Measurement 42 (2009) 1176–1187

Contents lists available at ScienceDirect

Measurement journal homepage: www.elsevier.com/locate/measurement

Measuring functional recovery of hemiparetic subjects during gentle robot therapy Maura Casadio a, Pietro Morasso a,*, Alessandro Noriaki Ide a, Vittorio Sanguineti a, Psiche Giannoni b a b

Neurolab, Department of Informatics, Systems and Telematics (DIST), University of Genova, Via Opera Pia 13, 16145 Genova, Italy ART Education and Rehabilitation Center, Genova, Italy

a r t i c l e

i n f o

Article history: Received 22 October 2007 Received in revised form 28 April 2008 Accepted 23 September 2008 Available online 1 October 2008

Keywords: Robot therapy Stroke patients Neurorehabilitation Measurement of performance Haptic interfaces

a b s t r a c t This paper presents a pilot, proof-of-concept study of robot arm therapy (RT) with a treatment protocol specifically designed for severe hemiparetic patients and integrated with a suitable performance measurement protocol. The robot is a planar haptic manipulandum, with low inertia, low friction and impedance control. The task is reaching, with targets arranged in the horizontal plane, in such a way to induce full extension of the arm. Targets are represented haptically, by means of an attractive force field applied by the manipulandum, and visually, by means of circles on a computer screen. The force field is smoothly activated until it reaches a preset intensity that is maintained until the target is reached. Such level of assistance is selected initially as the minimum level that allows each patient to fulfill the task. In each training session, two blocks of trials are alternated (with open and closed eyes, respectively). In the course of training, the level of assistance is reduced as performance improves. Functional recovery is evaluated by processing the kinematic measurements in order to express in quantitative terms the smoothness of the targeting movements. In particular, we defined four performance indicators or outcome measures: (1) the mean speed of the movements; (2) the number of sub-movements in which reaching is decomposed, (3) the remaining error after the first sub-movement, (4) the relative time of the first sub-movement with respect to the total reaching time. For these indicators we identified a measurement scale from the performance of a population of normal subjects performing the same task. The statistical analysis of the responses shows that the proposed protocol is capable to induce significant improvements in all the patients and the performance indicators are sufficiently stable to be chosen as candidates of future adaptive RT protocols in which the training and measurement protocols are designed in an integrated way. Ó 2008 Elsevier Ltd. All rights reserved.

1. Introduction Although the application of robot technologies to the rehabilitation of neurological patients is more than a decade old [1] and since then many robot systems for rehabilitation have been proposed, the number of clinical studies is still limited, as documented in a recent systematic re* Corresponding author. Tel.: +39 010 353 2749; fax: +39 010 353 2154. E-mail address: [email protected] (P. Morasso). 0263-2241/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.measurement.2008.09.012

view [2]. Furthermore, there is not yet a consensus on the critical features that must characterize protocols of robot therapy (RT) and the measurement schemes for evaluating functional recovery in terms of motor performance. With regards to the assessment of motor recovery after stroke, a number of clinical scales have been proposed. The most quoted one is the Fugl–Meyer (FM) scale [3,4]. As all clinical scales, based on human judgment, it is prone to subjective bias and is characterized by floor and ceiling effects, considering the wide variability of clinical conditions

M. Casadio et al. / Measurement 42 (2009) 1176–1187

even for subjects with a well diagnosed pathology like stroke-determined hemiplegia. Moreover, it cannot be used ‘‘on-line” in order to adaptively modify a training protocol on the basis of performance. Nevertheless, the FM scale as well as other scales that focus on different aspects of the motor disability, such as muscle spasticity [5], is appropriate for an initial characterization of the population of patients and for defining inclusion/exclusion criteria in clinical studies. With regards to the design of rational protocols of RT, a working hypothesis that we assumed in this study is that RT and PT (human physiotherapy) should complement each other. A natural meeting point between PT and RT relates to the very nature of haptic interaction between patient and therapist, in which force and energy must flow bi-directionally in an ordered and ‘‘informative” way. Accordingly, haptic interaction between patient and therapist must not be invasive and unidirectional, imposing movements in a passive way, but, rather, should aim at exploiting the innate plasticity of the neural system, present also in chronic, adult patients [6]: in other words, RT should be gentle in such a way to stimulate sensorimotor learning by means of a minimal degree of intervention. This general concept allows to identify, in the large class of robots that have been used in the biomedical field, the basic features that must characterize an ideal neuro-rehabilitation robot: it must be truly haptic, back-drivable, with very small inertia and friction. In contrast with PT, RT allows to integrate in the rehabilitation protocol a performance measurement protocol that allows making available, in an automatic way, an evaluation of the efficacy of the treatment. Here we report a pilot clinical study, in which chronic hemiparetic patients are treated with a protocol of RT, based on a reaching tasks and a highly compliant robot manipulandum [7]. The rehabilitation protocol is minimally assistive, in the sense that the forces provided by the robot are adapted to the specific patient and are kept to a minimum value; moreover, this value is reduced over training in an adaptive way. The type of assistance provided by the robot is instrumental for allowing the patients to initiate the reaching movements but in no way imposes the trajectory, the reaching time, and the speed profile. In this framework, the emergence of features that are typical of normal reaching movements (approximately straight paths and bell-shaped velocity peaks [8]) would suggest that the patients have recovered their ability to generate coherent active patterns and, at the same time, have reduced the hypertonus and have improved their perception of the paretic limb. Therefore, we also designed a performance measurement protocol which aimed at making explicit the ‘‘distance” of the motion patterns of the patients from the motion patterns of a reference population of normal subjects, taken as the standard. The measurement protocol focuses on four performance indicators or outcome measures, that characterize complementary aspects of the reaching movements analysed in this rehabilitation protocol: (1) the mean speed of the movements; (2) the number of sub-movements in which reaching is decomposed, (3) the remaining error after the first sub-movement, (4) the relative time of the first sub-movement with respect to the total reaching

