A total of 20 patients with PD were selected by the outpatient department at the Clinical Unit of Physical Rehabilitation, Istituto di Ricovero e Cura a Carattere Scientifico Istituto Nazionale Ricovero e Cura per Anziani (IRCCS INRCA). The patients were recruited if they were over 65 years old; able to provide informed consent; had a stage of Hoen and Yahr scale between 1 and 2 [12]; had a Functional Ambulation Category score ≥2 [13]; had a Rankin Scale score ≤3 [14]; had stability of drug treatment for at least 1 month; had a Geriatric Depression Scale 4-item score ≤1 [15]; had a Mini-Mental State Examination ≥24 [16]; and could maintain an upright posture≥30 seconds, evaluated by a trained physiotherapist during the recruitment. The evaluation of compliance with the inclusion and exclusion criteria was performed during the recruitment session. Once we completed this phase, informed consent was obtained. The patient’s clinical assessment was performed at the start and the end of the treatment. In particular, the evaluation consisted of the administration of the following scale: measurement of functional state with the Barthel Index [17]; physical performance with Tinetti’s Performance-Oriented Mobility Assessment (POMA) [18], the Short Physical Performance Battery (SPPB) [19], the 6-Minute Walking Test [20], and the Timed Up and Go test (TUG) [21]; evaluation of the quality of life with the 12-Item Short-Form Health Survey (SF-12) [22]; fear of falling with Falls Efficacy Scale-International (FES-I) [23]; and the assessment of the prognosis of PD with the Unified Parkinson Disease Rating Scale – III (UPDRS-III) [24].

Figure 1 shows the experimental setting in a protected environment (gymnasium) and the positioning of the technologies, the patient, and the physiotherapist, during the dance-based rehabilitation sessions.

A total of 16 therapy sessions of 50 minutes were conducted (2 sessions per wk, for 8 wks), involving 2 patients at a time. Cardiac and respiratory activity was monitored during dancing treatments to detect heart rate and breathing frequency. The participants were required to complete at least 80% of the sessions. Each session involved the following activities: (1) breathing, relaxation, and postural harmonization exercises (5 min); (2) active mobility and stretching exercises (5 min); (3) the SI-Robotics intervention (35 min) consisting of the AI-enhanced “Let’s Dance” game, and a socially interacting robot monitoring the execution of the physical exercises of patients; and (4) relaxation exercises (5 min).

After selecting a difficulty level, players are presented as dancers on the screen, along with footprints that suggest the movement to be done. Each task can vary from simple aerobic exercises (side steps, arm raising, hand clapping, etc) to choreographies. At the end of the game session, a score is presented to the patient. This would make a patient aware of the quality of the performance. This score can also be used by the therapist to assess the patient’s performance and the possible increase in the difficulty level of the game.

The next sections describe the technological components of the session in more detail.

The designed rehabilitation session involves three main actors: (1) a therapist, (2) a patient, and (3) a socially interacting robot. To stimulate engagement, sessions are organized in the shape of exergames where an avatar shows patients the movements they should perform to correctly carry out a rehabilitation exercise (eg, dancing steps). Given a set of data about the needs of a group of patients and about stimulation capabilities of known exercise, a game engine integrates an AI planner [25] and synthesizes exercises by selecting a (sub)set of movements (ie, dancing steps) that best fit the clinical objectives of the session (ie, of the current patient). The exergame is enriched with AP to synthesize suitable physical exercises, contextualized to the clinical objectives of the rehabilitation sessions [25]. Many works in the literature investigate the use of AI in health care and PD [11,26] AI is used for example to predict the wearing-off of symptoms [27], support decisions [28], or early diagnosis of PD [29]. The majority of these works adopt AI solutions based on deep or machine learning and focus on the diagnosis of the disease. Few works investigate the use of AP to support therapists in the synthesis of rehabilitation programs. The study by Gonzalez et al [30], for example, integrates AP into a control architecture to allow a social robot to physically show motions to a patient during physical rehabilitation. Baschieri et al [31] uses AP to synthesize simulation scenarios within a serious game for cognitive rehabilitation. This work pursues an objective similar to ours but in a different clinical scenario.

