Machine Learning in Biomechanism and Rehabilitation

Identification of robotics metrics for predicting outcome in stroke patients

Stroke is one of the leading causes of permanent disability in Canada, where nearly 80% of stroke survivors suffered from some form of disability. Around 50,000 individuals suffer from stroke every year, and there are an estimated of 300,000 individuals living with stroke side effects in Canada.

Classic clinical scores for stroke assessment tend to rely on observer-based ordinal scales, many of which have limited resolution and reliability. Recently, more advanced robotic technologies are capable of recording objective and highly reliable data for assessment of brain impairments that have been developed. KINARM (Kinesiological Instrument for Normal and Altered Reaching Movement) is a robotic device that quantifies many areas of brain dysfunction for stroke survivors.

Several tasks are presently performed on the KINARM robot for quantification of sensorimotor, proprioceptive and cognitive brain function. These include, but are not limited to, a visually guided reaching task, limb proprioceptive tasks, and a rapid target interception task (Figure 1).

The two main concepts of this project are:

1) The use of robotic technologies for prediction of stroke prognosis. The main concern regarding this concept is to determine whether the data collected using the KINARM robot can be used to predict the process of stroke subjects after rehabilitation program with the help of computational and mathematical techniques.

2) Time reductions on the current KINARM assessment protocol. As more tasks are incorporated on the KINARM protocol, the length of time to assess each subject continues to grow. This leads to the question of whether the length of overall assessment time can be reduced while retaining the maximal amount of information to quantify subject performance across broad range of neurological functions. For this purpose, the application of computational techniques will be considered to determine various methods by which the overall assessment time on the KINARM can be reduced.

Novel methods in SEMG-Force estimation


An accurate determination of muscle force is desired in many applications in different fields such as ergonomics, sports medicine, prosthetics, human-robot interactionand medical rehabilitation. Since individual muscle forces cannot be directly measured,force estimation using recorded electromyographic (EMG) signals has beenextensively studied. This usually involves interpretation and analysis of the recordedEMG to estimate the underlying neuromuscular activity which is related to the forceproduced by the muscle. Although invasive needle electrode EMG recordings haveprovided substantial information about neuromuscular activity at the motor unit(MU) level, there is a risk of discomfort, injury and infection. Thus, non-invasivemethods are preferred and surface EMG (SEMG) recording is widely used. However,physiological and non-physiological factors, including phase cancelation, tissue filtering,cross-talk from other muscles and non-optimal electrode placement, affect theaccuracy of SEMG-based force estimation. In addition, the relative movement of themuscle bulk and the innervation zone (IZ) with respect to the electrode attached tothe skin are two major challenges to overcome in force estimation during dynamiccontractions.

The objective of this work is to improve the accuracy of SEMG-based force estimationunder static conditions, and devise methods that can be applied to forceestimation under dynamic conditions. To achieve this objective, a novel calibrationtechnique is proposed, which corrects for variations in the SEMG with changing jointangle. In addition, a modeling technique, namely parallel cascade identification (PCI)that can deal with non-linearities and dynamics in the SEMG-force relationship is appliedto the force estimation problem. Finally, a novel integrated sensor that sensesboth SEMG and surface muscle pressure (SMP) is developed and the two signalmodalities are used as input to a force prediction model.

The experimental results show significant improvement in force prediction usingdata calibrated with the proposed calibration method, compared to using noncalibrateddata. Joint angle dependency and the sensitivity to the location of thesensor in the SEMG-force relationship is reduced with calibration. The SEMG-forceestimation error, averaged over all subjects, is reduced by 44% for PCI modelingcompared to another modeling technique (fast orthogonalsearch) applied to the same dataset. Significantly improved force estimation results are also achieved for dynamiccontractions when joint angle based calibration and PCI are combined. Using SMPin addition to SEMG leads to significantly better force estimation compared to using
only SEMG signals.

The proposed methods have the potential to be combined and used to obtainbetter force estimation in more complicated dynamic contractions and for applications such as improved control of remote robotic systems or powered prosthetic limbs.

Subscribe to RSS - Machine Learning in Biomechanism and Rehabilitation