1177

time. The experimental protocol was designed in such a way to satisfy two constraints: (1) the reference population of normal subjects could solve the task with a single movement, a null error at the end of the fist movement, and a stable mean speed; and (2) the population of patients was unable to solve the task without robot assistance at the beginning of the training. Thus, the measurement protocol has a clearly identified measurement scale and measurement range. The level of robot assistance was selected, different for each patient, as the minimum level that allowed the patient to solve the task. Moreover, this level was progressively decreased, in steps, as performance increased. With regards to the uncertainty of the measurements of each patient’s performance, we should take into account that since the measurement protocol is integrated in the rehabilitation protocol, the measurand is actually modified by the underlying learning process, thus making repeated measurements somehow dependent on each other. However, learning is unlikely to progress in a linear way and thus we can assume that the effects of learning are relevant in inter-session fashion (the training protocol was characterized by one session per week) but can be neglected intrasession (each session lasted about 1 h). In summary, the reliability of the measurements and the influence of the training protocol on functional recovery were evaluated at the same time by means of an analysis of variance of the measurements (type A uncertainty), with the simplifying assumption (intended to reduce type B uncertainty) that learning occurs inter-session, not intra-session. With respect to the previously quoted systematic review of the effect of robot-aided therapy on recovery of the hemiparetic arm after stroke [2], our study is mainly aimed at severely affected stroke patients, for which only an ‘‘active-assisted protocol” is appropriate. The novelty of our approach was to design a ‘‘gentle” interaction scheme that exploits the compliant features of the robot in order to allow the patients to be trained in two conditions, namely with or without vision: in the latter condition the role of the proprioceptive channel in the recovery of active motor patterns is greatly enhanced. Research in this field is relatively young and few controlled clinical trials have been conducted. Therefore, the factors that might affect the outcome of robot-aided therapy and bias current research findings are still unclear. The present article presents a proof-of-concept study with regards to the feasibility of the gentle, proprioceptive-enhanced robot therapy approach.

2. Materials and methods 2.1. Subjects Ten hemiparetic subjects (3 males, 7 females, age 53 ± 13 years) participated in this study. Subjects were recruited among those followed as outpatients of the ART

1178

M. Casadio et al. / Measurement 42 (2009) 1176–1187

Rehabilitation and Educational Center in Genova, associated with IBITA (International Association of Bobath Instructors). When subjects entered the robot therapy protocol, they were attending regular rehabilitation sessions (on average twice per week) since at least 6 months. The inclusion criteria were: chronic conditions (at least 1 year after stroke), stable clinical conditions for at least 1 month before entering robot therapy. The exclusion criteria were inability to understand instructions about the exercise protocol and other neurocognitive problems. Preference was given to patients with a high degree of motor impairment. Time since onset was 46.3 ± 41.5 months (range 12–154), with a majority of ischemic etiology (7/10). Subject 3 quit after the 6th session. Patient impairment level was evaluated by means of the Fugl–Meyer score, limited to the arm section (FMA) [3,4]. Five subjects had a severe impairment (FMA < 10/66); 3 patients had an intermediate impairment level (10 < FMA < 20); 2 patients had a mild impairment (FMA > 20). The average FMA score was 15 ± 13 (range 5–41). The average Ashworth score of muscle spasticity [5] was 2.1 ± 0.8 (range 1–3). Table 1 reports the demographic data for all the subjects. The research conforms to the ethical standards laid down in the 1964 Declaration of Helsinki, which protects research subjects, and to ethical bylaws of IBITA. Each subject signed a consent form that conforms to these guidelines. We also carried out a preliminary test of the assistance protocol with four control subjects, aged 29.8 ± 2.5 years, in order to obtain a reference data set for the performance measurement procedure (Section 2.4). The previously quoted systematic review of the effect of robot-aided therapy on recovery of the hemiparetic arm after stroke [2] conducted a systematic search of articles from 1975 to 2005 in the PubMed, Cochrane Controlled Trials and other relevant clinical databases. Seventeen clinical trials were identified but only eight of them were retained according to criteria of methodological quality, based on the Maastricht–Amsterdam criteria for clinical trials [9]. The present study agrees with such criteria and the size of our experimental group (10) falls inside the range of the studies selected as eligible by Prange et al. [2]: 3–42.

