The final assessment sessions A_F -i and A_{F 2} were completed one week and one month after the end of the therapy. During the assessment sessions, all 18 targets of the point-to-point reaching task were presented to the patients in a randomized order. The patients were asked to perform as many repetitions of the point-to-point reaching task as possible (on average 27.5 (54.0) repetitions per session at A_|,1-2 (A_{F .2}) for P01 , 36.0 (145.5) repetitions for P02 and 26.0 (54.0) repetitions for P03). In addition to the robotic assessment, patients were also assessed using the upper extremity section of the Fugl-Meyer assessment (FMA-UE) (Fugl-Meyer et al. 1975) and the Jamar dynamometer for grip strength (Hamilton et al. 1992). The procedures for the assessment sessions were identical for all patients. However, different protocols were applied during the treatment sessions, in which two different ways of selecting the training targets were used.

Standard robotic rehabilitation protocol

Patients P01 and P02 followed a standard robotic rehabilitation protocol in which a physical therapist selected eight training targets at the beginning of each session. The choices were solely based on the therapist’s evaluation of the patient's performance. The selected targets were presented in the same repetitive order and were not changed within a training session. The total amount of repetitions for each session was determined by the therapist while encouraging the patients to perform as many movements as possible (on average 51.5 repetitions per treatment session for P01 and 82.6 repetitions for P02).

Personalized robotic rehabilitation protocol

The developed personalization algorithm was validated in a pilot test with a patient (P03) undergoing the personalized robotic rehabilitation training. In this case, the training targets were not fixed beforehand, but selected by the algorithm in real-time, and inserted and/or removed accordingly within a training session. Following the initial assessment sessions A_{| .2}, we identified the patient-specific difficulty for each of the 18 targets. The eight targets with the best values of the performance measures MV, SAL and SUCC were classified as easy and selected as the initial training targets. The remaining targets were placed in the training queue and ordered by increasing difficulty based on their values for MV, SAL and SUCC. Motor improvement was continuously estimated for each subtask separately. The replacement of a training target based on the motor improvement estimates followed the same procedure as presented above (see 2. Personalization routine). The current set of training targets was saved after the completion of each training session, ensuring continuity between sessions. The total amount of repetitions for each session was determined by a physical therapist while encouraging patient P03 to perform as many movements as possible (on average 34.6 repetitions per treatment session).

6.2 Experimental protocol healthy participants

The healthy participants attended a single experimental session that consisted in seven blocks of reaching movements (figure 5d). Breaks were allowed between the blocks to prevent muscle fatigue.

The session started with an initial assessment block similar to the assessment sessions of the patients. The initial assessment block consisted of three runs (A_| -i_₃) and during each run all 18 targets were presented once in a randomized order. The purpose of the assessment block was (i) to familiarize the subjects with the robotic system and the reaching task and (ii) to record a baseline for the performance measures. The initial assessment block was followed by five blocks B-I_₅ in which the visual feedback was inverted (i.e., an upward movement was displayed as downward and vice versa, likewise for left/right and forward/backward movements). At the onset of these five blocks, participants were not informed about the feedback manipulation, but were told that the task difficulty was changed. We expected that this modification would increase the difficulty of the point-to-point reaching task, especially in the early exposure. However, previous studies on visually manipulated motor tasks showed that with training most participants were able to cope with similar manipulations (Shabbott et al. 2010, Bock et al. 2010, Krakauer 2009, Miall et al. 2003, Harris 1965). Therefore, we hypothesized that performance would drop after the introduction of the inverted visual feedback (i.e., movements would become slower and jerkier), but would then gradually improve and eventually reach a plateau - a behaviour which is similar to the one observed in the rehabilitation of stroke patients. Using this setup, we wanted to test whether the developed personalization algorithm was indeed capable of tracking these improvements in real-time and whether it was able to personalize the training by identifying“recovered” movements. Each of the five blocks B-I_₅ consisted of five runs, each one composed of eight point-to-point reaching movements. The initial set of training targets for each healthy participant was generated following a semi-random procedure. It always contained the six on-axis targets (i.e., targets 1 ,3,5,7,10 and 13, see figure 5b) and two other randomly selected off-axis targets (i.e., targets 2,4,6,8,1 1 ,14,15, 16, 17 and 18). The remaining ten off-axis targets were placed randomly in the training queue. Previous studies (Werner et al., 2010) using planar setups demonstrated that participants showed better performance for targets lying on the axis of the inversion, since for these targets no adaptation was necessary. Although in this study we used a three-dimensional setup, we also expected the subjects to have less difficulties with the on-axis targets, as they involved inversions in only one dimension (in contrast to two dimensions for the off-axis targets). The personalization algorithm was used during the five blocks B-I_₅ and a target was removed from the current set of training targets if the motor improvement estimates for this subtask satisfied the conditions presented above (see 2. personalization routine). In this case, it was replaced by the next target from the training queue.

