Neural net identification of thumb movement using spectral characteristics of magnetic cortical rhythms.

Neural nets have shown great promise as tools for reducing and examining multi-dimensional data. When carefully tuned with selected data sets of individual subjects neural nets have indisputable potential in identifying distinct stages of voluntary finger movements. However, robust, automatized data description methods would be needed to eventually extend the use of neural networks into visualization of brain activity during more complex, multimodal tasks where the cortical processes are not equally well understood. We explored the suitability of a self-organizing map (SOM) in the widely studied case of voluntary finger movements (left and right thumb), using as input such spectral characteristics that showed systematic task-dependent changes when averaged over repeated movements. SOMs constructed without individual fine-tuning and with generally chosen training parameters from these spectral features identified correctly 85% of the ongoing movements but, somewhat surprisingly, not the side of thumb movement. Even for this inclusive choice of input, the neural nets were sensitive to transient signals, but focused fine tuning, based on a priori known subgroups in the data, is clearly required for more detailed classification. Thus, a neural net visualization is likely not the most attractive first approach for characterization of cortical processing during complex multimodal tasks.