2.2. Experimental apparatus The robot – Braccio di Ferro (BdF) is a planar manipulandum with 2 degrees of freedom, which is fully described elsewhere [7]. Its most relevant features are: (i) large planar workspace (80  40 cm ellipse); (ii) very rigid parallelogram structure with direct drive of two brushless motors; (iii) low intrinsic mechanical impedance at the end-effector (inertia <1 kg; neglectable viscosity and friction); (iv) good isotropy (manipulability index = 0.23 ± 0.02; force/ torque ratio = 2.21 ± 0.19 N/Nm); (v) large available force at the handle (continuous force >50 N; peak force >200 N); (vi) impedance control scheme, with a 16-kHz sampling rate of the current loop, a 1-kHz rate of the impedance control loop, and a 100-Hz rate of the virtual reality, data logging loop. The subjects sat in a chair, with their chest and wrist restrained by means of suitable holders, and grasped the BdF handle. A light, soft support was connected to the forearm that allowed low-friction sliding on the horizontal surface of a table. Therefore, only shoulder and elbow could move and motion was restricted to the horizontal plane, with no influence of gravity. The height of the seat was adjusted so that the arm was kept approximately horizontal. The position of the seat was also adjusted in such a way that the farthest targets (see the next section) could only be reached with an almost extended arm. A 1900 LCD screen was positioned right in front of the patients at a distance of about 1 m in order to display the positions of hand and target (see below) by means of circles of different colors, with a diameter of 2 cm. The visual scale factor was 1:1. 2.3. Training protocol The training protocol was specifically focused on facilitating active execution of outward movements because most patients were unable to actively perform such movements, whereas had little difficulty with inward movements, The task consists of reaching a set of targets, arranged in the horizontal plane (Fig. 1) according to three

Table 1 Demographic and clinical data of subjects S

Age (years)

To (months)

FMA (0–66)

ASH (0–4)

Sex (M/F)

E (I/H)

PH (R/L)

Finit (N)

Ffin (N)

S1 S2 S3 S4 S5 S6 S7 S8 S9 S10

72 59 53 69 57 34 30 46 53 55

28 39 154 25 40 24 12 26 39 76

6 5 8 12 17 13 6 6 41 36

3 3 3 1+ 3 1+ 2 2 1 1

M F m f m f f f f f

I I I I I I I H H H

L R L R L R L R R L

25 22 18 15 12 12 9 9 5 5

10 9 12 4 3 3 3 3 0 0

Mean SD

52.8 13.4

46.3 41.5

15.0 13.0

2.1 0.8

13.2 6.8

4.7 4.1

S, subjects; To, time since onset of the disease (months) at the beginning of robot therapy; FMA, arm portion of Fugl–Meyer score (0–66), at the beginning of robot therapy; ASH, Ashworth scale of muscle spasticity (0–4), at the beginning of robot therapy; E, etiology, which is either ischemic (I) or hemorrhagic (H); PH, paretic hand (Right/Left); Finit (N), initial assistance level; Ffin(N), final assistance level.

M. Casadio et al. / Measurement 42 (2009) 1176–1187

1179

Fig. 1. A view from above of a subject holding the handle of the BdF haptic robot. The subject’s shoulders are strapped to a chair; the forearm is attached to a sliding support; the wrist is stabilized by means of a skateboard wrist brace and the hand grasp by means of a Velcro holder. The targets are arranged on three shells: A, B and C. The C shell is just in front of a virtual wall. The basic sequence of target activation is A ? C ? B ? A and it is repeated 3  7  3 = 63 times in random order.

shells: inner (A, 3 targets), middle (B, 3 targets), and outer (C, 7 targets). Reaching the outer targets requires almost full extension of the arm. The distance between adjacent shells is 10 cm; the distance between targets on the same shell is, respectively, 6.26 cm (shell A), 8.77 cm (shell B), 5.65 cm (shell C). Targets are circular, with a 2 cm diameter, chosen in such a way that control subjects can reach them in one shot. Target sequences are generated according to the following scheme: A ? C ? B ? A. Outward movements are performed in one step (A ? C), whereas inward movements are performed into two steps (C ? B and B ? A). In this way, the task emphasizes the training of wide, outward movements with respect to the return to the initial flexed posture. After a patient reaches a target there is a 1-s delay before the activation of the next one. When a target is presented to the patient, the robot generates an assistive force field, i.e. a force vector directed from the current position of the hand xH to the target, xT. In order to avoid an unpleasant, sudden interaction between the robot and the patient, the assistive force is not activated abruptly but smoothly, with a ramp-and-hold profile, R(t) (rise time: 1 s), up to a force value FA. For each patient, this force is set by the therapist as the minimum value that evokes a functional response. The force is switched off as soon as the patient hits the target. The next target is presented after a pause of 1 s. In addition to the assistive force field, in order to improve the interaction between patients and robot we added a mild viscous force field (to damp occasional hand oscillations without significantly affecting the voluntary reaching patterns) and a virtual ‘‘wall” to prevent going be-

yond the C shell of targets and to provide an additional feedback about the successful achievement of the outward target. The force field generation law is expressed by Eq. (1):

FðtÞ ¼ F A

ð x T  xH Þ  RðtÞ  Bx_ H  K W ðxH  xW Þ j x T  xH j

ð1Þ

where xT is the vector that identifies the target position in the plane, xH and x_ H are, respectively, the hand position and velocity vectors; B is the viscous coefficient, set to 12 N s/m, and KW is the stiffness coefficient of the wall, set to 1000 N/m; hxH  xW i indicates the degree of ‘penetration’ of the hand inside the wall, and is zero outside the wall. It should be noted that such a control scheme does not impose any constraints on either the timing of the reaching movement and/or the trajectory that patients choose to follow in order to reach the target, except for the occasional ‘sliding’ movements along the virtual wall. This was done on purpose, together with the choice of keeping as low as possible the level of robotic assistance, in order to be sure that the observed responses were mainly driven by active patterns, not by robot actions. In inward movements (C ? B and B ? A) the force field could have been turned off for most of the patients, even in the early phase of training, but we chose to keep it running to maintain the protocol as uniform as possible, so that they could focus their attention on the task and progressively release the hypertonus that is associated with the pathological flexion pattern. The pause of 1 s in the activation profile of the force field was introduced on purpose: a normal subject is likely

