Conference Agenda

Myoelectric prostheses are still far from fully replacing the complete functionality of the lost limbs. Particularly challenging are the situations in which joint stiffness must be varied during task execution. The human ability to dynamically adjust the mechanical properties of the joints in response to environmental demands is fundamental for a natural interaction with the surroundings. The integration of this feature in upper-limb prostheses is currently under studies. Most of the approaches proposed in literature are focused on developing mechatronic devices with variable stiffness. Whereas, the possibility of changing prosthetic joint impedance via software could enhance the integration of the approach into various commercial devices.
EMG-driven musculoskeletal models are promising for estimating joint stiffness due to their high accuracy and ability to extract additional parameters, such as joint torques and angles. They typically assume Hill-type muscle behavior and use EMG signals to estimate muscle force. However, they suffer of two main drawbacks: 1) the computational cost of the most accurate approaches, such as Opensim models, often exceed the response time required for online prosthesis control (< 150ms). 2) these models require calibration datasets with EMG signals and force measurements, typically obtained through subject specific experimental setups, limiting calibration process due to the need for specialized instrumentation.
The contribution of this work is twofold. First, we implement an innovative musculoskeletal model calibration pipeline using kinematic data instead of torque and force information, in order to reduce experimental setup complexity. Second, for elbow joint stiffness estimation, we use the MuJoCo environment instead of OpenSim, since it offer same accuracy and allows much faster simulations, enabling real-time estimation.
The proposed calibration pipeline gives in input to MuJoCo model the muscles activation level and compares its output (i.e. elbow angle) with data obtained from an optoelectronic system. The resulting Root Mean Square Error (RMSE) is minimized through a Powell optimization process, which adjusts the parameters of the muscle model to better replicate the reference data. This loop continues until the simulated output closely matches the real movement. To validate the proposed approach, experiments were conducted in order to analyse how stiffness varies across different loads and elbow-shoulder configurations. Ten participants (aged 29 ± 3) performed static acquisitions in various shoulder (0, 90°) and elbow configurations (45, 90, 135°), as well as under different loading conditions (0, 1.5, 3.0 kg). For each combination of shoulder-elbow-load configuration, the task was repeated 3 times.
Results show that the estimated stiffness values fall within the range reported in the literature (0–40 Nm/rad). Furthermore, statistical and biomechanical analyses were conducted to assess the efficacy of the proposed approach. The results demonstrate that the calibrated musculoskeletal model successfully distinguishes stiffness variations across different joint configurations and loading conditions.
In conclusion, a new methodological approach is proposed to integrate the musculoskeletal model information into prostheses control. This will allow to have a more intuitive prosthesis control.

Model explainability is becoming crucial for machine learning (ML) to unveil the reasoning behind the system’s decision and justification of its response. In health domains, these aspects are relevant as they can provide arguments to the evidence-based medicine and can become determinant when providing decisions on patients management and care after discharge.

Neurological diseases, such as stroke, although extremely variable under etiological and pathological perspectives, are characterised by clinical characteristics of often severe motor and cognitive impairments affecting everyday life activities. During the disease progression, the prediction of motor recovery may contribute to identify timely therapeutic decisions.

The European Medicine Agency has recently underlined the potential benefits of digital health technologies to improve the quality of the care standards, in particular adopting wearable devices, telemedicine and robotics in medical applications.

Artificial Intelligence and ML techniques may play a significant role in interpreting data coming from these applications in order to provide reliable tools to support clinicians' decisions.

The aim of the PREDICTOR project, funded under the PRIN national programme, is to develop and validate a proof-of-concept of an innovative technological framework based on the integration of robotics and wearable sensors for the prediction of motor recovery in the early subacute phase of stroke survivors, who will undergo an upper limb robot-assisted rehabilitation program, by means of explainable ML techniques.

Data from a) a planar end-effector robotic device for upper limb rehabilitation, b) wearable sensors (for home monitoring), and c) clinical scales will be analysed and used to train specific ML algorithms whose final expected output is represented by the prediction of 1) upper limb motor recovery in terms of scores of clinical outcome scales and 2) parameters (such as viscosity, weight and stiffness) in order to improve the patient-robot interaction. The former is a long-term prediction (e.g. prediction at 3/6 months after the admission), the latter is a short-term prediction (e.g. prediction for the subsequent rehabilitation session).

Stroke remains a leading cause of disability worldwide, often resulting in long-term motor impairments that necessitate intensive physical rehabilitation. In recent years, robotic systems have been increasingly integrated into rehabilitation programs, offering objective kinematic measurements and the ability to deliver consistent, high-intensity therapy. However, despite their promise, there remains a pressing need to better personalize treatment pathways and optimize resource allocation in clinical settings. This presentation introduces a predictive model developed to forecast the rehabilitation outcomes of post-stroke patients undergoing robot-assisted therapy, utilizing an integrated set of clinical, demographic, and robotic kinematic data.

Our model draws from a rich dataset comprising clinical scales (e.g., Fugl-Meyer Assessment, Modified Ashworth Scale), demographic variables (e.g., age, gender, time since stroke), and detailed kinematic parameters captured during robot-mediated therapy sessions. Using advanced machine learning techniques, we trained and validated the model to predict patients' functional gains over the course of rehabilitation, providing clinicians with valuable foresight into expected recovery trajectories.

Beyond prediction, the model also addresses a critical clinical need: dynamic personalization of rehabilitation exercises. By analyzing a patient's motor capabilities and real-time performance data, the system suggests optimal robot parameters that tailor the difficulty of exercises to each individual’s current ability, following previous choices of expert clinicians. This ensures that therapy remains appropriately challenging, promoting neuroplasticity and maximizing functional recovery.

The proposed approach offers several benefits to healthcare systems. First, by accurately predicting rehabilitation outcomes early in the therapy process, clinicians can make informed decisions about resource allocation, prioritizing intensive interventions for patients most likely to benefit. Second, the model's ability to recommend exercise parameters helps therapists fine-tune rehabilitation plans without relying solely on trial and error or subjective assessments, thereby saving time and enhancing treatment efficacy. Lastly, this personalized strategy fosters a patient-centered rehabilitation process, empowering patients through appropriately scaled challenges and enhancing engagement and motivation.

In a retrospective study across multiple clinical sites, our model achieved a good predictive accuracy and demonstrated robust generalizability across different patient profiles. Moreover, preliminary feedback from clinicians indicated that the model's parameter recommendations were both feasible and helpful in clinical decision-making.

This presentation will detail the architecture of the predictive model, the methodologies employed for feature selection and model training, and the validation results. Additionally, we will discuss case studies highlighting how model-driven parameter suggestions were accepted on average by the clinical personnel. Finally, we will address potential limitations, such as data quality challenges and the need for prospective validation, and outline future directions.

By providing a data-driven framework for both outcome prediction and exercise personalization, our work moves towards a future of smarter, more efficient, and more individualized rehabilitation for post-stroke patients.