A Decision Support System to Provide an Ongoing Prediction of Robot-Assisted Rehabilitation Outcome in Stroke Survivors.

Clinicians often deal with complex robotic platform and serious games in stroke patients rehabilitation contexts, and they face two main problems: 1) the interpretation of either the performance in game or measures of a robotic system from the motor recovery point of view, and 2) the duration and complexity of clinical scales administration that makes repetitive assessments during the therapy unpractical. In this paper, a Random Tree Forest based system was trained and tested to provide a prediction of different clinical outcomes (i.e. FMA, ARAT, and MI) along the whole therapy duration, having non-clinical measures only as inputs, acting as a simulated decision support system. The dataset includes 30 post-stroke patients, that underwent a 30-session robot-assisted rehabilitation treatment. Results have shown that the system is able to produce very accurate and reliable predictions about the motor recovery of the patient at the end of the therapy, already in the first phases of the rehabilitation (i40% of therapy execution), just using robotic platform measures. Such a tool would provide a great benefit in terms of rehabilitation objectives planning, as a decision support tool for highly personalized rehabilitation treatments.

A Random Tree Forest decision support system to personalize upper extremity robot-assisted rehabilitation in stroke: a pilot study.

Robotic-based rehabilitation administered by means of serious games certainly represents the frontier of rehabilitation treatments, offering a high degree of customization of therapy, to meet individual patients' needs and to tailor a proper rehabilitation therapy. Despite the rush on developing complex rehabilitation systems, they often do not provide clinicians with long-term information about the outcome of rehabilitation, thus, not supporting them in the initial set-up phase of the therapy. In this paper, a Random-Forest based system was trained and tested to provide a prediction at discharge of several clinical scales outcomes (i.e. FMA, ARAT, and MI), having clinical scale scores and measures from the robotic system at the enrollment as inputs. The dataset includes 25 post-stroke patients from different clinics, that underwent a variable number of days of rehabilitation with a robotic treatment. Results have shown that the system is able to predict the final outcome with an accuracy ranging from 60% to 73% on the selected scales. Also results provide information on which variables are more relevant for the prediction of outcome of therapy, in particular clinical scales scores such as FMA, ARAT, MI, NRS, PCS, and MCS and robotic automatically extracted measurements related to patient's work expenditure and time. This supports the idea of using such a system in a clinical environment in a decision support tool for clinicians.