Knowledge Discovery strategy over patient performance data towards the extraction of hemiparesis-inherent features: A case study

Abstract

Aiming to perform an extraction of features which are strongly related to hemiparesis, this work describes a case study involving the efforts of patients in upper-limb rehabilitation, diagnosed with such pathology. Expressed as data (kinematic and dynamic measures), patients' performance were sensed and stored by a single InMotion Arm robotic device for further analysis. It was applied a Knowledge Discovery roadmap over collected data in order to preprocess, transform and perform data mining through machine learning methods. Our efforts culminated in a pattern classification with the abilty to distinguish hemiparetic sides with an accuracy rate of 94%, having 8 features of rehabilitation performance feeding the input. Interpreting the obtained feature structure, it was observed that force-related attributes are more significant to the composition of the extracted pattern

Machine-learning-based biomarkers towards customization of robotic rehabilitation treatments for stroke patients

“…As previously highlighted in Section 2.1.2, both feature selection and extraction techniques are possible to perform over raw data. For the sake of interpretability, regarding feature extraction, in most of the cases it is preferred to keep the underlying semantics from the raw input (MORETTI et al, 2016;DIPIETRO et al, 2012;COLOMBO et al, 2012;JUNG;GLASGOW;SCOTT, 2008) avoiding spatial transformations. However, it is sometimes the case that more than one technique, including spatial transformations (BOSECKER et al, 2010) for dimensionality reduction, is required.…”

“…In previous work (MORETTI et al, 2016), yielding 24 features purely based on descriptive statistics, we calculated the statistical moments of the distribution of kinematic and dynamic measures, such as positions, velocities and forces. Different from Dipietro et al (2012) and Bosecker et al (2010), in which the same device was used, we considered every movement (backwards or towards center) of the star-like pattern, so every feature is a descriptor of the entire cycle of movements composing such pattern.…”

“…The robotic device MIT-MANUS was used in (MORETTI et al, 2016;DIPIETRO et al, 2012;BOSECKER et al, 2010;KREBS et al, 2003), also collecting data from the same gaming environment. Displaying eight targets around the edge of a clock-like interface, having its positions arranged at every 45 degrees, and a central target, the goal is to perform reaching movements towards every indicated target, and backwards, towards center.…”

“…A machine learning strategy (JUNG; GLASGOW; SCOTT, 2008) aims to propose a measure based on the outputs of a hierarchical ensemble classifier, considering performance data from both healthy and impaired limbs. Also using machine learning, in (MORETTI et al, 2016), as a preliminary work, we aimed the feature extraction for the spatial separability of features, inherent to hemiparetic sides, for further studies in the obtained space for proposing a way to quantify progress. Bosecker et al (2010) assess patients conditions through estimates of clinical scales as a function of patient behavior.…”

“…Enhancing our previous results (MORETTI et al, 2016), this strategy, aiming the definition of the hemiparesis-inherent data space, is organized in the following sections, similarly as established in the KDD roadmap (Figure 1).…”