In order to investigate whether the adaptation to the inverted environment was dependent on the visual feedback, we introduced blind trials at every fifth reaching movement. During a blind trial, subjects were asked to reach the presented target without any visual feedback of the cursor position during the first two seconds of the movement. If a participant was not able to reach the presented target within those two seconds, the visual feedback was reactivated, in order to ensure the completion of the reaching task.

The inversion blocks B-I_₅ were followed by a final assessment block which was composed of three runs (A_F,I- ₃) and followed the same procedure as the initial assessment block (i.e., neither visual manipulation nor personalization were applied).

In order to further investigate the adaptation of the healthy subjects to the inverted environment, an additional performance measure was introduced, which was not used by the algorithm: the directional error (DE). DE was computed as the angular difference between the ideal and the actual movement direction at peak velocity during the first two seconds of the movement. The ideal movement direction was defined as the vector connecting the starting position of the movement and the position of the displayed target. DE was used only for validation purposes and did not contribute to the estimation of the motor improvement.

7. Statistical analysis

A Student’s t-test was used to analyse the statistical significance of the correlation between the performance measures and the clinical evaluations. A two-sample t-test was used to analyse the statistical significance of the performance differences in the experiments with healthy participants. A two-way ANOVA was used to analyse the effects of the visual manipulation on the healthy participants. We used a significance level of 0.05 for all analyses. All analyses were performed using MATLAB (The MathWorks, Natick, Massachusetts).

Results

Example 1

Analysis of performance measures

The data recorded from two patients (P01 and P02) were used to identify the performance measures for the state-space model. Those two patients completed the robot-assisted therapy following the standard robotic rehabilitation protocol (see 6. Experimental protocols). In this case, a physical therapist was in charge of setting the task difficulty at the beginning of each training session.

The overall evaluation of the performance measures highlighted an improvement for both patients during the rehabilitation training (figure 6a). This improvement can be summarized by comparing the average values of the assessment sessions right before (A_{! 2}) and right after (A_{F 1}) the rehabilitative treatment. P01 and P02 improved in terms of movement velocity (MV) by 1 16% and 46%, respectively, movement smoothness (SAL) by 20% and 32% and rate of SUCC by 59% and 5%. These improvements were equally reflected by a change of +7 points and +12 points in the Fugl-Meyer assessment for upper extremities (FMA-UE) between A_| 2 and A_{F 1}. A strong correlation was found between most of the performance measures and the FMA-UE scores: the Pearson correlation coefficient between those variables for P01 and P02 was respectively 0.96 (p=0.046) and 0.95 (p=0.034) for MV, 0.97 (p=0.032) and 0.99 (p=0.006) for SAL and 0.99 (p=0.005) and 0.92 (p=0.084) for the rate of SUCC. As evaluated with the Jamar dynamometer, we also observed an increase in grip strength for both patients: between A_{! 2} and A_F,I , P01 and P02 improved their grip strength by 5.6 kg and 2.3 kg, respectively. We also observed a strong correlation between some of the selected performance measures and the assessed grip strength, for both patients: the Pearson correlation coefficient between those variables for P01 and P02 were respectively 0.98 (p=0.016) and 0.95 (p=0.052) for MV, 0.84 (p=0.160) and 0.99 (p=0.014) for SAL and 0.93 (p=0.067) and 0.92 (p=0.084) for the rate of SUCC.

The time courses of MV, SAL and SUCC for the movements of P01 towards two selected targets (targets 6 and 16) are provided as examples (figure 6b). Using the recorded data and an offline implementation of the proposed algorithm, we computed offline estimates of motor improvement (Ml_0ffiine) for both examples. Those offline estimates (figure 6b, bottom row) suggested a target replacement following 9 repetitions for target 6, while 19 repetitions were necessary for target 16. These suggestions were supported by the evolution of the performance measures in both examples. The need of a different number of repetitions for the two targets was supported by the evolution of the performance measures and highlighted the capacity of our approach to detect different time courses of motor improvement at a subtask level.

Example 2

Experimental validation on healthy participants

We tested the developed personalization algorithm with a group of 17 healthy participants, who performed a modified version of the point-to-point reaching task (see 6. Experimental protocols). In particular, the visual feedback was manipulated during the five inversion blocks B-_|-B₅ (i.e., an upward movement was displayed as downward and vice versa, likewise for left/right and forward/backward movements). Using this setup, we aimed at testing whether the algorithm was capable of tracking improvements in real-time and whether it was able to personalize the training by identifying “recovered” movements (i.e., sufficiently well executed subtasks) and by replacing them with more difficult ones.