1180

M. Casadio et al. / Measurement 42 (2009) 1176–1187

to stay on target after the force field is turned off, whereas we may expect the patients to perform a ‘‘rebound movement”, with a velocity peak that is higher, the higher the hypertonus of the patient. This gives us an oppurtunity to detect if the hypertonus of the patients goes up or down as a side effect of the treatment, by looking at the velocity peak of the rebound movement. One block of trials included repetitions of the A ? C ? B ? A sequence with different targets in random order, for a total of 3  3  7 = 63 movements; 21 of them were large amplitude, outward movements and 42 movements had smaller amplitude and were directed inward. The protocol started with a test phase, in which individual patients familiarized with the apparatus and with the range of assistive forces. This phase was supervised by a physical therapist, who observed the response to the different force levels and selected the minimum level Ftest capable to induce at least a hint of active response in the direction of the target. Each block of trials could be performed with or without vision. In the latter case, the patients were blindfolded but could still feel the target direction through proprioception, by means of the assistive force and the virtual wall. In other words, target representation was both visual and haptic. In trials without vision the patients could only rely on haptic targets and thus were forced to focus their attention on the proprioceptive channel. The subjects knew which targets were to be reached from the robot-generated force field itself because, for each time instant, the force applied by the handle to the hand of the subject was directed toward the target. They also knew when the target was reached by the fact that the force was suddenly switched off and a sound was activated. The first training session initiated with two blocks of trials (vision, no-vision), using the same level of force determined in the test session (F1), for a total of 126 movements. After a little rest, the therapist considered the level of performance and asked the patient about fatigue. The decision could be (1) to terminate the session, (2) to continue with the same force level, (3) to continue with a reduced force, F2 (10–20% less than F1). The procedure was iterated until the decision to stop was agreed by the patient and the therapist. In the second session, two blocks of trials were carried out with F1, two blocks with F2 (if indeed the subject worked with the reduced assistance) and then, if possible, the level of assistance was decreased to F3. In following sessions the training always started with two blocks of F1, two blocks of F2 ... down to the latest assistance level. If patients reached a level of assistance with a force below 4 N, the no-vision blocks were eliminated because that force level is very close to the perceptual threshold. The robot training protocol consisted of 10 sessions (1 session/week), plus the initial test session. During that time, all patients continued their physiotherapy program with the same schedule and the same routine. The robot training sessions were carried out at the Neurolab of the Department of Informatics, Systems, and Telematics of the University of Genova, under the supervision of a physical therapist. The physiotherapy sessions were carried out at the ART premises.

2.4. Data analysis and performance measurement protocol Hand trajectories and the forces generated by the robot were recorded at a 100 Hz sampling rate. Hand position was measured from the 17-bit encoders of the motors with a precision better than 0.1 mm in the whole workspace. Hand speed was estimated by using a 4th order Savitzky–Golay smoothing filter (with an equivalent cut-off frequency of 6 Hz). The analysis was limited to the outward movements (A ? C) because these are more important from the rehabilitation point of view. From the raw kinematic measurements, we designed a performance measurement scheme aimed at making explicit the evolution of the recovery process. Motor performance is a complex concept that cannot be reduced to a single quantity or measurand. We chose to focus on four complementary aspects that characterize ‘‘normal” reaching movements: (1) the average speed of the movement, (2) the number of sub-movements of the reaching pattern, (3) the error at the end of the first sub-movement, (4) the relative duration of the first submovement with respect to the total reaching time. According to the so called Fitts’ law [10], the duration of reaching movements is a logarithmic function of the required accuracy, defined as the ratio between movement amplitude and target size: for large accuracy values, the trajectory tends to be broken down into a sequence of sub-movements of decreasing amplitude, whereas for sufficiently small values (accuracy of the order of 10/1 or less) the movement is characterized by a bell-shaped speed profile [8] with a rather constant duration. We chose a target size (2 cm in diameter) that appears to satisfy the unimodality constraint of the speed profile. The first step of the measurement protocol was to obtain a reference standard. On this purpose, in a preliminary experimental session, the normal subjects (3 females and 2 males, age 25.2 ± 3.5 years) were exposed to the same experimental environment and the same protocol of the patients, but without robot assistance. The analysis of the responses shows that, in the defined experimental environment, the average reaching speed of normal subjects is Vref = 19.29 ± 4.92 cm/s. We also verified that, for the designed experimental setup, the following conditions hold (1) the reaching movements of all the control subjects were characterized by a single velocity peak; (2) the normal error at the end of the first sub-movement was null; and (3) the duration of the first sub-movement coincided with the total reaching time. The performance measurement scales were then defined in relation with the standards above. We chose to keep four different indicators instead of a single one because they represent different aspects of the performance, possibly with different patterns in different patients. Such indicators express, in different ways, the degree of smoothness of the movements. This choice is motivated by the fact that studies on the recovery from neural injury [11] suggest that smoothness is a result of a learned, coordinative process rather than a natural consequence of the structure of the neuromuscular system. The four indicators are defined as follows:

