The study cohort included eight stroke patients (5 women and 3 men) ranging in age from 39 to 60 years old who were right-handed according to the Edinburgh Handedness Inventory (13). They were recruited through the registries of stroke survivors who agreed to be contacted for stroke recovery studies that are maintained at Massachusetts General Hospital. Each had suffered an ischemic stroke affecting the left middle cerebral artery territory at least 6 months prior to recruitment. According to their medical records, they presented with acute unilateral loss of hand strength (Medical Research Council scale score <4, on 0-5 scale in which 5 is normal) that lasted for >48 h. They did not have hearing, vision, language, or cognitive deficits. Institutional review board approval of the study was granted by the Massachusetts General Hospital Human Research Committee (protocol number 2005P000570) and all participants provided written informed consent.

We evaluated a total of 928 game rounds distributed over 386 training sessions. Patients were trained with a robotic hand rehabilitation system coupled to an interactive game for 45 min per day, 3 days per week, over a 10-week period. Each training session consisted of four 8-minute-long scenarios separated by 1- to 5-minute rest breaks. One session was conducted on each of three training days per week. Training was conducted at the patients' homes under supervision to ensure compliance. Motor performance was assessed prior to rehabilitation training (Pre), after finishing the treatment period (Post), and again 1 month after finishing the treatment in a follow-up assessment (FU) intended to probe the persistence of motor benefits over time. Motor performance was assessed via four clinical motor scales: The Fugl-Meyer scale, which we used to assess sensorimotor impairment (includes upper extremity, wrist, hand, coordination, sensation, passive joint motion, joint pain, and total scores) (14); the Action Research Arm Test (ARAT), which includes grasp, grip, pinch, gross movement, and total scores (15); the Modified Ashworth scale to assess spasticity of the elbow, wrist, fingers, and thumb (16); and the Box and Blocks test of gross manual dexterity (17).

Patients underwent training with a newly re-designed robotic hand rehabilitation device called the Magnetic Resonance compatible Hand-Induced RObotic Device, version 3 (MR_CHIROD v3) (18-20). It was engineered to provide adjustable levels of force for handgrip exercising (Fig. 1). Continuous, grip-opposing, restoring force was provided by a low-friction glass cylinder/graphite piston assembly under the control of an electronic pneumatic pressure regulator with a portable air compressor that allowed the MR_CHIROD v3 to be used in a home environment. The MR_CHIROD v3 device was linked to a low-cost, Arduino-compatible microcontroller via a simple USB serial interface (57.6-kbps data transfer rate) thereby allowing pressure control, force/displacement acquisition, and interface with a laptop computer.

A serious game was developed in Microsoft Visual C++ for use on a Microsoft Surface (Microsoft Corp., Redmond, WA). The user plays the game via an avatar, namely a small green alien in a flying saucer that travels through a linear scrolling labyrinth (right to left scrolling). The speed of the saucer starts slow and increases linearly to a final level for each stage. The saucer flies through a scene, earning points for collision-free flight. The scene is populated with obstacles and rewards, collisions with which contribute negatively and positively to the user's score, respectively. During the game, the patient uses the MR_CHIROD handle to control the altitude of the saucer such that the handle position controls saucer height on the screen directly, with the fully open and fully closed positions corresponding to the lowest and highest screen positions. The goal of the game is to maximize one's overall score by avoiding obstacles and collecting rewards (Fig. 2). To increase user involvement, a soundtrack is looped and several sound effects are played upon collision with obstacles and rewards.

The game scenario was parametrized through an Extensible Markup Language (XML) script (‘stage script’) that specifies what the game engine loads in the form of sprites (graphic elements such as the avatar, obstacles, and rewards), the background picture, and game information. The stage script specifies obstacle and reward point values as well as all game components and mechanisms, including speed, sounds, and supportive visuals. Obstacles can be avoided with large, low accuracy motions to reach a safe position, whereas reward contact requires fine position control within a small range of motion. However, for a greater challenge in high-level play, users face more closely-spaced obstacles that leave narrow passages that require fine control to negotiate. The combination of obstacles and rewards allows the design of precise and complex trajectories for training gross motor and fine grip motions.

Prior to the commencement of each stage of the game, the prescribed pressure command is sent to the MR_CHIROD v3. When the game starts, the game engine polls the MR_CHIROD interface to retrieve the handle position in encoder counts and normalizes the position based on the range of motion set by end-stops on the device appropriate to each patient, and then updates the vertical and horizontal position coordinates of the avatar.