2.1 Task adaptation on a subject level

The results of the experiments on healthy participants revealed a division of the subjects into two groups, according to their performance in the visually manipulated reaching task. Participants were grouped based on the number of new targets that were introduced during the inversion blocks B-I_₅. We found significant differences (p<0.001 ) between the groups and grouped participants into fast (n=9, 7.7±1.2 new targets) and slow learners (n=8, 2.6±2.0 new targets). We then carried on the analysis for the two groups separately.

Both groups already showed remarkable performance differences at the initial assessment A_| -I_₃: interestingly, the average values for MV were significantly higher (p<0.001 ) for the slow learners (0.171 m/s) in comparison to the fast learners (0.159 m/s). In contrast, the values for SAL were significantly lower (p<0.001 ) for the slow learners (-3.134) compared to the fast learners (-2.807). For the rate of SUCC no significant difference (p=0.060) was found between slow learners (95%) and fast learners (97%). As expected, the performance measures dropped for both groups after the introduction of the visual manipulation (figure 7). However, the drop was remarkably lower for the fast learners: between A_{! 3} and the first run of B-i, the values for MV worsened by 29% for the fast learners (41 % for slow learners), SAL by 50% (157% for slow learners) and the rate of SUCC by 57% (87% for slow learners). A two-way ANOVA validated the different impact of the visual manipulation on both groups for MV (F_{1 42}I = 10.62, p=0.001 ), SAL (F_{1 42}I = 1 12.31 , p<0.001 ) and rate of SUCC (F_{1 42}I = 30.38, p<0.001 ). Both groups gradually improved during the experiment (i.e., from block B-i to B₅), although they did not reach their initial baseline level. However, a comparison between the last run of B₅ and the last run of the initial assessment A_{! 3} showed that the fast learners were more successful in restoring their initial performance: MV was lowered by 9% (30% for slow learners), SAL by 22% (25% for slow learners) and rate of SUCC by 15% (50% for slow learners). During the entire experiment, the slow learners were outperformed by the fast learners, who reached better final values for all performance measures (19% higher values for MV, 15% for SAL and 31 % for rate of SUCC at the last run of B₅, values compared between groups).

In order to investigate whether the adaptation to the inverted environment was dependent on the visual feedback, we introduced blind trials at every fifth reaching movement. For MV and SAL, both groups exhibited similar trends for regular (i.e., trials with visual feedback) and blind trials (figure 10a-b), indicating that the visuomotor adaptation was retained even when the visual feedback was not provided. Although the trends were similar, the absolute values for MV and SAL during the blind trials were slightly lower compared to the regular trials. The reduced movement speed and smoothness was probably induced by the movement corrections performed after the reappearance of the cursor (i.e., after 2 seconds). Due to the increased difficulty to complete the reaching task without visual feedback, most participants exceeded the given time threshold t_th when they eventually reached the target and the rate of SUCC in blind trials was almost 0% for all participants (figure 10c).

In order to better characterize the visuomotor adaptation of the healthy participants to the inverted environment, we introduced an additional performance measure, calculated independently from the measures used by the algorithm: the directional error (DE, see 6.2 Healthy participants). Also for DE (figure 7d) we observed remarkable differences at the initial assessment A_| -_|.3 between the groups: the average values of DE were significantly lower (p<0.001 ) for the fast learners (23.4°) compared to the slow learners (34.0°). An increase of DE was observed in both groups after the introduction of the inversion, which represents a larger angular difference between the ideal and the actual movement direction (figure 7d). Between A_{! 3} and B-i , DE worsened by 71 % (125%) for the fast (slow) learners, which followed the same evolution we observed from MV, SAL, and rate of SUCC. A two-way ANOVA validated the different effects of the visual manipulation on the values of DE for both groups (F_{1 42}I = 22.77, p<0.001 ). Specifically, at the onset of the inversion phase (first run of B-i), the fast learners showed an increased value of DE (36.7°), that was eventually reduced and reached a plateau after two runs (DE of 28.3° at the third run of B-i). This value was slightly higher than the average baseline level recorded in A_| -_|.3 (23.4°). In contrast, the slow learners exhibited a considerably slower adaptation to the inversion. They started from a much higher DE value at the first run of B-i (80.4°) and fifteen runs were necessary to reach values similar to the ones of the fast learners (27.9° at the first run of B₄) - remarkably, these values were better than their average baseline values from A_| -_1.3 (34.0°). Finally, both groups showed slight after effects at the beginning of the final assessment, in which the inversion was removed: between B₅ and A_{F 1} the values of DE again increased by 16% (19%) for the fast (slow) learners. Both groups readapted to the regular point-to-point reaching task during the three runs A_F,I-₃ and approached their initial baseline values at the last run of the final assessment A_{F 3} (28.0° for fast learners, 37.8° for slow learners).

For both groups, the evolution of DE showed a similar trend for normal and blind trials, again highlighting that the visuomotor adaptation was retained when no visual feedback was provided (figure 10d). A slight increase of DE was observed after the third block (B₃), in particular for the blind trials (on average increase of 23% for fast learners in B₄ and B₅, 17% for slow learners).

