Date of Award

Spring 2005

Document Type


Degree Name

Doctor of Philosophy in Biomedical Engineering - (Ph.D.)


Biomedical Engineering

First Advisor

Stanley S. Reisman

Second Advisor

Ronald H. Rockland

Third Advisor

Richard A. Foulds

Fourth Advisor

Benjamin Martin Bly

Fifth Advisor

Matthias H. Tabert

Sixth Advisor

Wen-Ching Liu


The advance of blood oxygen level dependent function magnetic resonance imaging, (BOLD fMRI), allows researchers to non-invasively investigate the functioning human brain. The BOLD fMRI response to brief stimuli is called the hemodynamic response function (HRF), which can vary across brain regions and across subjects.

Models of the HRF are used to increase sensitivity of statistical maps; however, they often don't account for spatial and temporal variance. Physiological effects, such as learning, fatigue or habituation, introduce mismatch between statistical models and the data. Methods that use minimal a priori information and track time varying signals are able to show the processing of information over time and thereby elucidate such effects.

The method of Kalman filtering was employed to characterize mismatches occurring between statistical models and BOLD data. The Kalman filter operates on data point by point. This contrasts regression techniques, that use blocks of data to find a single estimate.

Functional MRI data was collected from ten subjects at Columbia University while they engaged in three visual experiments and four olfactory experiments. The Kalman filter was used to distinguish between the fMRI response to a 2 second and a 12 second visual stimulus. The results from this analysis showed the extracted responses from the two stimuli significantly differed. The same analysis was also used to distinguish between primary and secondary olfactory cortices. These brain regions have shown differential temporal responses to odorants. The extracted responses were not significantly different.

Extracted responses from one stimulus (visual or olfactory) were used to test if this subject specific information would predict the next experimental session, better than standard a priori models of the data. The results of this analysis showed this not to be the case. The extracted response over time to the odorant stimuli were tractable with the Kalman filter, and shown to decay as predicted from the literature. This temporal change was hypothesized to decrease predictability from one session to the next, causing the null result. To alleviate this, models were tested for their predictability across hemisphere, within session. The results showed that inclusion of subject specific information improved this fit over other a priori models.

The implications of this analysis are the ability to extract temporally varying fMRI responses over an experiment without knowledge of the expected response to a stimuli. Results of such analyzes offer a look into how the brain responds and processes stimuli over the course of an experiment. This contrasts method that offer summary, or average, results from an experiment.