As shown in Umbrico et al [25] SI-Robotics proposes AP to support the synthesis of physical rehabilitation programs for patients with PD . A key aspect of the developed planning framework is the combined reasoning about spatial and clinical effects of stimuli (ie, motions or dancing steps) needed to synthesize plans that are technically valid and effective from a clinical point of view. Compared with a manual definition of the rehabilitation session, the use of a planner aims to improve the quality, accuracy, and engagement of the resulting rehabilitation programs. A therapist provides a planner with data about the rehabilitation session (song time, song rhythm, and difficulty level) and the clinical objectives. Given this input, the planner decides a sequence of steps, optimized according to an objective function encoding the specified clinical objective. Planned steps are thus chosen according to the rehabilitation needs of the participating patient (personalization).

In addition, several devices enrich the session to extract useful data about the health conditions and the performance of patients. The generated data streams are gathered by the robot that embodies the AI-based reasoning modules. Overall, three types of perception components were considered: (1) wearable sensing devices (eg, sensorized shirts) that constantly produce data streams about physiological parameters of patients (eg, heart rate) [32]; (2) a video-processing component elaborating 3D camera data to extract kinematics about motions and produce data about body posture and body equilibrium of patients and; (3) a video-processing component elaborating 3D camera data to analyze patients’ behaviors and produce data about the correctness of performed motions (ie, feedback). Figure 2 below shows a conceptual overview of the designed session pointing out a subset of the main hardware and software components and the data streams generated during the execution of a rehabilitation session.

In addition, a socially interacting robot should support the execution of a session by recognizing the current state of a session according to data gathered from the environment, autonomously detecting events that would trigger different phases, and setting the behavioral goals needed to support a session. The robotic platform consists of a novel social robotic platform designed within SI-Robotics and developed by Co-Robotics [33]. Figure 3 shows the lifecycle determining the assistive behavior of the robot during each rehabilitation session.

A rehabilitation session starts with a welcoming state where the robot introduces patients to the session. Patients do not know the structure of the rehabilitation and are not familiar with the involved technology. The objective of this state is therefore to prepare patients for the therapy by explaining the structure of the session and the technology used. When all patients are ready to start a session, the robot enters into a “monitoring” state gathering data from the environment to monitor the performance and health state of patients. When a rehabilitation exercise ends the robot enters into the “reporting” state to interact with the therapist to comment on the exercise and the observed performance. This state is especially useful to enrich the subjective experience of the therapist by enhancing the awareness of the therapist with objective feedback on the quality of the exercises. It is indeed meant as a support for the decisions on how to proceed with the session (eg, change or repeat the exercises, difficulty of the next exercise, etc). If the therapist decides to start a new exercise, then the robot enters again into the monitoring state. Otherwise, the robot enters into the “closing” state to finalize the session. In this latter case, the robot interacts with patients to ask their personal feelings about the session and to show a qualitative assessment of their session.

Different types of interaction are foreseen during an exercise, depending on the type of situation recognized by the robot. Two types of interactions have been considered characterized by the motivating and the alerting states in the automata. The motivating phase allows the robot to play the role of a coach during the session stimulating patients to perform better or correcting their behavior. The robot is supposed to enter the motivating phase when it detects a heart rate below some minimum “training threshold” or when it detects a wrong posture of the body, for example, bad alignment or asymmetrical motion of lower or upper limbs.

The alerting phase allows the therapist (and the robot) to intervene when some critical situation concerning the health state of a patient is detected during the session. In such case, the robot is supposed to warn the therapist about the danger and simultaneously interact with the involved patient, even interrupting the exercise if necessary. Specifically, the robot enters the alerting phase when it detects the risk of a fall or an anomalous heart rate for a patient. Both these situations are critical for the considered target of patients and need a prompt intervention.

A key aspect (and still an open challenge) in the design of autonomous robotic agents is the integration of different AI technologies and Robotics [34,35]. SI-Robotics proposes an AI-based control architecture with twofold objectives: (1) to support the abstraction and reasoning capabilities needed to recognize health-related situations and assistive objectives and (2) to coordinate robotic skills to “act” in the environment and (autonomously) achieve contextualized assistive objectives. Figure 4 shows an overview of the designed architecture with the main functional components involved in the control process. Two main architectural layers are considered and are organized according to a cognitive architecture inspired by the Dual Process Psychological Theory [36].