As the amount of data was not sufficient to obtain proper motor improvement estimations (see 1. Motor improvement model), no new training targets were introduced during blocks the B-i and B₂. For the fast learners, new training targets were inserted starting from B₃ (3.0±1.9 new targets on average). In contrast, for most of the slow learners the algorithm started to insert new targets in B₄ (1.5±1.7 new targets on average). The insertion of new training targets during B₄ is in accordance with the evolution of DE for this group. Indeed, the slow learners reached DE values (33.7°) similar to the ones of the fast group (31 .5°) during B₄. The insertion of new targets at different time points during the training (i.e., in block B₃ for the fast learners and in B₄ for the slow learners) highlights the capacity of our approach to detect motor improvement at a subject level and to adapt the motor task accordingly.

2.2 Task adaptation on a subtask level

In order to analyse the ability of the personalization algorithm to track motor improvement at a subtask level, we further evaluated which of the initial training targets were replaced by the algorithm during the inversion blocks and when this replacement occurred (figure 8). The analysis revealed that despite the differences in the overall performance, the sets of easy (1 ,7 and 10) and difficult (3,5, 13 and off-axis) targets appeared to be similar for fast and slow learners. As hypothesized in the experimental design, movements towards the off-axis targets (2,4,6,8,1 1 ,14,15,16, 17 and 18) seemed to be more difficult: on average, these targets were replaced by the algorithm for 12.5% of the slow learners and for the 77% of the fast learners. The on-axis targets (1 ,3,5,7,10 and 13) instead, were replaced for 38% of the slow learners and 87.0% of the fast learners. We also observed differences within the on-axis targets: on average, targets 3, 5 and 13 were replaced for 13% of the slow learners and for the 74% of the fast learners, while a replacement for targets 1 ,7 and 10 was achieved by 63% of the subjects in the slow learner group and by 100% of the fast learners. Furthermore, we observed an earlier replacement of easy targets for the fast learners: 56% of the targets from the easy set were replaced in B₃ (4.16% for slow learners), 33% were replaced in B₄ (33% for slow learners) and 1 1 % were replaced in B₅ (13% for slow learners). In contrast, for the difficult targets, also the fast learners needed more time to achieve a replacement (if replaced at all): 26% of the difficult targets were replaced in B₃ (3% for slow learners), 35% were replaced in B₄ (5% for slow learners) and 14% were replaced in B₅ (5% for slow learners).

In order to better understand the behaviour of individual participants on a subtask level, we present the data of an exemplary subject selected from each group (figure 9). Here we show the results for the movements of both participants towards the same two targets. We selected one target from the set of easy (target 10) and one target from the set of difficult (target 13) targets. The examples clearly illustrate the differences between both groups: although the easy target was replaced for both subjects, the slow learner needed three more repetitions to achieve the replacement. For the fast learner, all performance measures quickly reached a plateau, while the slow learner showed difficulties until the fourth repetition, reflected particularly by SAL and SUCC. However, starting from the fifth repetition the slow learner also managed to adapt the movements to the visual manipulation and finally reached the conditions for the target replacement at repetition 13. The difficult target instead, was more challenging for both subjects. For this target the fast learner showed an improvement in the performance, in particular SAL and SUCC, only after tenth repetition and finally reached the conditions for the target replacement at the very last repetition. In contrast, the slow learner did not reach the conditions for a replacement. Despite an improvement for all measures, the performance of the slow learner was not sufficient at any time point to allow a replacement of the target.

This example emphasizes the capability of the algorithm to capture individual time courses of improvement at a subject level. Moreover, it once again illustrates that using the algorithm, motor improvement can also be tracked on a subtask level. With regard to improving robotic rehabilitation therapy, this result is particularly interesting, since it allows to automatically identify subtask-specific strengths and weaknesses.

Example 3

Real-time validation in pilot test

We finally validated the algorithm in a pilot test with patient P03, who completed the therapy following the personalized robotic rehabilitation protocol (see 6. Experimental protocols). In this case, the difficulty of the point-to-point reaching task was automatically adapted by the developed algorithm, based on a continuous evaluation of the motor improvement estimates for each training target.

P03 was able to make great progress during the rehabilitation training and eventually achieved a replacement of all 18 targets (figure 11 ). At this point, previously removed targets were reintroduced as training targets and presented alternatingly in the order in which they were replaced by the algorithm. This led to a dynamic setting, enabling the patient to continue the training in a more diversified modality. We observed that the number of repetitions needed for a replacement was very low for some targets (e.g., target 9 (10 repetitions) and target 12 (8 repetitions)), while for others it was remarkably higher (e.g., target 10 (19 repetitions) and target 15 (21 repetitions)) compared to the average value (13.3±3.1 repetitions). This observation once more emphasized the importance of an algorithm able to detect motor improvement at a subtask level and demonstrated the abilities of our approach to differentiate between various subtask- specific time courses of improvement.

