Developing an assistive interface for individuals with spasticity disorders cd.

5. MITIGATING SPASTICITY

To mitigate spasticity, patterns of motion have to be accurately identified by extracting features of motion. This helps in establishing user intent as to what button is targeted as well as minimize extraneous motion associated with CP.

There are many methods present in the literature that would help identify and mitigate spastic data. One such method is Principle Component Analysis (PCA) [12], in which the dimensionality of the feature space is reduced by developing a scatter matrix. As popular as PCA is in pattern recognition, it would not be an ideal choice since excess dimensionality is not an issue associated with our feature space. Another popular method is the use of a Hidden Markov Model (HMM) [12] to identify paths of most probable motion. While this method is effective, it still is computationally taxing and would require the user to define and train the model.

One last method that was attempted was to minimize the feature space to just the cursor locations and train a Bayesian classifier for the various data. A new GUI, MitSpacPoC, has been developed that would allow for both the training and classification of the data. The MitSpacPoC console is depicted in Fig. 3.



5.1 TRAINING

Training mode is based on Bayes’ Theorem


where P(B | w) is the posterior probability, P(w | B) is the class-conditional, P(B) is the prior, and P(w) is the overall evidence. Given that the discriminant function is monotonically increasing, eq. 1 can be simplified by removing the normalizing factor to


5.2 CLASSIFICATION

Once the data has been sufficiently trained, classification can be performed, allowing for the estimation of user intent. Classification is done utilizing the Maximum A Posteriori (MAP) for the best fit. Some classification rules are presented below:
– If g(B1 | wi) > g(B2 | wi) and g(B3 | wi) then choose B1,
– If g(B2 | wi) > g(B1 | wi) and g(B3 | wi) then choose B2,
– If g(B3 | wi) > g(B1 | wi) and g(B2 | wi) then choose B3.