M. Casadio et al. / Measurement 42 (2009) 1176–1187

 Mean speed indicator: V (in cm/s), which is computed from the time of target presentation to the time at which the subject reaches the target. The range is between 0 (when the subject is unable to reach the target) to the normal performance Vref = 19.29 ± 4.92 cm/s.  Number of sub-movements indicator: N, which is the number of peaks in the speed profile. In order to eliminate spurious velocity peaks at the output of the Savitzky–Golay smoothing filter, we used the two following criteria: (1) a threshold on the speed (0.01 m/s), (2) a threshold on the time interval between one peak and the next one (0.3 s). The range of the indicator is between the reference value Vref = 1 and 1 (when the patient is unable to reach the target). In practice, even the most severely impaired patient in our population does not exceed a value of N = 15–16.  Endpoint error after the first sub-movement: E (in cm), which is the distance between the hand position and the position of the target at the end of the first submovement. It ranges between Eref = 0 and the maximum distance from the initial position of the hand to the designated target (26 cm).  T ratio: TR, which is defined as the ratio between the duration of the first sub-movement (identified by two consecutive points of minimum in the speed profile, one before and the other after the point of peak) and the total reaching time, from the instant of target onset to the instant in which the hand enters the target area. The range is between 0 (when the subject is unable to reach the target) and the reference value TRref = 1. With regards to the uncertainty of each indicator, we considered, for each session, the standard deviation of the mean and the corresponding variation coefficient, for the level of assistance initially selected for each patient, with the underlying assumption that the mean level of performance remained constant within each session but changed intra-session. 2.5. Evaluation of the efficacy of the treatment In order to evaluate the efficacy of RT, we performed a statistical analysis of the evolution of the performance indicators throughout sessions. In particular, we performed a 2-way, repeated-measures ANOVA, with two factors (session: 1–10; vision: yes/no), for the assistance force level that was present in all sessions and that, according to the protocol, was administered at the beginning of each session. To assess whether the session effects actually corresponded to an increase of performance over sessions, for each indicator we ran a contrast analysis (linear contrast) of the session effect and of the session  vision interaction. To quantify the magnitude of the change in a more direct way, for each indicator we also compared the performance (at the initial minimum force level) in the first and last session. The analysis above evaluates session and condition effects for a fixed level of robot assistance. However, a more complete picture of training performance can only be obtained by considering all the different levels of force experienced by each subject. This cannot be done with standard

1181

ANOVA, because the number of force levels as well as and their magnitudes are not the same for each subject. Rather, we used a mixed-effect model, with three fixed-effect factors (session, vision and force) and two additional random factors (subject and target), to properly account for the correlations among repeated measures from the same subject. The deterministic part of the model consists of a linear regression:

indicator ¼ b0 þ b1  ðsession  1Þ þ b2  force þ b3  vision

ð2Þ

where ‘‘session” is an integer (from 1 to 10), ‘‘force” is the field intensity (in Newton), and ‘‘vision” is a binary variable that designates whether the exercise was carried out with open or close eyes. Model coefficients may be interpreted as follows: (1) b0 = ‘‘baseline performance level”, i.e. the theoretical performance at the initial session with zero assistive force; (2) b1 = ‘‘session by session improvement rate”; (3) b2 = ‘‘rate of improvement related to the assistance level”; (4) b3 = ‘‘bias introduced by presence of vision”. 3. Results As shown in Table 1 (penultimate column), the initial level of assistance selected for the subjects ranged between 25 and 5 N and it was generally higher for patients with lower initial levels of the FMA score, although the correlation between the two variables was not very high (0.68). The outward reaching movements (for the A ? C targets) could be segmented into a large number of submovements (up to 9) and, as a consequence, the reaching time was quite high (up to 1 min). In spite of this, the patients did reach the targets and they felt that, to a great extent, this achievement was due to their effort. This was a great motivation to proceed in the training and gradually improve the smoothness and the effectiveness of the control patterns. In contrast, inward movements were usually characterized by a single, higher peak in the speed profile because they are ‘‘favoured” by the pathological syndrome itself. One of the goals of the training was indeed to decrease the asymmetry between inward and outward movements. Fig. 2 displays the modification of the hand speed profile during the 10 training sessions for a typical subject and for the initially chosen assistance level: in the early training sessions the first peak is followed by a number of smaller peaks; after the 6th training session, the subject can reach the target with a single movement, characterized by a higher speed peak. Moreover, the height of the dominant peak becomes higher and higher as training proceeds. In spite of the segmented nature of the reaching movements, the individual sub-movements exhibit the typical bell-shaped velocity profile that characterizes normal reaching movements. This is important for two reasons: (1) it makes us confident that, in spite of the severity of the impairment, the subjects have an underlying potential of recovery because the basic control patterns are still there, though ‘hidden’ by the dominating pathological scheme; (2) it assures us that for the selected level of assistance the observed movements are mainly determined by

1182

M. Casadio et al. / Measurement 42 (2009) 1176–1187

Fig. 2. Session by session modifications of the speed profile of movements to a single target. Subject 7, force level 6 N, open eyes condition. Please note that recording is stopped when the subject hits the target and this occurs at higher and higher hand speed as the training proceeds.

active motor patterns and do not merely reflect passive responses to the robot action. Moreover, Table 1 (last column) shows that at the end of training the required assistance level is more than halved. The four performance indicators defined in the previous section were estimated for all the trajectories in all the experimental blocks, and identified by session number and force level. We first analyzed the evolution of the performance indicators for the level of assistance that was present in all sessions and in both situations (vision and no-vision): mean speed (Fig. 3), number of sub-movements (Fig. 4), end-point error (Fig. 5), T ratio (Fig. 6). Black and grey curves correspond to vision and no-vision blocks, respectively. For all the patients a general improvement of performance is apparent. With regards to the mean speed indicator, Fig. 3 shows that for all of the patients there is a positive trend, confirmed by the a 2-way, repeated-measures ANOVA with two factors (session and vision). The session effect is strongly significant (p < 0.001); moreover, the contrast analysis of the this factor confirms that there is a highly significant linear trend (F(1, 8) = 15.47, p = 0.004), which almost doubles the mean speed: from 5.61 ± 0.94 cm/s (mean ± standard error) at session 1, to 10.00 ± 0.78 cm/s at session 10 (F(1, 8) = 16.78, p < 0.003). With regards to vision, Fig. 3 shows that different subjects exhibit different behaviours: (a) in the majority of subjects (S2, S5, S6, S7 and S8) there is no clear difference between the open-eyes and closed-eyes cases; (b) in some subjects (S1, S3 and S4) there is a tendency for the no-vision paradigm to prevail;

