Work

Differential Projection Pursuit for Large-scale Biomedical Data, with Applications to Flow Cytometry

Year: 2026
Type: dissertation
Abstract: Modern biomedical research generates increasingly large and high-dimensional datasets that present major challenges for conventional statistical approaches. First, complex biological populations frequently exhibit differences that are not reflected by conventional summary statistics. Second, analysis of large datasets can be computationally expensive. This dissertation develops Differential Projection Pursuit and associated computational methods for identifying and characterizing such complex population differences. Projection pursuit (PP) is a method in multivariate data analysis for finding low-dimensional projections that uncover interesting structures in the data. Differential projection pursuit (DPP) is proposed as an extension of PP that finds projections that maximize the difference between two or more distributions. The proposed method compares projected probability density functions, allowing detection of differences that extend beyond changes in mean. Theoretical analysis and simulation studies characterize the behavior of DPP under different distributional conditions. Under homoscedastic Gaussian assumptions, DPP preserves the ordering produced by Linear Discriminant Analysis (LDA), while providing greater flexibility when those assumptions do not hold. Further, Data Nuggets were incorporated into the DPP framework to dramatically reduce the computational cost of analysis. Data Nuggets reduce datasets containing millions of observations to a few thousand weighted representatives while preserving the underlying structure of the original data. The utility of the proposed framework is shown through analysis of high-dimensional flow cytometry data. An initial proof-of-concept analysis shows that DPP can identify cellular regions exhibiting biologically expected differences. The methodology is then incorporated into an automated analytical workflow combining cell selection, subject-level baseline centering, Data Nugget compression, weighted DPP, and downstream characterization of identified differential regions. Applied to a complete flow cytometry dataset, the workflow reduces approximately 11 million multivariable cellular observations to 4,858 Data Nuggets. The analysis identifies localized regions enriched in different treatment groups and reveals distinct protein-expression patterns and cellular subpopulations within these regions, illustrating the ability of DPP to serve as a method for detecting multivariate population differences. A complementary biomedical application examines racial and ethnic disparities in treatment and outcomes following ST-segment elevation myocardial infarction (STEMI). Using a large statewide cardiovascular registry from New Jersey, temporal patterns in percutaneous coronary intervention utilization and mortality are evaluated among Black, Hispanic, and White patients before and after implementation of the United States Initiative to Eliminate Racial and Ethnic Disparities in Health. Over the study period, PCI utilization increased and disparities decreased, particularly among Hispanic patients, while one-year all-cause and cardiovascular mortality declined following implementation of the initiative. Some differences among racial and ethnic groups nevertheless persisted. Together, this work develops a progression from statistical methodology to computational implementation and biomedical application. More broadly, this dissertation demonstrates how statistical methodology can identify interpretable differences between populations as biomedical data continue to increase in both dimensionality and scale.
Language: English
Open Access status: green
OpenAlex ID: W7235116316
Cites:
Cited by:
Related to: 10