Research on these projects advanced on three fronts.

(1) Real-time neural interface and decoding [2002]: Using linear and nonlinear discriminant analyses, we developed the capability to categorize individual rat behaviors on the basis of the observed neural spiking activity alone. This involved isolating spikes from many neurons simultaneously and training either linear (canonical discriminant analysis) or nonlinear (supervised backpropogation artificial neural network) models to recognize various neuronal states. After such models were available, we decoded ensemble firing activity in real time by passing neural data from behaviring rats implanted with microwire electrode arrays across an ethernet link to a PC executing Matlab code running our model. Successful decoding of neural activity was demonstrated by successful classification of behavioral events at greater than 90%.

(2) Characterization of rhythmic hippocampal activity in a short-term memory task [2003]: The significance of rhythmic modulation of firing activity within and across brain regions is becoming clear in many lines of neuroscience research, from olfactory encoding to a putative role in consciousness.  Theta oscillations (4-12 Hz) in spatial representation in rat hippocampus are well characterized, though their functional significance has been less convincingly described. We sought to understand the role of theta and gamma oscillations during cognitive, non-spatial processing that occurs during the short-term memory test paradigm DNMTS (delayed-nonmatch-to-sample). We applied a variety of tools to this problem including spectral and modulation analyses, spike train cross-correlations, directed coherence , ICA, and classification testing. 

(3) Independent Components Analysis of human fMRI [2001]: We investigated (2001) independent components analysis (ICA) developed by Bell and Sejnowski at the Salk Institute (and described for fMRI data by McKeown, now at Duke). This method has the potential of revealing novel dynamics of brain functioning during memory tasks because, like nonlinear time series analysis of neuron recordings, it enables continuous, ongoing analysis of patterns of neuron firing activity even when that  activity is not correlated with any obvious, observable event or stimulus. An interactive Matlab software shell was developed around the Bell and Sejnowski neural network implementation of ICA adapted for fMRI data (largely by McKeown) for analysis of fMRI data acquired at Baptist Medical Center under the Biomedical Engineering Program. Capabilities were added including standard fMRI preprocessing, algorithmic out-of-brain voxel removal, multi-frame alignment, spatial- and temporal smoothing and filtering, principle components analysis (PCA), and various visualization and significance testing capabilities. The primary objective of this study was to establish ICA capabilities and to benchmark ICA against our in-house statistical parametric mapping (SPM) capabilities.