Game scores are calculated continuously based on the following formula:

where the instantaneous score S(t) is equal to the previous score S(t - 1) plus any reward collected, R(t), minus any penalty for hitting an obstacle, O(t), plus the constant g, which is the incremental score for each time interval. The scoring algorithm was tuned by adjusting R, O, and g in the stage script. The game was programmed to save each user's position trajectory, events at each time interval, score, handle pressure, and timestamp into a secondary XML script, which we used for offline analysis of raw and processed information. The stage script was individualized such that the handle force setting was set to ~75% of the user's maximal grip strength, thereby requiring our patients to exert effort and generate correspondingly large motor cortex activation, while not being so high as to impede completion of the training sessions.

For the purposes of the present study analyses, motor impairment was classified according to a three-level Fugl-Meyer upper extremity (FM UE) scale scheme (21), in which an FM UE score <20 was classified as severe, a score in the range of 20-40 was classified as moderate, and a score >40 was classified as mild. Improvement of at least 4.25 on the FM UE scale (22) and/or a smallest real difference of at least 5.5 blocks/min on the Box and Blocks test (23) were considered signs of clinically meaningful improvement in grip ability.

The mean force and the Pearson correlation coefficient between handle position and force (an indirect measure of grip control) were calculated for each round. These device metrics and the game metrics (final score, collisions, and rewards) for each of four scenarios were averaged for each session. Associations of device and game metrics with clinical motor scale scores were assessed with the Spearman ρ correlation coefficient after Bonferroni correction of multiple comparisons. Stepwise linear regression analysis was used to assess which metrics were significant predictors of motor performance, as indexed by FM UE, ARAT grip, and Box and Blocks scores. All statistical analyses were performed in SPSS version 23.0 (IBM, Inc.). Analyses with two-tailed P-values <0.05 were regarded as statistically significant.

Notably, the MR_CHIROD v3 system was designed with features that facilitate its use and dissemination. The system's USB serial connectivity and available WiFi compatibility allow it to be used with multiple commercial operating systems and game consoles. The master-side of the polling-based serial communication protocol can also be implemented in different programs. To enable ease of replication and wider use of the system, it has multiple 3D-printable and uncomplicated machinable components as well as low-cost electronic and pneumatic components.

The strong magnetic fields characteristic of the modern MR machines limit the range of actuators, sensors, and materials that can be used in an MR-compatible rehabilitation devices. Ferrous materials, electromagnetic actuators, and unshielded conductors maybe forbidden for use in MR scanners altogether for safety reasons or poorly suited due to imaging data distortion. The MR-compatibility of the MR_CHIROD v3 robotic device-made possible by plastic parts (3D-printed and machined), shielded sensors/cables, and MR-compatible ball bearings and other components-mean that patients brain activation can be analyzed during its use. Generally, upper-extremity rehabilitation devices are not made to MR-compatible (35); for exceptions see (36). In addition to being MR-compatible, the currently used MR_CHIROD version 3 (manuscript describing the design and testing is in preparation) is portable and suitable for standardized use across clinical and home environments (18). Whereas the high cost of fabricating our prior design of the MR_CHIROD device (version 2) was limiting (19), the version 3 redesign yielded fabrication cost reduction as well as overall simplification of the fabrication process.

Finally, it is our view that the combination of robotic technologies with serious gaming can be applied as a rehabilitation tool as well as in multiparametric modeling of motor performance and/or of the rehabilitation outcome. Whereas device metrics are strongly linked to motor skill rehabilitation per se, serious gaming metrics can reflect cognitive factors, such as attention, working memory, and decision-making processes. Although such cognitive factors are not typically regarded as relevant for understanding motor performance, cognitive abilities represent fundamental constraints on learning and execution of movements (37,38).

This study has a couple of notable limitations. Firstly, the small number of participants was the most important major limitation of this pilot study. Studies with larger cohorts are required to validate our preliminary results and allow clinical trials focused on new rehabilitation approaches. Secondly, the applicability of the MR_CHIROD v3 is limited by the simplicity of its grip functions and the lack of multiple degrees of freedom. Notwithstanding, the device's grip functions are critical for stroke patients given that, from a neurology point of view, grip rehabilitation is the most important goal for achieving independence in daily life, including the ability to reliability grasp and hold onto objects without dropping them. In this context, it is important to note that although robotic therapy has been shown to improve arm motor function after stroke (39-43), apart from a few reports (43-45), these efforts have not focused on the hand (46) as the presently examined system does.

The present results demonstrate that MR_CHIROD v3 device output parameters and linked serious game score elements can be used as indices of hand motor function during post-stroke hand rehabilitation training. These robot and game metrics showed especially marked improvements during the initial 2 weeks of rehabilitative training in the present study sample. Clinical motor scale scores were found to associate very strongly with these indices, with the best correlated metric being average force. Game score elements and device output parameters can be combined to build a more reliable predictive model of clinical motor performance and, potentially, replace commonly used unreliable self-assessment methods. Larger studies are needed to inform the development of models that can be used reliably for remote monitoring and patient prognosis applications.