Similar to the results of P01 and P02, we observed an improvement for movement velocity (MV), movement smoothness (SAL) and robot assistance dependency (SUCC) throughout the training also for P03 (figure 12a). Comparing the values of the assessment sessions right before (A_{! 2}) and right after (A_{F 1}) the rehabilitative treatment, P03 improved MV by 77%, SAL by 21 % and the rate of SUCC by 22%. These improvements were equally reflected by a change of +8 points in the Fugl-Meyer assessment for upper extremities (FMA-UE) between A_F2 and A_{F 1}. A correlation tendency was found between the performance measures and the FMA-UE scores: the Pearson correlation coefficient with FMA-UE was 0.81 (p=0.188) for MV, 0.93 (p=0.070) for SAL and 0.93 (p=0.070) for the rate of SUCC. We also observed a slight improvement in grip strength using the Jamar dynamometer: between A_F2 and A_{F 1} , the grip strength of P03 improved from not measurable to 1.0 kg.

P03 was able to make great progress during the rehabilitation training and eventually achieved a replacement of all 18 targets. From then on, previously removed targets were reintroduced as training targets and presented in a randomized order. Interestingly, we observed that the number of repetitions needed for a replacement varied for the different targets, again highlighting the capability of the algorithm to differentiate motor improvement on a subtask level. As examples, we present the time courses of MV, SAL and SUCC for the movements towards two example targets - one which was replaced quickly (target 12) and one which was kept for a longer period (target 15) as a training target (figure 12b). Once again, we observed a target-specific time course for the different performance measures, which was well captured by the motor improvement estimates provided by the model. The estimates of the motor improvement related to these examples suggested a target replacement following 12 repetitions for target 12, while 21 repetitions were necessary for target 15. Accordingly, the performance measures for these targets started to approach a plateau after 7 and 15 repetitions respectively. Interestingly, we also observed that performance was retained when the targets were reintroduced as training targets.

With regard to the clinical relevance of the proposed approach, the obtained results of this pilot test are highly promising. The well-timed and autonomous replacement of training targets made the rehabilitation therapy more diverse and motivating to the patient. The fact that the performance improved and was retained when a target was reinserted, indicates the usability and efficacy of the chosen approach.

Example 4

Supplementary data

In order to demonstrate the capability of the proposed model to capture varying dynamics of the performance measures, we simulated different rehabilitation scenarios under varying conditions (figure 13). Therefore, data was generated for the three variables MV, SAL and p_k using an exponential equation:

where j = 1,2,3 are the different performance measures and k = 1, 2, ...30 are the different repetitions of a movement. h_:bhά and h _3ίagί are parameters used to set the desired initial and final values of each performance measure t,- is the individual time constant for each performance measure. The equation was used to simulate the data of MV, SAL and p_k for 30 repetitions of the movement towards the same target. The values for SUCC were deduced by using the values of p_k and a Bernoulli distribution model. We ran the simulations under four conditions: in the first three conditions, the time constant of one performance measure was reduced to t = 5, while the other two were kept at t = 15. In the fourth condition the time constants for all three measures were reduced to t = 5. For all conditions, we obtained approximations of the simulated data by inserting the estimates of the unknown model parameters into the observation equations. Moreover, we calculated the 95% confidence intervals of the approximations and the corresponding motor improvement estimates. The results of the simulations illustrate the capability of the proposed model to capture varying dynamics of the performance measures properly. The simulated data lies within the 95% confidence intervals of the approximations for the most part. Moreover, the only condition where the requirement for a replacement is met, is the one where all performance measures are simulated with low time constants and quickly reach a plateau, highlighting the fact that a replacement is only suggested by the algorithm when no further improvement is expected.

References

Badesa, F.J., Morales R., Garcia-Aracil, N.M., Sabater, J.M., Zollo, L, Papaleo, E. and Guglielmelli, E., (2016), "Dynamic Adaptive System for Robot-Assisted Motion Rehabilitation," in IEEE Systems Journal, vol. 10, no. 3, pp. 984-991

Balasubramanian S, Melendez-Calderon A, Burdet E. A robust and sensitive metric for quantifying movement smoothness. IEEE Trans Biomed Eng. 2012;59(8):2126-36.

Bergamasco M, Salsedo F, Lenzo B. An exoskeleton structure for physical interaction with a human being. WO2013186705, 2013.

Biernaskie J. Efficacy of Rehabilitative Experience Declines with Time after Focal Ischemic Brain Injury. J Neurosci. 2004;24(5): 1245-54.

Colombo R, Pisano F, Mazzone A, Delconte C, Micera S, Carrozza MC, et al. Design strategies to improve patient motivation during robot-aided rehabilitation. J Neuroeng Rehabil. 2007;4:3.