(c) in other subjects (S9 and S10) the vision case is clearly better. Therefore, it is not surprising that, at the population level, no vision effect or vision  session interaction appears to be statistically significant. However, what is significant in terms of motor control theory is that in most subjects (8 out of 10) the deprivation of an importance sensory channel like vision does not necessarily imply a deterioration of the reaching performance: this further supports the working hypothesis that an experimental design that emphasizes the role of the proprioceptive channel has a relevant rehabilitation value. A similar picture emerges from the statistical analysis of the other perfomance indicators. The degree of movement segmentation, indicated by the number of sub-movements, steadily decreases from initial values as high as 9 to values that are quite close to the 1 (i.e., the normal value). Contrast analysis of the session factor confirms that there is a highly significant linear trend (F(1, 8) = 16.93, p = 0.003), with a decrease of the number of sub-movements from 5.71 ± 1.04 at session 1, to 1.69 ± 0.21 at session 10 (F(1, 8) = 16.01, p = 0.004). The endpoint error after the first sub-movement steadily decreases, approaching the nominal value of 0. Statistical analysis again shows a highly significant trend of the session factor (F(1, 8) = 14.29, p = 0.005), with a statistically significant improvement from 5.05 ± 1.22 cm at session 1, to 3.06 ± 0.33 cm at session 10 (F(1, 8) = 16.98, p = 0.003). Likewise, over sessions the T ratio approaches a value quite close to the nominal value of 100%. The session factor exhibits a significant trend (F(1, 8) = 19.96, p = 0.002), with

M. Casadio et al. / Measurement 42 (2009) 1176–1187

1183

Fig. 3. Each panel shows the evolution, over the training sessions, of the mean speed of the outward reaching movements for the different subjects, at a level of assistance selected in the initial session. Subject 3 quit after the 6th session; Subjects 8 and 9 initiate no-vision blocks in the second session. The reference value (average speed of the normal population) is Vref = 19.29 ± 4.92 cm/s.

a statistically significant improvement from 47.06 ± 8.89% at session 1, to 83.77 ± 5.44% at session 10 (F(1, 8) = 29.28, p = 0.0006). All together, the analysis of the performance indicators shows a linear improvement trend, without any hint of saturation, thus supporting the choice of a non-monotonic profile of reduction of the assistance level: the linear trend means that the patients remain well responsive, throughout the training protocol, to the patterns of robotic assistance. With regards to the uncertainty of the performance measurements we considered the variation coefficient of each indicator. As could be expected it improves with training: it ranges between 16.7% and 24.1%, at the beginning of the training protocol, and between 6.49% and 12.4% at the end. Of the four indicators, the first and the fourth one appear to be the most stable. To account for the joint effects of session and the level of assistive force, we applied the mixed-effects model described in the methods: all indicators appear to improve as a function of both the session and the level of assistance. This is confirmed by the linear regression analysis of all the indicators (see Eq. (2)). More specifically, we found highly significant effects of session for the mean speed

(p = 0.0068), the number of sub-movements (p = 0.0011), and the T ratio (p = 0.0027), but not for the endpoint error (p = 0.067). With regards to the level of assistance, all indicators display a significant effect (p < 0.0001, p = 0.0006, p = 0.0009 and p < 0.0001 for, respectively, speed, number of peaks, endpoint error and T ratio). The b1 and b2 coefficients are positive for the two indicators (mean speed, T ratio) for which the improvement is indicated by an increase of the indicator (0.364 ± 0.097 cm/s per session and 2.98 ± 0.77% per session, respectively), and negative for the two indicators (number of sub-movements and endpoint error) for which improvement is indicated by a decrement (respectively, 0.30 ± 0.09 peaks per session and 0.29 ± 0.08 cm per session). 4. Discussion Before analyzing in some detail the results of this pilot experiment, we address some general points that are relevant from the measurement point of view. First of all, it is worth mentioning that this study was carried out within the scope of TC 18 (Measurement of Human Functions) that shares with TC13 (Measurements in Biology and Medicine) the common interest on the analysis of biological

1184

M. Casadio et al. / Measurement 42 (2009) 1176–1187

Fig. 4. Each panel shows the evolution, over the training sessions, of the number of sub-movements of the outward reaching movements for the different subjects, at a level of assistance selected in the initial session. Subject 3 quit after the 6th session; Subjects 8 and 9 initiate no-vision blocks in the second session. The reference value (average number of sub-movements of the normal population) is Nref = 1.

