Learning to Classify Biomedical Signals Miroslav Kubat,Irena Koprinska and Gert Pfurtscheller
Contents Two Medical Domain Decision-Tree Based Initialization of Neural Networks Tree-Based Initialization of RBF Networks Experiments Discussion
Two Medical Domains Sleep Classification Hypnogram : Horizontal axis~Time Vertical axis~ different sleep state To draw a hypnogram: EEG(electroencephalogram) ~ brain activities EOG(electrooculogram) ~ eye movements EMG(electromyogram) ~ muscle contractions
Two Medical Domains Brain-Computer Interface Recognition of Motor Commands From EEG Signals
Problem The complicated nature of domains impossible to use symbolic machine learning tech. such as rule or decision trees. Multilayer neural network sensitive to proper initialization of topology and weight.
Decision-Tree Based Initialization of Neural Networks General Idea of TBNN
Decision-Tree Based Initialization of Neural Networks Initialization of Weights and Full interconnection of Adjacent Layers OR-neuron
AND -neuron
Decision-Tree Based Initialization of Neural Networks Softening Intervals and Neural-Network Tuning
Tree-Based Initialization of RBF Networks RBF Networks and their Parameters
Tree-Based Initialization of RBF Networks Decision Tree Based Parameter Setting
Experiment Data set domain 1 ~ 8 data files 770~990 examples 15 attributes domain 2 ~ 3 data files 150~250 examples 11 attributes ~ 44 attributes
Experiments Results
Result and Discussion First domain(sleep state) The accuracy achieved by TBNN and TB-RBF is not worse that that of human experts Second domain(brain computer interface) The utility is obvious because the patterns of desynchronization of EEG are difficult of describe by rules, and learning appears to be only way to accomplish the task.