Coscia M, Cheung V, Tropea P, Koenig A, Monaco V, Bennis C, et al. The effect of arm weight support on upper limb muscle synergies during reaching movements. J Neuroeng Rehabil. 2014; 1 1 (22):1— 15. Cramer SC. Repairing the human brain after stroke: I. Mechanisms of spontaneous recovery. Ann Neurol. 2008;63(3):272-87.

Frisoli A, Borelli L, Montagner A, Marcheschi S, Procopio C, Salsedo F, et al. Arm rehabilitation with a robotic exoskeleleton in Virtual Reality. 2007 IEEE 10th Int Conf Rehabil Robot ICORRO7. 2007;0(c):631-42.

Fugl-Meyer AR, Jaasko L, Leyman I, Olsson S, Steglind S. The post-stroke hemiplegic patient. 1. a method for evaluation of physical performance. Scandinavian Journal of Rehabilitation Medicine. 1975. p. 13-31.

Fuhrer, M. J. and R. A. Keith (1998). "Facilitating patient learning during medical rehabilitation: a research agenda." Am J Phys Med Rehabil 77(6): 557-561.

Guadagnoli, M.A. and Lee, T.D. (2004). Challenge point: a framework for conceptualizing the effects of various practice conditions in motor learning. Journal of Motor Behavior, 2004.

Guerrero, C. R., J. F. Marinero, J. P. Turiel and P. R. Farina (2010). Bio cooperative robotic platform for motor function recovery of the upper limb after stroke. Engineering in Medicine and Biology Society (EMBC), 2010 Annual International Conference of the IEEE, IEEE. Hamilton GF, McDonald C, Chenier TC. Measurement of grip strength: validity and reliability of the sphygmomanometer and jamar grip dynamometer. J Orthop Sports Phys Ther. 1992; 16(5):215-9.

Harris CS. Perceptual adaptation to inverted, reversed, and displaced vision. Psychol Rev. 1965;72(6):419- 44.

Jezernik, S., Colombo, G., and Morari, M. (2004). Automatic gait-pattern adaptation algorithms for rehabilitation with a 4-DOF robotic orthosis. IEEE Transactions on Robotics and Automation, 20, 574-582.

Kan, P., Huq, R., Hoey, J., Goetschalckx, R., & Mihailidis, A. (201 1 ). The development of an adaptive upper- limb stroke rehabilitation robotic system. Journal of NeuroEngineering and Rehabilitation, 8, 33. http://doi.Org/10.1 186/1743-0003-8-33

Klamroth-Marganska, V., J. Blanco, K. Campen, A. Curt, V. Dietz, T. Ettlin, M. Felder, B. Fellinghauer, M. Guidali and A. Kollmar (2014). "Three-dimensional, task-specific robot therapy of the arm after stroke: a multicentre, parallel-group randomised trial." The Lancet Neurology 13(2): 159-166.

Koenig, A., X. Omlin, J. Bergmann, L. Zimmerli, M. Bolliger, F. Mdller and R. Riener (201 1 ). "Controlling patient participation during robot-assisted gait training." J Neuroeng Rehabil 8(14.10): 1 186.

Krakauer, J. W. (2006). "Motor learning: its relevance to stroke recovery and neurorehabilitation." Curr Opin Neurol 19(1 ): 84-90.

Krakauer JW. Motor learning and consolidation: the case of visuomotor rotation. Adv Exp Med Biol. 2009;629:405-21.

Krebs, H. I., J. J. Palazzolo, L. Dipietro, M. Ferraro, J. Krol, K. Rannekleiv, B. T. Volpe and N. Hogan (2003). "Rehabilitation robotics: Performance-based progressive robot-assisted therapy." Autonomous Robots 15(1 ): 7-20.

Krebs HI, Ferraro M, Buerger SP, Newbery MJ, Makiyama A, Sandmann M, et al. Rehabilitation robotics: pilot trial of a spatial extension for MIT-Manus. J Neuroeng Rehabil. 2004;1 :5.

Lo HS, Xie SQ. (2012). Exoskeleton robots for upper-limb rehabilitation: State of the art and future prospects. Med Eng Phys. 2012;34(3):261-8.

Maciejasz P, Eschweiler J, Gerlach-Hahn K, Jansen-Troy A, Leonhardt S. (2014). A survey on robotic devices for upper limb rehabilitation. J Neuroeng Rehab. 2014; 1 1 :3.

Maclean N, Pound P, Wolfe C, Rudd A. Qualitative analysis of stroke patients’ motivation for rehabilitation. Bmj. 2000;321 (7268): 1051-4.

Maclean N, Pound P, Wolfe C, Rudd A. The Concept of Patient Motivation: A Qualitative Analysis of Stroke Professionals’ Attitudes Niall Maclean, Pandora Pound, Charles Wolfe and Anthony Rudd. 2002;444-9.