A skill control layer encapsulates the control components responsible for providing the robot with the “low-level” capabilities needed to perceive the environment and interact with the environment. The Navigation Skills component encapsulates functionalities that allow a robot to robustly navigate within the environment and autonomously reach known locations. The Dialog Skills encapsulate natural language processing functionalities that allow a robot to talk with patients and therapists. This component in particular enables an affective interaction [37] by adapting the way the robot talks with a patient to the detected emotional state and known personality traits.

While some level of perception is needed for both navigation and dialogue skills, the Perception Skills component emphasizes the additional capability needed to process data from deployed environmental devices. It encapsulates low-level data processing to provide the higher decisional level with refined information. A behavior control layer supports AI-based control features to contextualize observations, trigger assistive goals, and coordinate underlying skills to implement the desired assistance. The Autonomy and Skill Coordination component encapsulates an AI plan–based controller that supports behavior continuity coordinating robotic skills. The Goal Reasoning component instead is in charge of contextualizing observations to automatically set assistive goals. The combination of goal reasoning and a plan-based controller is the key architectural feature enabling proactive and autonomous goal-oriented behaviors of the robot [38].

The AI-based architectural elements composing the coordination layer of Figure 4 have been embedded into the robotic platform and integrated through the Robot Operating System (ROS) using the ROS Melodic distribution [39]. The coordination component responsible for the synthesis of the assistive behavior has been implemented in the shape of an AI plan–based controller relying on AP. The component uses the open-source framework ROXANNE [40], which implements goal-oriented acting capabilities into an ROS [41]. ROXANNE allows the robot to receive high-level assistive goals and (autonomously) synthesize and execute navigation and dialogue commands or tasks to realize a desired assistive behavior. The goal reasoning component correlates data received from the perception component with the autonomous behavior of the robot. It endows the robot with the cognitive capabilities needed to understand the evolution of a rehabilitation session autonomously select suitable planning goals and consequently decide the assistance needed in a particular context.

Embedded AP models the behavioral constraints needed to correctly support a rehabilitation session. Behavioral constraints are specified in the shape of temporal constraints and require interacting tasks performed through available skills. In this regard, Table 1 shows the modeled assistive goals and the parameters defined to contextualize interactions. Furthermore, the table briefly describes the behavioral constraints specified in the model to correctly implement the desired assistance and support the associated goal. It is worth noting that AI planning facilitates robot programming since it supports an “easy specification” of desired behavioral dynamics without hard-coding the implemented skills.

All assessment procedures adhere to a standardized protocol. Specifically, the primary focus of the study revolves around enhancing balance, and gait, and alleviating the fear of falling among elderly patients with PD. This is gauged through the utilization of the 3 POMA scales (POMA balance, POMA Gait, and POMA Total) after the 10 treatment sessions, as a result of the Irish dance intervention. In addition, secondary measures include evaluating the gait speed of elderly patients with PD, their fear of falling (assessed via FES-I), their physical performance (SPPB), their autonomy in daily living activities (UPDRS-III), and their overall physical and psychological well-being (SF-12) together with the evaluation of the Unified Theory of Acceptance and Use of Technology (UTAUT) that is composed by 10 subscales (anxiety; attitude; facilitating conditions; intention to use; perceived adaptability; perceived enjoyment; perceived ease of use; perceived usefulness; social influence; and trust).

Continuous variables were presented as either mean and SD or median and IQR, depending on their distribution, which was determined using the Kolmogorov-Smirnov test. Categorical variables were expressed as absolute numbers and percentages. To test statistically significant differences (P<.050) between pre- and postconditions, Wilcoxon signed-rank test (for non-normal distribution) or paired test (for normal distribution) were used, alongside simple descriptive statistics such as means, medians, and SDs as appropriate. The statistical analysis was performed using SPSS software.

The study was approved by the Ethics Committee of the Istituto Nazionale Ricovero e Cura per Anziani, (IRCCS INRCA) on June 17, 2021 (CE INRCA 21004). The protocol is registered on ClinicalTrials.gov with trial registration number NCT05005208 (October 23 , 2023). All participants signed the informed consent and data processing consent. The data are anonymised so that the identity of the subject cannot be traced. No compensation is provided for participation in the study.

