An optimised region and boundary classifier for head movement classification
- Publication Type:
- Thesis
- Issue Date:
- 2009
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
The use of Artificial Neural Network (ANN) classifiers in real-time applications that
could be considered critical, such as head movement classification for the control of a
powered wheelchair, naturally raises questions over safety due to performance accuracy.
One inherent characteristic of an ANN is that the placement of classification boundaries
within the decision space is arbitrary and ultimately unknown. The result of this
characteristic is to give unpredictable results when presented with data outside the
boundary of the training set and is therefore considered as a major source of uncertainty.
The problem is one of accuracy and predictability and to address this a novel algorithm
termed an Optimised Region and Boundary Classifier (RBC) was created to provide
improvements in accuracy and predictability over a conventionally trained ANN.
To improve accuracy the RBC requires the formation of effective boundaries, which
relies on the training set containing data that is both representative of all types of data
likely to be input to the classifier and is complementary (data either side of an implied
boundary). To achieve this the original training set consisting of commands only
(forward, back, left, and right) was augmented with “data outside the boundary”, which
consisted of seventeen types of non-command data that the classifier was likely to see.
To improve predictability the RBC uses explicit boundaries, which is achieved using kmeans
clustering techniques to define Hyper-Rectangles. Also, two additional regions
(vertical and horizontal null regions) are extracted from the Hyper-Rectangles and
added to the classifier.
To further improve accuracy and predictability an optimisation process is used that
expands the Hyper-Rectangles for each of the classes to be classified (forward, back,
left, and right) until the associated training error for sensitivity and specificity is
optimal.
To show that the RBC could provide improvements in performance comparisons were
made between the RBC trained on the augmented training set and ANN’s trained on
both the original training set and the augmented training set. The performance for each
type of classifier was assessed using the 0.632+ Bootstrap Method and Receiver Operating Characteristics (ROC) analysis (area under the curve, sensitivity, and
specificity).
Results showed that the RBC provided significant improvements in performance when
compared with an ANN trained on the original training set (up to 9% improvement in
mean sensitivity and 30% improvement in mean specificity). When compared with the
ANN that was trained on the same augmented training data only small improvements in
mean sensitivity and specificity could be seen (up to 3%). However, the RBC was
clearly the best performing classifier algorithm overall.
The significance of the Optimised Region and Boundary Classifier is that it addresses
both accuracy and predictability and therefore it is potentially an inherently safer
classification algorithm. This would enable its use with more confidence in applications
where safety is critical, such as in Medical Devices.
Please use this identifier to cite or link to this item: