Recently, it was shown [1] that a myoelectric signal (MES), recorded during the initial phase of a contraction, has considerable structure which is distinct for contractions producing different limb functions. This has enabled distinctive features to be extracted from the signals which were used to distinguish between contraction types using an artificial neural network [1]. In the present paper, we show that parallel cascade identification [2] can be used to distinguish between contraction types, with the following benefits. The parallel cascade requires very little data to learn to distinguish contraction type accurately. Training is very rapid, and feature extraction is unnecessary so that sampled raw MES data can be used. Results of using parallel cascade identification to distinguish between medial and lateral humeral rotation are presented.Introduction
Until recently, it was believed that the myoelectric signal (MES) underlying a muscle contraction was stochastic and the instantaneous value of the signal contained no information. However, it has now been shown [1] that there is considerable structure in the signal during the initial phase of muscle contraction, and that this structure differs for contractions producing different limb functions. This specificity of the structure of the myoelectric signal during the initiation of a muscle contraction was a significant discovery. It made possible the extraction of distinctive features from the MES, which were used to train a neural network to distinguish between contraction types. Thus, a single channel of MES was used to control a number of functions of a prosthetic limb [1].
In particular, the authors [1] demonstrated excellent classification accuracy by the trained neural network in distinguishing the MES for four different contraction types, from both limb deficient and normally-limbed subjects. A number of repetitions of each contraction type were used to train the neural network. A separate network was trained to classify the signals for each subject. Retraining via an updating procedure was used to adapt to long term changes in generated signals [1].
In the present paper, we use parallel cascade identification [2] to distinguish between the contraction types, and show the following benefits in classifying MES. First, the parallel cascade is able to "learn" and "generalize" from very little data. Typically one example of each contraction type to be distinguished suffices to train the parallel cascade to achieve excellent classification accuracy for a large number of subsequent test signals. Second, the parallel cascade trains very rapidly, so that it can be retained conveniently whenever necessary to adjust to changing characteristics of the contraction types. Third, the parallel cascade readily trains on, and can then classify with high accuracy, sampled raw MES data. This means that the selection of appropriate features to represent the contraction types, which till now has been a critical issue to achieve high classification accuracy [1], is no longer required.
Methods
The procedure described in [1] was employed to choose the region of the MES to be sampled: 200 samples at 1 ms spacing were extracted starting 50 ms before a 100 point moving average of the absolute value of the amplified signal first exceeded 100 mV. (The MES data used in our study were supplied to us by the authors of [1], and had been amplified by a factor 5000 as described therein.)
Only binary classification by parallel cascade was to be tested, since a multi-state classifier can clearly be structured in terms of a series of binary classifiers. The second sequence from the series of MES recordings for medial (pronation, Fig. 1) and lateral (supination, Fig. 2) humeral rotation were selected. For training the parallel cascade, 200 successive samples from each of these 2 signals were extracted using the 100 mV threshold. The MES data had been gathered sequentially [1], but the first records in each series were not used for training because the 100 mV threshold did not cleanly select the entire structured region of the pronation signal.
The parallel cascade required about 2 s to train, on a 90 MHz Pentium. It was then tested on classifying all subsequent pronation and supination records, as well as the first signal in the 2 sets which had preceded the training pair. Again only 200 points at 1 ms spacing were used from each signal.
Results
The parallel cascade which had been trained on pronation and one supination signal correctly classified all 38 subsequent pronation records, as well as the pronation record preceding the one used for training. It correctly classified all corresponding supination records except for the 29th and 32nd records following the training signal. When the parallel cascade was retrained on the 24th pronation and supination records following the original training pair, it correctly classified all subsequent signals.
References
B. Hudgins, P. Parker and R.N. Scott, "A new strategy for multifunction myoelectric control," IEEE Trans. Biomed. Eng., vol. BME-40, pp. 82-94, 1993.
M. Korenberg, "Parallel cascade identification and kernel estimation for nonlinear systems," Ann. Biomed. Eng., vol. 19, pp. 429-455, 1991.
Acknowledgment
We thank Drs. Bernard Hudgins, Philip Parker, and Robert N. Scott for providing us with the MES data records.
[Figure 1 omitted]
[Figure 2 omitted]