This study aims to investigate the usefulness of the SI-Robotics intervention based on AI robotic assistance and exergame stimulation, in the context of PD. This study aims to assess the ability of a dance-based physical therapy program enriched by AI-based exergames and contextual robotic assistance in improving motor function, balance, gait, and quality of life in patients with PD. The findings of this study demonstrate the potential usefulness of a dance-based physical therapy program using a robotic platform to deliver exergames in improving motor function, balance, gait, and quality of life in patients with PD. Results suggest that integrating dance therapy with engaging technology may lead to greater improvements in PD symptoms compared with conventional physical therapy alone. Although the statistical analysis revealed significant improvements in motor function, balance, and gait (eg, UPDRS-III, POMA, and TUG scores), the effect sizes for several outcome measures were relatively small (Cohen d<0.2 for POMA and SPPB scales). This suggests that, while the observed changes were statistically significant, the practical impact of the intervention on certain motor and functional outcomes may be limited. These small effect sizes could be partly attributed to the small sample size, which may have reduced the power to detect larger effects.

This study adds to the growing body of research supporting the use of dance therapy and exergaming as adjunctive treatments for PD. Previous studies have shown that both dance therapy and exergaming can independently improve motor function, balance, and mobility in patients with PD [4,7]. By combining these approaches, our intervention aimed to target the multifaceted nature of PD symptoms, including motor, cognitive, and psychosocial aspects.

The results are consistent with previous findings demonstrating the benefits of dance therapy and exergaming in PD management. For example, Duncan and Earheart [4] reported significant improvements in motor function and balance among patients with PD participating in community-based dancing, while Barry et al [7] found that exergaming interventions were effective in improving mobility and reducing fall risk in patients with PD. These studies, along with ours, highlight the potential of integrating dance therapy with exergaming to address the complex needs of patients with PD.

One novel aspect of our intervention is the use of the SI-Robotics system, a robotics-based platform designed to engage patients with PD in Irish set dancing. This innovative approach combines the physical and cognitive benefits of dance therapy with the interactive and adaptable nature of exergaming. By incorporating personalized avatars and choreography, the “Let’s Dance” game provided a stimulating and enjoyable exercise experience for participants, potentially enhancing motivation and adherence to the intervention.

Regarding the analysis of technology acceptance among the population studied, the results indicate a generally positive perception and interaction with technology. In particular, users displayed a positive attitude towards the technology, indicating favorable feelings toward it. They also found the system adaptable, enjoyable, easy to use and useful showing that users experienced pleasure while using the system. Finally, trust in the system was moderately high, indicating that users believed it to be reliable and trustworthy. These highlight how the system successfully meets user expectations by offering an engaging, straightforward, and beneficial experience.

Our study focused on elderly patients with PD with mild to moderate disease severity, as reflected by their scores on the Hoen and Yahr scale, Functional Ambulation Category, Rankin Scale, Geriatric Depression Scale, and Mini-Mental State Examination. These inclusion criteria aimed to ensure that participants were physically and cognitively capable of engaging in the intervention safely and effectively. Future research could explore the applicability of our intervention to a broader range of patients with PD , including those with more advanced disease stages.

Limitations of our study include the small sample size and lack of a control group for comparison. While our results provide preliminary evidence of the usefulness of the intervention, larger-scale randomized controlled trials are needed to confirm these findings and establish the optimal dosage and timing of dance-based physical therapy with exergames for PD management. In addition, longer-term follow-up assessments could help evaluate the sustainability of the intervention effects over time.

In conclusion, this study highlights the potential of integrating dance therapy with exergaming integrated with an interactive robotic platform as a novel approach to improving motor function, balance, gait, and quality of life in patients with PD. The SI-Robotics intervention offers a promising avenue for future research and clinical practice in the management of PD symptoms. To strengthen the validity of the current findings and ensure their broader applicability, future research will focus on replicating the study with a larger sample. This will help to verify the generalizability of the results providing a more comprehensive understanding of the observed effects. Moreover, further studies are warranted to explore the long-term effects and feasibility of implementing this intervention in diverse clinical settings.