systems but differs markedly for the nature of the measurements and their purposes. TC18 addresses voluntary movements and goal-oriented behaviour, whereas TC13 focuses on specific biological functions at a lower level of integration; TC13 is mainly intended for diagnostic purposes, whereas TC18 is more oriented to modelling (normal and abnormal) behaviour and rehabilitation. In particular, the purpose of measuring in rehabilitation is to monitor in an objective way the recovery of function and modulate accordingly the treatment in order to optimise it. We also wish to point out that the subject of this article is not a (randomised) controlled clinical trial (RCT) but a proof-of-concept study. RCTs involve the random allocation of different interventions (or treatments) to subjects, in order to ensure that confounding factors are evenly distributed between treatment groups. RCTs are used to establish average efficacy of a treatment as well as learn about its most frequently occurring side-effects, while taking into account a number of concerns: (1) the effects of a treatment may be small and therefore undetectable except when studied systematically on a large population; (2) humans are complex organisms and do not react to the same stimulus in the same way, which makes inference from single clinical reports very unreliable; (3) some conditions

will spontaneously go into remission; (4) the simple process of administering the treatment may have powerful psychological effects on the patient, (placebo effect). The problem is that RCTs are very expensive, particularly in terms of time in the case of rehabilitation that requires long treatment and follow-up phases; therefore it is worth running them only after suitable proof-of-concept studies have achieved enough preliminary evidence as regards efficacy and have provided empirical support for the design of stable experimental protocols. Likewise, statistical analysis has a different meaning in RCTs and proof-of-concept studies: in the former case it is meant to evaluate the absolute or relative efficacy of a given treatment protocol, whereas in the latter case it is supposed to outline a trend for motivating the selection of parameters in the design of treatment protocols, to be validated thereafter on the basis of specific RCTs. Within the framework outlined above, we can say that an important qualitative result of our experimental approach is that the described protocol of gentle robot therapy is well tolerated by the patients, thus addressing a side-effect concern of any treatment approach. When the subjects are put in front of a machine or a robot, the first risk that must be avoided is to ‘‘bother” them by forcing an endless repetition of uninteresting actions or to

M. Casadio et al. / Measurement 42 (2009) 1176–1187

1185

Fig. 5. Each panel shows the evolution, over the training sessions, of the endpoint error after the 1st sub-movement of the outward reaching movements for the different subjects, at a level of assistance selected in the initial session. Subject 3 quit after the 6th session; Subjects 8 and 9 initiate no-vision blocks in the second session. The reference value (average endpoint error of the normal population) is Eref = 0 cm.

‘‘frighten” them with noise, vibrations, and rigid, imposed motions. This was avoided, in particular, because the freedom left to the subjects with regards to time and path introduced enough variability in the exercise to make it interesting. BdF was indeed perceived by our subjects as a benign and patient entity. The protocol also included elements of challenge, for example facing reduced levels of assistance that motivated the repetitions in order to improve performance. Another risk is to induce muscle spasticity, which is velocity- and acceleration-dependent. This was avoided by the criterion of minimal assistance, which kept the movement velocity at relatively low levels, without any sharp acceleration peaks, and by the fact that each session started as a ‘‘warm-up”, using the assistance levels mastered in the previous sessions, before attempting a more challenging task. A related risk is fatigue, both muscular and cognitive, but the experience with this pilot study suggests that the challenge of the task was well tolerated, even by the most severely impaired patients. In quantitative terms, we can say that the results confirm the working hypothesis that robot therapy, based on a minimal level of assistance paradigm, is can improve the reaching movements of chronic post-stroke patients. The improvement of performance, revealed by the de-

scribed indicators, was accompanied by a general regularization of the movements. In particular, we observed a remarkable reduction in the degree of segmentation into a number of sub-movements and the emergence of normal reaching patterns. As training proceeded, kinematic indicators showed a tendency toward movements that were faster, smoother, and with more symmetric velocity profiles. It may be asked to what extent such improvements are also indicative of functional recovery. Studies on the recovery from neural injury [11] have suggested that smoothness is a result of a learned, coordinative process rather than a natural consequence of the structure of the neuromuscular system. Additionally, there is some evidence that the segmented structure of arm movements in stroke patients can be attributed to a deficit of inter-joint coordination [12]. Therefore, smooth movements result from an improved coordination, which is a conditio sine qua non for functional recovery. A peculiar feature of our proposed approach is that assistance is kept at a minimum level, thus avoiding as much as possible the risk that movements are performed in a passive way: the movements must be assisted, not enforced by the robot therapist. Our results suggest that even severely impaired patients benefit from gentle robot therapy, to an extent that

1186

M. Casadio et al. / Measurement 42 (2009) 1176–1187

Fig. 6. Each panel shows the evolution, over the training sessions, of the T ratio (duration of the 1st sub-movement divided by the total movement duration) of the outward reaching movements for the different subjects, at a level of assistance selected in the initial session. Subject 3 quit after the 6th session; Subjects 8 and 9 initiate no-vision blocks in the second session. The reference value (average T ratio of the normal population) is TRref = 100%.

is at least comparable to those with mild to intermediate levels of impairment. This is partly in contrast with other studies, which suggest a greater benefit for mildly impaired subjects [13–15]. Thus, the present article is a contribution to an open debate, about the extent to which physical therapy (either by humans or robot assistants) may really have an impact on the functional recovery of severe, chronic patients. At least, our results provide some evidence of a positive prognostic outcome that motivates further investigation. With mildly impaired subjects, we noticed a kind of ceiling effect: some performance indicator exhibited no significant performance improvement, although the patients subjectively ‘‘felt better” and the physical therapist perceived an improvement of the prognostic horizon. It may be that for these patients the task used in this study is too simple and more complex interaction schemes would have been more appropriate. In a sense, this could be taken as a criterion for adaptively modifying the interaction scheme in order to match the degree of difficulty to the capabilities of each individual patient, while taking always into account that it must be the patient who has to solve the problem: the robot must remain in the background, helping, assisting, evaluating, reinforcing but never forcing.

