An Approach To Evaluate Disability Using Machine Learning And Sensor-Equipped Instruments Demonstrated On Hand Grip Force Data

The goal of this research is to improve the process of rehabilitation. Even though the current tools and technologies are meeting the requirements of clinical use, there is a lot of room for improvement. Improving the tools used in therapy can help reduce costs and save time for the patient and the therapist by giving a valuable insight into the progress of the therapeutic process. Combining therapeutic tools with cognitive tools such as machine learning can assist a therapist to determine and predict functional ability and help patients keep track of their progress. The long-term goal for instrumented outcome measures (IOM) development is to use the captured clinical data in analytics to predict a functional score. The process starts by selecting what type of measure (grip, flexion/extension, balance, walk, etc.) is to be predicted or categorized as “normal” or “not normal/condition”. Then we determine the type of sensor that will facilitate the measurements and the placement of sensors. The performance observed for a functional measure is determined from the time-varying signals recorded for each task (gripping, flexion, extension, gait, etc.), either from collected data or derived from the collected measure. In this research, an off-the-shelf hand dynamometer (Vernier) was used to collect the grip data from 253 participants: 206 young adults of ages between 18-30 and 47 older adults of ages 65 and above. The data was collected from two experiments. During the maximum force experiment, each participant had to grip to the dynamometer for 5 seconds at the start of a tone and release it after a 2nd tone. During the tracking experiment, each participant had to grip the dynamometer to manipulate a moving block on a computer screen. Various markers related to the grip characteristics were extracted from the data. The extracted markers were (1) examined through descriptive statistics and (2) fed into machine learning algorithms using Microsoft Azure Machine Learning Studio. Different learning models were tried and evaluated for their performance in predicting ability vs. disability from both individual markers and combinations of markers