【Prosthetic Hands】20 Third Generation - Experiments

 In this article, I will introduce two primary experiments.
1. Comparison of different feature extracting method-and-classifier combination

2. Comparison of Decision Strategies

In both experiments, the experiment setup is as follow. The subject sits comfortably on the chair, with left hand holding the prosthetic hand and the right hand wearing MYO armband. In the first experiment, I acquired 4,000 EMG signals from eight motion types, including Power Grasping, Wrist Extension, Lateral Pinch, Flattened Hand, Centralized Grip, Hook Grasping, Three-jaw Chuck, and Flexion (each for 500 times). Half of them are used to generate the classifier with 5 folds of cross validation to prevent overfitting; the other half are treated as the data to be predicted. The way I evaluate the effectiveness of a classifier is the Confusion Matrix and F-socre [1].

The test result of the combinations of 7 feature extracting methods and 3 types of classifiers are shown below.

The combination of Medium Gaussian SVM and RMS reaches the best result, roughly 96.2%. Though the results seems promising, some aspects of practical usage are neglected. The experiment only includes one subject's EMG data, so it is uncertain that the result will be as good as what I obtain here.

In the second experiment, I separated the motions into two types - single grasping task and multiple grasping task. In single grasping tasks, users only have to perform grasping once (like lifting a water bottle), whereas in multiple grasping tasks, users have to move objects several times (like moving fruits into the refrigerator). The grasping motions used in the experiment include Power Grasping, Lateral Pinch, Three-Jaw Chuck, Hook Grasping, and Indexing Hand.

There are some minor experiment settings that ensure the consistency of each test, such as the duration of the motion signals and the distance between objects. Subjects have to first perform each motion for 10 times to build up their customized classifiers before they conduct three rounds of grasping tasks. The first round and the third round (using Practical Strategy) are used to examine whether subjects performance can be improved, and the second round (using Intuitive Strategy) is used to compare different strategies.

(The "KNN model" should be corrected to "SVM model")

The results are shown below. The average time for subjects to finish single grasping tasks is shorter when using Practical Strategy than using Intuitive Strategy. This is because less motion signals are needed in Practical Strategy (only three, compared to four in Intuitive Strategy). However, because in Practical Strategy, user must perform three motion signals in every grasping, it takes longer for them to finish multiple grasping tasks (user can stop at any motion in Intuitive Strategy, while they have to perform all three motion in Practical Strategy).

Comparison of two strategies - Total time cost of each task

To conclude, Practical Strategy is indeed more efficient when performing single grasping task. Also, the users also imply that performing two identical motion signals is simpler. The double insurance mechanism in Practical Strategy also enables them to take action when they realize that the hand is performing the motion to return to resting position.

Nonetheless, the experiments are still relatively simple.  The number of subjects and types of grasping tasks should be increased to get a more precise result. Also, more scenarios can be added, such as grasping while standing or grasping while walking, so that the functionality of the hand in all daily situations can be examined.  

[1] M. Sokolova, G. Lapalme, “A Systematic Analysis Of Performance Measures For Classification Tasks,” Information Processing & Management, Vol. 45, No. 4, pp. 427- 437, 2009.