Our study contradicts the traditional assumption that most recovery occurs within the first 3 to 6 months after stroke with no further improvements later on [16,17] and it agrees with several more recent publications that claim that chronic patients (i.e., more than 6 months post-stroke) can improve upper-limb function [18,19]. In general, we think that RT requires to design, in a coordinated way, the training protocol and the performance measurement protocol. Changing the task, for example from reaching to tracking, would require a redefinition of the both protocols. In other words, we do not think that a universal performance standard can be defined for carrying out on-line adaptive RT protocols. However, in both cases we think that the analysis of performance of normal subjects, in the same experimental framework, is essential for defining a reliable measurement scale. Acknowledgments This work was supported by two Research Projects of National Relevance (PRIN) grants awarded by the Italian Ministry of University and Research to P. Morasso and V. Sanguineti. We thank Ms Liliana Zerbino, PT, for the help

M. Casadio et al. / Measurement 42 (2009) 1176–1187

in the selection of the patients and the evaluation of the FMA score.

[10]

References [11] [1] D. Khalili, M. Zomlefer, An intelligent robotic system for rehabilitation of joints and estimation of body segment parameters, IEEE Trans. Biomed. Eng. 35 (1988) 138–146. [2] G.B. Prange, M.J. Jannink, C.G. Groothuis-Oudshoorn, H.J. Hermens, M.J. Ijzerman, Systematic review of the effect of robot-aided therapy on recovery of the hemiparetic arm after stroke, J. Rehabil. Res. Dev. 43 (2006) 171–184. [3] D.J. Gladstone, C.J. Danells, S.E. Black, The Fugl–Meyer assessment of motor recovery after stroke: a critical review of its measurement properties, Neurorehabil. Neural Repair 16 (2002) 232–240. [4] T. Platz, C. Pinkowski, F. van Wijck, I.H. Kim, P. di Bella, G. Johnson, Reliability and validity of arm function assessment with standardized guidelines for the Fugl–Meyer Test, Action Research Arm Test and Box and Block Test: a multicentre study, Clin. Rehabil. 19 (2005) 404–411. [5] R.W. Bohannon, M.B. Smith, Interrater reliability of a modified Ashworth scale of muscle spasticity, Phys. Ther. 67 (1987) 206– 207. [6] R.J. Nudo, Mechanisms for recovery of motor function following cortical damage, Curr. Opin. Neurobiol. 16 (2006) 638–644. [7] M. Casadio, P.G. Morasso, V. Sanguineti, V. Arrichiello, ‘‘Braccio di Ferro: a new haptic workstation for neuromotor rehabilitation, Technol. Health Care 13 (2006) 1–20. [8] P. Morasso, Spatial control of arm movements, Exp. Brain Res. 42 (1981) 223–227. [9] M.W. Van Tulder, W.J. Assendelft, B.W. Koes, L.M. Bouter, Method guidelines for systematic reviews in the Cochrane

[12] [13]

[14]

[15]

[16]

[17]

[18]

[19]

1187

collaboration back review group for spinal disorders, Spine 22 (1997) 2323–2330. P.M. Fitts, The information capacity of the human motor system in controlling the amplitude of movement, J. Exp. Pshychol. 47 (1954) 381–391. B. Rohrer, S. Fasoli, H.I. Krebs, R. Hughes, B. Volpe, W.R. Frontera, J. Stein, N. Hogan, Movement smoothness changes during stroke recovery, J. Neurosci. 22 (2002) 8297–8304. M.F. Levin, Interjoint coordination during pointing movements is disrupted in spastic hemiparesis, Brain 119 (Pt. 1) (1996) 281–293. R. Colombo, F. Pisano, S. Micera, A. Mazzone, C. Delconte, M.C. Carrozza, P. Dario, G. Minuco, Robotic techniques for upper limb evaluation and rehabilitation of stroke patients, IEEE Trans. Neural Syst. Rehabil. Eng. 13 (2005) 311–324. S.E. Fasoli, H.I. Krebs, J. Stein, W.R. Frontera, N. Hogan, Effects of robotic therapy on motor impairment and recovery in chronic stroke, Arch. Phys. Med. Rehabil. 84 (2003) 477–482. J.L. Patton, M. Kovic, F.A. Mussa-Ivaldi, Custom-designed haptic training for restoring reaching ability to individuals with poststroke hemiparesis, J. Rehabil. Res. Dev. 43 (2006) 643–656. H.S. Jørgenson, H. Nakayama, H.O. Raaschou, J. Vive-Larsen, M. Støier, T.S. Olsen, Outcome and time course of recovery in stroke. Part II: Time course of recovery. The Copenhagen stroke study, Arch. Phys. Med. Rehabil. 76 (1995) 406–412. H.T. Hendricks, J. Van Limbeek, A.C. Geurts, M.J. Zwarts, Motor recovery after stroke: a systematic review of the literature, Arch. Phys. Med. Rehabil. 83 (2002) 1629–1637. J.G. Broeks, G.J. Lankhorst, K. Rumping, A.J. Prevo, The longterm outcome of arm function after stroke: results of a follow-up study, Disabil. Rehabil. 21 (1999) 357–364. S.J. Page, D.R. Gater, P. Bach-Y-Rita, Reconsidering the motor recovery plateau in stroke rehabilitation, Arch. Phys. Med. Rehabil. 85 (2004) 1377–1381.