Alan D. Kaplan (Washington University in St. Louis), “Information Processing for Biological Signals: Application to Laser Doppler Vibrometry” (2011)

Alan D. Kaplan (Washington University in St. Louis), “Information Processing for Biological Signals: Application to Laser Doppler Vibrometry”, Advisor: Prof. Joseph A. O'Sullivan (2011)

Signals associated with biological activity in the human body can be of great value in clinical and security applications. An approach for the modeling and processing of biological signals is developed using graphical models. Graphical models describe conditional dependence between random variables on a graph. When the graph is a tree, efficient algorithms exist to compute sum-marginals or max-marginals of the joint distribution. Some of the variables correspond to the measured signal, while others may represent the hidden internal dynamics that generate the observed data. Three levels of hidden dynamics are outlined, which enable models to be constructed that track internal dynamics on differing time scales using a hierarchical graphical model structure. For cardiovascular signals, the proposed time scales correspond to intra-beat dynamics, the dependence across beats, and state transitions occurring across days.

Experimental results of this approach are presented for a novel method of recording bio-mechanical activity using a Laser Doppler Vibrometer. The LDV measures surface velocity on the basis of the Doppler shift. This device is targeted on the neck overlying the carotid artery, and the proximity of the carotid to the skin results in a strong signal. Vibrations and movements from within the carotid are transmitted to the surface of the skin, where they are sensed by the LDV. Changes in the size of the carotid due to variations in blood pressure are sensed at the skin surface. In addition, breathing activity may be inferred from the LDV signal.

Individualized models are evaluated systematically on LDV data sets that were acquired under resting conditions on multiple occasions. These models are applied to the problem of identity verification using the LDV signal. Identity verification is an important problem in which a claimed identity is either accepted or rejected by an automated system. The system design that is used is based on a loglikelihood ratio test using models that are trained during an enrollment phase. A score is computed and compared to a threshold. Performance is given in the form of False Nonmatch and False Match empirical error rates as a function of the threshold. Confidence intervals are computed that take into account correlations between the system decisions.

For details, please access the full thesis or contact the author.