• electromyogram
• myokinetic interface
• multiuser interface
• pattern recognition
• transient control
• prosthetic hand
• transfer learning

Research pattern recognition myoelectric controls that include several movements require too much extensive algorithm training sessions to be clinically suitable. The patient has to reproduce different movements several times to build a data set for training the algorithm and the length of the process is proportional to the number of movements to train. In order to reduce the training time, we introduced transfer learning to the already existing machine learning algorithm to build a general common model from different subjects that can be tailored to a new user. The chosen transfer learning algorithm is Canonical Correlation Analysis (CCA) for its simplicity and effectiveness. This technique makes two views of the same set of objects and projects them onto a different space in which they are maximally correlated. It is an attempt to create a unified-style-space, i.e., attempt to make all training users' style similar to that of the expert. The novel user requires minimal training effort to obtain the projection matrices which will transform the user's data into a space where it is maximally correlated to the expert user.

The data points are extracted at the moment of the EMG transient. This way there is no continuous classification, movements are classified only when an Onset Detection Algorithm (ODA) recognizes a transient, which makes the whole system less prone to errors. Additionally, since the contraction precedes the actual movement, the response time of the transient classifier is faster than that of a conventional continuous classifier.

In this work, two different data sets were explored, surface and intramuscular EMG data set. Intramuscular EMG data set had signals recorded from 5 channels and divided into 5 classes, with 4 repetitions per each class. After processing the data, several subjects had to be removed due to noise and inconsistent contractions, leaving only 4 subjects that were considered for training and testing. On the other side, surface EMG data set had 8 channels and 8 classes with 20 repetitions per each class. The number of subjects which were included into training and testing is 15.

Firstly, the chosen machine learning algorithm, Support Vector Machine (SVM) was trained on each subject individually from both data sets in a form of crossvalidation, leaving one repetition per each movement for testing (CV LORO). Intramuscular EMG data set had an average accuracy across all subjects around 88\%, while surface EMG data set had more than 90\% average accuracy. It is expected that the surface EMG data set has greater accuracy due to more data available, for each class surface data set had 20 data points compared to 4 for intramuscular data set.

Subsequently, SVM was trained on data from all subjects combined. For finding the threshold for ODA, four different methods were observed: finding one threshold for all training data, finding individual threshold for each subject in the training set and choosing the minimum/median threshold for the test subject and choosing the optimal threshold for each subject whether it be in training or testing set. Here crossvalidation was implemented by leaving one subject for testing and training the SVM on the rest of the subjects (CV LOSO). Intramuscular data set under-performed with the average accuracy of 45\%. Class "Three digit pinch" had an accuracy of less then 10\%. This is due to low number of data and subjects, hence this data set was no longer considered for future testing. Surface EMG data set had similar average accuracy for all four different methods of finding the threshold for the ODA and was around 57\%. The maximum accuracy achieved was 59.7\% when median threshold was used for ODA. This was still much lower compared to the accuracy obtained from SVM trained on each individual subjects, which was over 90\%, therefore transfer learning algorithm was introduced.

Canonical Correlation Analysis was used on the feature set before the training of the SVM. Due to the use of the EMG transient there is less data compared to when EMG steady state is used, which means less data to adequately represent the feature space of the new user. To this end data was oversampled by extracting random points in the range of each feature. Without oversampling, the surface EMG data set had very low performances, with average accuracy of 29\% for when 5 repetitions per movement was used for calibration. The more the data was oversampled, the greater the performance was, with 76.5\% of average accuracy for when 5 repetitions per movement were oversampled to 100 repetitions per movement. This is grater than the accuracy obtained with SVM without CCA. However, the accuracy obtained from SVM trained on each individual subject with CV LORO when only 5 repetitions per movement was used for training and testing is slightly over 80\%, which means that the CCA does not increase the performance of the SVM enough.

Upon observing the accuracies per class, it was noticed that some classes have very low performances, like the "Side grip" class with around 20\% class accuracy. Therefore 3 classes were removed: "Pronation", "Supination" and "Side grip". CV LORO was implemented again on the surface EMG data set with the reduced number of classes, as well as the CV LOSO with CCA and the performances were compared. It has been found that for low number of calibration data for CV LOSO and training data for CV LORO (3, 4, 5 samples per class) and large oversampling (up to 100 samples per class) the CV LOSO with CCA outperforms the CV LORO by few percentages.

To further try to increase the performance, the SMOTE algorithm was introduced as a data augmentation method. The calibration data was oversampled from 5 to 100 samples per class and performance of the SVM was compared to that of the SVM with data oversampling with random extraction. Both methods reached to around 90\% average accuracy, meaning that the oversampling method does not contribute to the performance of the CCA or the SVM.

Finally, an SVM adaptation was implemented. The SVM was retrained with the support vectors obtained from the SVM trained on subjects in the training set and the oversampled calibration data from the test subject. Calibration data was oversampled from 5 to 100 samples per class with SMOTE algorithm. The adapted SVM with and without CCA was compared, with former outperforming the later with more than 5\%, showing the benefits of the CCA. Maximum average accuracy obtained with 5 repetitions per movement for calibration was achieved when SMOTE algorithm was used to oversample data to 100 samples per class, using CCA as transfer learning algorithm and SVM adaptation, reaching almost 92\%. This is by 7\% greater than the accuracy obtained with CV LORO, proving the applicability of these algorithms.

The CCA and SVM adaptation was performed again to surface EMG data set when only class "Side grip" was removed. The highest performance is for adapted SVM with CCA when calibration data was oversampled from 5 to 100 samples per class with average accuracy of 87.4\%. Considering the large number of classes (in this case 7), and that the data from different subjects was used for training, this accuracy quite sufficient.

In conclusion, transfer learning algorithm CCA and SVM adaptation contribute greatly in increasing the performance of the machine learning algorithm SVM when it comes to training data from different sources. Using CCA will significantly reduce the training time of the novel prosthetic user, increasing its applicability.