Date of Award
Master of Science in Computer Science - (M.S.)
Jason T. L. Wang
Jorge P. Golowasch
To further an understanding of how a neuronal population generates patterns of rhythmic activity, the temporal dynamics of the group of neurons must be formalized. Essential to this pursuit, is the ability to reliably detect and separate the classes of single-unit neuronal activity from multi-unit extracellular signals recorded in a single channel. This study proposes a unified approach to automatically detect and classify single-unit bursts, and to observe the precise onset and offset of burst activity. Existing approaches to the problem fundamentally depend on the statistics of spike waveform variability, both extrinsic and intrinsic to the neuron. In contrast, the proposed approach depends on statistics that characterize the burst variability. An unsupervised learning procedure is implemented using hierarchical clustering to derive a complete and natural description of the variability in terms of clusters of bursts that possess strong internal similarities. Redundant solution vectors are used to parameterize each cluster, and a fuzzy classification approach assigns each burst to a class. Accuracy of the technique is demonstrated on in vivo and in vitro recordings of the triphasic pyloric rhythm in stomatogastric ganglion of crab Cancer borealis. The results, evaluated against a widely used manual classification approach, show that the technique performs detection and classification with comparable accuracy and quantifiable certainty, and is robust to background activity and noise.
Natarajan, Rama, "Uber-Claws : unsupervised pattern classification for multi-unit extracellular neuronal burst extraction" (2003). Theses. 655.