BIOST 2025: Biostatistics Seminar Notices

The work discussed in this talk is motivated by the need for a procedure that can predictepileptic seizures from EEG data which take the form of non-stationary time series.  We approach this problem by developing a spectral based classification rule to discriminate between baseline brain EEG and EEG preceding the onset of a seizure.  Existing methods in non-stationary time series classification assume that the second moment properties of times series from the same population are the same.  This is usually not true in realapplications and can lead to misclassification. A model for a family of time series is introduced that uses a hierarchical structure on their log-spectra. This model assumes thatwhile a family of time series share some similarity characterized by the population-average spectrum, each time series has its own characteristics modeled by the unit-specific deviation in terms of its log-spectrum. We then propose nonparametric methodsto estimate the population-average log-spectrum and the between-unit variance function.We develop a quadratic rule for discriminating between different populations based on the estimated mean log-spectra and the variance functions.

The value of observational research to the pharmaceutical industry is growing.  Observational research is becoming increasingly important for all phases of drug development as well as for post marketing activities.  Most recently the passage of the Food and Drug Administration Amendments Act of 2007 (FDAAA) requires manufacturers to ensure that the benefits of a drug outweigh its risks, further elevating the value of observational research. This talk will give an overview of the drug development process, describe the types of data and observational research necessary at each phase of drug development, and give a broad overview of some of the cutting edge statistical techniques being developed to overcome some of the problems inherent in observational studies of effects of drug exposures.

The DuPont Company has maintained an epidemiology program since 1955 that conducts routine occupational health surveillance for its worker population.  The presentation will describe epidemiologic practices that support a global manufacturing company with a range of products and services including agriculture and food, building and construction, electronics and communications, industrial chemicals, and transportation.  Fundamental and reconsidered approaches to occupational epidemiologic methods including the healthy worker effect, empirical Bayes adjustments for standardized mortality ratios, and  exposure-response analyses will be illustrated.

Suggested References

DuPont Epidemiology

1.  Pell S, O'Berg MT, Karrh BW.  Cancer epidemiologic surveillance in the DuPont Company.  J Occup Med 1978;20:725-40.

2.  O'Berg MT, Burke CA, Chen JL, Walrath J, Pell S, Gallie CR.  Cancer incidence and mortality in the DuPont Company: an update.  J Occup Med 1987;29:245-52.

3.  Symons JM, Kreckmann KH, Sakr CJ, Kaplan AM, and Leonard RC.  Mortality among workers exposed to acrylonitrile in fiber production: an update.  J Occup Environ Med 2008;50:550-60.

4.  Sakr CJ, Symons JM, Kreckmann KH, Leonard RC.  Ischemic heart disease mortality among workers with occupational exposure to ammonium perfluorooctanoate. Occup Environ Med 2009; doi:10.1136/ oem.2008.041582.

Occupational Health Surveillance Methods

5.  Carpenter LM, Maconochie NES, Roman E, Cox DR.  Examining associations between occupation and health by using routinely collected data.  J R Statist Soc A 1997;160:507-21.

6.  Steenland K, Bray I, Greenland S, Boffetta P.  Empirical Bayes adjustments for multiple results in hypothesis-generating or surveillance studies.  Cancer Epi Bio Prev 2000;9:895-903.

7.  Checkoway H, Pearce N, Kriebel D.  Selecting appropriate study designs to address specific research questions in occupational epidemiology.  Occup Environ Med 2007;64:633-8.

In the health care setting, data are often obtained using administrative billing databases or through retrospective chart review.  Thus, the data have many limitations, including non standardized definitions, free text as opposed to discrete data, missing data, inter-rater reliability issues for chart review, and data elements that are needed but not recorded consistently or at all.  In addition, many independent databases exist within our institution and collect data on the same patient populations but lack standardization and adequate ontologies.  We recognized the tremendous opportunities to share resources, share data, increase collaboration with respect to research studies, learn from each other to reduce redundant efforts and collect standardized, valid and reliable data that can be easily accessed.
               
The methods to develop a a proactive approach to data collection will be discussed consisting of: 1) A database needs assessment survey; 2) An evaluation of existing systems; 3) An evaluation of vendors for purchasing software or outsourcing; 4) Physician Input on final options;  5) Pilot testing a proposed solution 6) building clinical outcomes registries.  In addition, the traditional data quality and statistical analysis issues using administrative data or retrospectively collected will be discussed and compared to the new method of prospective data collection to support statistical analyses.

Currently, the conditions included in the pediatric Clinical Outcomes Registry at our institution are as follows: neonatal surgeries urologic surgeries; scoliosis; diabetes; cleft palates/lip; cardiothoracic surgeries; pain management and otitis media and voice and airway patients.

The main message of my talk is to show how statisticians can approach important biological questions using basic analytical tools. Two published studies and further work will be presented to illustrate the point. Here are the ideas:

1) Detection of disease gene interaction effects among the enormous array of single nucleotide polymorphism (SNP) combinations represents the next frontier in genome-wide association (GWA) studies. We propose to look at the interactions in GWA datasets on the basis of the pattern and nature of the interaction, which can be classified as essential (EI) or removable (RI). I’ll present the analytical framework, including the qualitative conditions for screening EIs/RIs and a RI-to-EI likelihood ratio score to quantitatively measure the effects. In application to several real GWA studies, we find that the scores follow an exponential distribution, except in the upper 10^-8 tail region in which the scores become irregular. Even though several issues remain to be solved, I wish to convince you that this approach is conceptually simple and computationally efficient, and therefore, it may potentially be useful in detecting interactions in GWA data that can be visualized and unequivocally interpreted.

2) Dendritic cells (DCs), known as the professional antigen presenting cells, play a key role in initializing immune responses. Knowledge involving DCs generation and maintenance is important in many aspects of organism’s heath. To examine cell divisions and cycles, experiments are often laborious and have their limitations. To this end, mathematical/statistical modeling can help to provide insights into this dynamic system. We collaborate with Dr. Liu at Rockefeller who has providing us some experimental data and relevant immunological backgrounds to construct a meaningful model on the relationship of DC’s replenishment and death rates during cell cycles. First, we consider the steady state under which cell population sizes, flow and death remain balanced. We construct the model based on three differential equations and derive the analytical solutions DC input rate, cell division rate and death rate. Our results were subsequently validated by Dr. Liu. With that, the next task is to consider a perturbed system. We explore this through extensive computer simulations. Finally, by combining the properties of both steady- and perturbed-states we derive a unified model for DC’s dynamic life cycle. The profound utility of these computational undertakings is at least two-fold: they allow immunologists to effectively design their experiments, which, in turn, may provide the mechanistic insights for DC-related disorders including cancer and autoimmune diseases.