BCIG SPEAKER EVENT: “Pattern
Profile Classification in Multivariate Data: Diagnosis and Bioinformatics
Applications”
- view the seminar archive
Clinical
Center (Building 10) Medical Board Room (Room 2C116)
DESCRIPTION: Classification of pattern profiles in biomedicine is of
increasing importance due to the recent surge of available multivariate data
sets. Proper analysis of these sets includes a focused scrutiny of validation
and testing methodology. An ideal application, if possible, would be to utilize
serum samples to determine the strength of existing profiles that could indicate
disease states. This talk encompasses an application of machine learning to a
suite of MALDI-TOF data, performed with collaborators at NHLBI. The study used
minimally treated samples that retain the original spectrum of protein content
and could be attained quickly for use in an emergency care facility. Blood serum
samples were obtained for 76 patients presenting in the emergency room at the
time they were being evaluated for venous thromboembolism (VTE). Mass
spectroscopic data were subsequently obtained for the samples, processed with
the pattern recognition methodology and the resulting VTE classifications were
compared with conclusive clinical tests performed later in the patients’
treatment. Classification performance of the pattern recognition methodology was
compared with alternative methods, current standard-of-care assays and analyzed
with ROC plots. Generalization performance of the final system, a trained neural
network, was assessed using an independent test set withheld until the end of
analysis. Using ROC plot analysis, a standard metric of classifier performance,
the neural network had an area under the curve (AUC) of 0.85, exceeding the
performance of the d-dimer assay currently in clinical use. The method for
construction of the classifier was further validated by comparative analysis on
two data sets serving as controls: (1) a random simulated set of data with no
profile patterns and (2) a data set with “spiked”, or known patterns. A
reproducible, multivariate profile was detected in the complex spectra from
minimally processed samples and the resulting classifier out-performed the
current, clinically available assays for the diagnosis of VTE. The conclusion of
this study was that machine learning methods show promise for successful
classification of patients with VTE, using MALDI-TOF data from human serum. The
conclusion of this talk will be a general discussion of methodology for testing
and validation of pattern profile classifiers and the role of such classifiers
in biomedicine and defense.
|
 |
3:00 - 4:30 pm October 11, 2007
Judy Dayhoff, Ph.D.,
Guest Scientist,
National Library of Medicine
JUDITH E. DAYHOFF: holds a Ph.D. in Biophysics from the University of
Pennsylvania and a B.A. in Chemistry from Duke University. She is currently a
guest scientist at the National Library of Medicine. Her research interests span
protein structure and interaction in bioinformatics as well as topics in
advanced computational science such as multivariate pattern profiling and
capability analysis of neural networks. She is currently classifying protein
interaction sites via evolutionary methods to test and analyze existing
interaction databases. She has recently developed and applied novel analytical
techniques for pattern profiling of multivariate data and participated in their
application to MALDI-TOF mass spectrometry data, thus enabling a test of the
classification of disease versus non-disease patients based on blood serum
samples. In a paper about to appear in Neural Computation, she has extended the
synfire chain neural network concept to have full computational capabilities.
She has applied artificial neural networks to the analysis of prostate cancer
data, and pioneered the uses of artificial neural networks for clinical decision
support, researching a series of concepts on the validation and testing of
multivariate classifiers. She co-developed the adaptive time-delay neural
network for spatiotemporal pattern recognition, and has researched timing
relationships in the electrical activity of nerve cells, including developing
new methodology for detecting synchronous firing in neuronal ensembles. Dr.
Dayhoff is author of the book Neural Network Architectures: An Introduction (Van
Nostrand Reinhold, 1990), and co-editor of the book Neural Networks and Pattern
Recognition (Academic Press, 1998), as well as over 65 research papers. She has
taught graduate courses in neural networks at the National Institutes of Health
and at the University of Maryland. Dr. Dayhoff has been a visiting research
faculty member at the Naval Research Laboratory, the Naval Surface Warfare
Center, and at the Air Force Phillips Laboratory. She is on the College of
Fellows for the International Neural Network Society, and was President in 1997.
She has previously been on the engineering faculty at the University of Maryland
and on the mathematics faculty at Rutgers University, and has been a Visiting
Scholar at Stanford University.
Previously: Chief Scientist, CRSI, Silver Spring, MD and Contractor, NHLBI
Email: jed2277@earthlink.net
|