Marchal-Crespo, L. and D. J. Reinkensmeyer (2009). "Review of control strategies for robotic movement training after neurologic injury." J Neuroeng Rehabil 6: 20. Metzger, J.-C., O. Lambercy, A. Califfi, D. Dinacci, C. Petrillo, P. Rossi, F. M. Conti and R. Gassert (2014). "Assessment-driven selection and adaptation of exercise difficulty in robot-assisted therapy: a pilot study with a hand rehabilitation robot." Journal of neuroengineering and rehabilitation 1 1 (1 ): 154.

Miall RC, Jenkinson N, Kulkarni K. Adaptation to rotated visual feedback: A re-examination of motor interference. Exp Brain Res. 2004; 154(2):201-10.

Milot M-H, Spencer SJ, Chan V, Allington JP, Klein J, Chou C, et al. (2013). A crossover pilot study evaluating the functional outcomes of two different types of robotic movement training in chronic stroke survivors using the arm exoskeleton BONES. J Neuroeng Rehab. 2013; 10(1 ):1 12.

Novak, D., M. Mihelj, J. Ziherl, A. Olensek and M. Munih (201 1 ). "Psychophysiological measurements in a biocooperative feedback loop for upper extremity rehabilitation." Neural Systems and Rehabilitation Engineering, IEEE Transactions on 19(4): 400-410.

Octavia, J. R., & Coninx, K. (2014). Adaptive Personalized Training Games for Individual and Collaborative Rehabilitation of People with Multiple Sclerosis. BioMed Research International, 2014, 345728. http://doi.Org/10.1 155/2014/345728

Panarese, A., R. Colombo, I. Sterpi, F. Pisano and S. Micera (2012). "Tracking motor improvement at the subtask level during robot-aided neurorehabilitation of stroke patients." Neurorehabilitation and neural repair 26(7): 822-833.

Papaleo, E., L. Zollo, L. Spedaliere and E. Guglielmelli (2013). Patient-tailored adaptive robotic system for upper-limb rehabilitation. Robotics and Automation (ICRA), 2013 IEEE International Conference on, IEEE.

Pirondini E, Coscia M, Marcheschi S, Roas G, Salsedo F, Frisoli A, et al. Evaluation of the effects of the Arm Light Exoskeleton on movement execution and muscle activities: a pilot study on healthy subjects. J Neuroeng Rehabil. Journal of NeuroEngineering and Rehabilitation; 2016; 13(1 ):9.

Sadaka-Stephan A, Pirondini E, Coscia M, Micera S. Influence of trajectory and speed profile on muscle organization during robot-aided training. IEEE Int Conf Rehabil Robot. 2015;2015-Septe:241-6.

Shabbott BA, Sainburg RL. Learning a visuomotor rotation: Simultaneous visual and proprioceptive information is crucial for visuomotor remapping. Exp Brain Res. 2010;203(1 ):75-87.

Werner S, Bock O. Mechanisms for visuomotor adaptation to left-right reversed vision. Hum Mov Sci. Elsevier B.V.; 2010;29(2): 172-8.

Wittmann F., Lambercy, O., Gonzenbach, R., van Raai, M., Hover, R., Held, J., Starkey, M., Curt, A., Luft, A. and Gassert R., (2015). "Assessment-driven arm therapy at home using an IMU-based virtual reality system," 2015 IEEE International Conference on Rehabilitation Robotics (ICORR), Singapore, pp. 707-712.

Wu, W., Wang, D., Wang, T., and Liu, M. (2016) "A personalized limb rehabilitation training system for stroke patients," 2016 IEEE International Conference on Robotics and Biomimetics (ROBIO), Qingdao, pp. 1924- 1929.

Claims

1. A method for operating an apparatus for restoring voluntary control of upper and/or lower limbs in a subject suffering from a neuromotor impairment, wherein said apparatus comprises at least a robotic- assistive device, one or more sensors, a processor controlling said robotic-assistive device and means for data storage, said method comprising the following steps:

a. storing on a data medium a list of motor tasks and/or subtasks to be performed by the subject; b. performing an initial assessment of the subject-specific level of ability for any motor task and/or subtask of said list and ordering said tasks and/or subtasks by increasing difficulty, so that subtasks with the lowest performance values are classified as difficult, then saving the obtained order on a data medium;

c. choosing from step b) the subset of tasks and/or sub-tasks having the easiest tasks and/or subtasks as the initial training subset to be performed by the subject and storing it as the current training subset on a data medium; storing the set of remaining tasks and/or subtasks as the training queue on a data medium;

d. setting the apparatus to apply the current training subset;

e. recording signals providing features of motion for each movement repetition of said subject executing said task or subtask from sensors integrated in said robotic-assistive device and/or sensors external from said robotic-assistive device;

f. computing on said processor the estimated motor improvement MI_k for any said task or subtask executed by the subject according to the following steps:

fi. modelling motor improvement as a random walk expressed by the following formula:

MI_k = MI_k- + e_k

wherein MI_k is the motor improvement, k = 1,2 ... K are the different repetitions of a movement required to accomplish the task and/or subtask and e_k are independent Gaussian random variables with zero mean and variance s ;

f₂. computing the continuous performance variables h , wherein j = 1, 2, ...] represents the different measures, from the data obtained in step e) using a multi-paradigm numerical computing environment installed on said processor and storing the results on a data medium; f₃. defining the log-linear probability model for the continuous performance variables h \ log {r_{J k}) = cc_j + b_]MI + 5_{j k}

where oc_j

are unknown parameters, S_{j k} are independent Gaussian random variables with zero mean and variance

f₄. tracking the completion of any task or subtask by using a binary discrete variable n_k e (0,1), with 1 meaning that the task is performed successfully and 0 meaning failure, f₅. defining an observation model for n_k as the following Bernoulli probability model:

PrOⁱ Pk) = R (! ^_ Pk)^1_n¾

where the probability p_k of performing the task successfully at repetition k is related to the motor improvement MI_k by the logistic function:

exp( MI_k)

k 1 + exp( MI_k )

ensuring that p_k is constrained in [0,1];

f₆. estimating the parameters (a_jt b_], s_{d .}, s_e, p_k) using the values of r_{j k} and n_k previously recorded by applying statistical estimation methods executed in a computer program loaded on said processor, so that the estimation of the parameters results in an estimate of the motor improvement MI_k,

wherein said step f) is only executed if K > -g- , whereas if K < - - step e) is repeated; g. evaluating for each subtask whether the motor improvement values are higher than 0 and the difference between two consecutive motor improvement values is smaller than 5% for at least four repetitions; in the affirmative case removing said task or subtask from the current training subset and replacing it by the next task or subtask from said training queue, unless the training queue is empty;

h. updating the current training subset and training queue according to task or subtask replacements performed in step g) and storing the updated current training subset and training queue on a data medium;

i. repeating steps d)-h) until the current training subset is empty.

2. The method according to claim 1 , wherein said binary discrete variable n_k is defined by the utilization of the robotic assistance during the execution of a sub-task, wherein it is“1” if the subject reaches a target without robotic support and it is“0” if the subject is completely dependent from robotic assistance.

3. The method according to claim 1 or 2, wherein said continuous performance variables are selected from (i) the average movement velocity and (ii) the spectral arc length.

4. The method according to anyone of claims 1-3, wherein the statistical method for the estimation of the unknown parameters in real-time is Bayesian Monte Carlo Markov Chain method.

5. The method according to anyone of claims 1-4, wherein the subtask removed in step g) is re-inserted into the training queue.

6. An apparatus for estimating motor improvement of a subject during robotic rehabilitation therapy and consequently adjusting said rehabilitation therapy, said apparatus comprising:

a. a robotic rehabilitation device, in particular a robotic exoskeleton, having at least 6 degrees of freedom, comprising one or more joints which can be actuated by integrated motors; b. sensors for recording signals providing features of motion of said subject;

c. at least one processor controlling said robotic device a) and receiving signals from said sensors b) and executing a rehabilitative motor task in a virtual environment comprising a three-dimensional workspace wherein one or more targets to be reached by the subject are set;

d. means for data storage;

wherein said apparatus is operated by the method of anyone of claims 1-5.

7. The apparatus according to claim 6, wherein said processor c) operates the method of anyone of claims 1-5 to estimate and track in real-time the motor improvement of said subject and consequently adjust in realtime the therapeutic protocol modifying said rehabilitative motor task.

8. The apparatus according to anyone of claims 6-7, wherein said sensors b) are integrated in said robotic rehabilitation device a).

9. The apparatus according to anyone of claims 6-8, wherein said sensors are force and/or position sensors.

10. The apparatus according to anyone of claims 6-9, wherein it further comprises a display wherein said targets are presented to the subject one after another.

11. A computer program for carrying out the method of anyone of claims 1-5.

12. A data medium comprising the computer program of claim 11.

13. A processor or a computer system on which the computer program of claim 11 is loaded.

14. Apparatus of anyone of claims 6-10 for use for restoring motor functions in a subject suffering from neuromotor impairment.

15. The apparatus according to the use of claim 14, wherein the apparatus is for use for the recovery of reaching and grasping abilities in one or both upper limbs in a subject suffering from neuromotor impairment.

16. The apparatus according to the use of claim 14, wherein the apparatus is for use for the recovery of locomotor functions in one or both lower limbs in a subject suffering from neuromotor impairment.

17. The apparatus according to the use of anyone of claims 14-16, wherein said neuromotor impairment is consequent to a stroke, a spinal cord injury, a neurodegenerative disease or a neurological disease.